PrototypeVerdict · Lab-backed troubleshooting
Your next safe step, before another fix.
Your app looks broken. Before another fix attempt, narrow the symptom, protect your data, and leave with a focused next message and a checklist.
Small unpublished apps · Oct 29 lab · Nov 12 fix-loop note · Suggested checks, not a diagnosis
The guided flow needs JavaScript and has not run. Use this safety summary and the closed sections below. Nothing here is a diagnosis, a security audit, or a launch clearance. This widget does not transmit your answers. The host page may still use its own analytics.
Safety stop
If someone can already see records they should not see, stop exploring them. Do not paste exposed contents into a prompt. Privately notify the authorized owner with a redacted description. Hiding navigation is not a full repair.
Symptom safety summary
| What you see | First safe move |
|---|---|
| Data saves here but is empty elsewhere | Preserve the original browser view. Do not clear site data or localStorage. A reload does not prove shared storage. No database migration was tested. |
| Login stalls or a page is blank | Record a redacted error category. A generic 500 is not a recursive-policy diagnosis. Do not disable access controls or expose elevated credentials. There is no SQL recipe here. |
| Cannot reach the admin area | Do not promote yourself. If you are not the owner, ask the authorized owner. Ordinary login working is different from a failed login. |
| Client controls or permissions do not match | If records are already exposed, use the safety stop. Inspect actual settings, not an AI “fixed” message. Preview simulation is not independent login. No plan upgrade is guaranteed to fix it. |
| Out of tokens or credits | An explicit quota message is different from a slow build. Read the current allowance display. Do not assume a reset time, a purchase, or a later CSV export. Ours stayed unrun. |
| The build is waiting for approval | Read the exact action. Do not send a blanket “continue” for deletion, access changes, or publishing. A code restore may not restore database records. |
| Suspicious or broken copy | Separate supplied facts from invented statistics, prices, testimonials, or escape characters. A working form does not validate marketing claims. No copy repair was tested. |
| Something else | There is no matched lab incident. Ask for one discriminating check before changing data or access. No match does not mean the app is safe. |
Read or print without a separate download
Saving a separate file uses buttons in the guided tool, and those buttons need JavaScript. The sections below are the same public text. The prelaunch kit is 24 proposed checks, not performed lab tests. Every result in the evidence worksheet starts Not checked. A completed checklist is not launch clearance. There is no link to a file beside this page.
Rescue worksheet text
PROTOTYPEVERDICT — APP BUILDER RESCUE WORKSHEET
Version 2026-11-07.1 | Original lab: Oct 29, 2026
Use privately. This worksheet does not connect to your app or send anything.
No diagnosis, automatic repair, security audit, or guarantee of recovery.
Use synthetic records and authorized test accounts, not real client data.
FIRST: STOP IF DATA IS EXPOSED
If someone can already see records they should not see, stop exploring them.
Do not paste exposed contents into a prompt. Privately notify the authorized
owner with a redacted description. The owner should follow their containment
process before further real-data use. Hiding navigation is not a full repair.
1. AUTHORITY AND CONTEXT (fill locally; do not include secrets)
Am I the owner or an authorized maintainer? [yes / no / unknown]
If no/unknown, ask the owner; do not change roles, records, or permissions.
Builder: [ ] App type: [ ] Test/live environment: [ ]
Same app/account/environment in both views? [yes / no / unknown]
Never include passwords, keys, session tokens, private records, or complete
email-confirmation/callback URLs in a shared handoff.
2. ONE REPRODUCIBLE SYMPTOM
What I did: [one step]
Expected: [one outcome]
Observed: [redacted symptom/error]
Last known working state: [description, not a token-bearing URL]
Known unexpected access? [yes / no known / not checked]
“No known” is not proof of access isolation.
3. CHOOSE A DISCRIMINATING CHECK
DATA: Preserve the original browser view. Do not clear site data/localStorage.
Check app, account, environment, and whether the origin changed. A reload pass
does not prove shared storage. Ask where saves/reads actually go before a
migration. No database migration or data recovery was tested in our case.
LOGIN: A generic 500 does not prove recursive RLS. Record the actual redacted
error category. Blank after email confirmation differs from “Admins only”.
Never disable RLS, expose elevated credentials, or broadly allow access.
ADMIN: Does ordinary login work? Does an authorized admin already exist?
Only the owner should establish an explicit first-admin bootstrap through a
trusted management path. No public promotion button or user-editable role fix.
PERMISSIONS: Inspect actual settings, not the AI's “fixed” message. Separate
page access, action visibility, and record restrictions. Preview simulation is
not independent client login. Do not publish an unsafe app merely to test it.
USAGE: An explicit quota gate differs from a slow build. Read the actual
allowance display. Do not repeatedly submit or assume an upgrade/reset fixes
it. Our Bolt CSV follow-up was blocked and remained unrun.
CHECKPOINT: Read exactly what will change before approving. Do not send a
blanket “continue” to a deletion/publication/access change. Code restore may
not restore database records. Our Base44 “continue” was logged as redundant.
COPY: Check unsupported numbers, testimonials, prices, and promises against
approved facts. A working form does not validate marketing claims. Ask for a
copy-only diff. Do not run an indiscriminate global escape-string replacement.
4. NEXT-MESSAGE DRAFT (suggested, not tested word for word; not sent)
What I did: [one step]. Expected: [one outcome].
Observed: [redacted symptom/error].
Please propose one safe discriminating check before changing anything.
Preserve existing records and access controls. Do not guess a root cause,
expose credentials, publish, delete data, or incur a new purchase. Explain the
smallest proposed change and its verification plan. Change nothing unrelated.
5. VERIFICATION — DEFAULT EVERY ITEM TO NOT CHECKED
For each applicable item record: NOT CHECKED / I CHECKED PASSED /
I CHECKED FAILED / NOT APPLICABLE. These are your records, not certification.
[ ] Intended function works with a synthetic record.
[ ] It still works after reload in the intended app/account/environment.
[ ] If shared storage is intended, another authorized browser/device sees
the same synthetic record. This was not tested in the storage lab case.
[ ] Signed-out visitors cannot reach protected pages or private records.
[ ] A separate non-admin cannot reach admin routes or actions.
[ ] Two independent clients cannot read or modify each other's fake records.
[ ] Ordinary users cannot modify their own administrative role.
[ ] Original records remain intact and any migration reconciles ownership,
counts, and totals before old storage is retired.
[ ] Forms, navigation, and unrelated functions still work.
A failed or unchecked access test means do not treat access as verified.
Even all self-recorded passes are not an audit or launch clearance.
6. EVIDENCE AND LIMITS
Observed: [historical statement + source]
Inferred: [plausible interpretation, not confirmed]
Suggested: [next check or message, not a tested fix]
Not verified: [explicit gaps]
Our small Oct 29 cases used fake data and unpublished previews, one run per
app per tool. They do not establish general builder reliability. No new app
tests were performed to create this tool. Bolt's independent non-admin check
was not run. Softr used Preview as, not independent login. The Lovable storage
migration and permanent confirmation-flow repair were not tested.
PUBLIC SOURCES
Historical case catalog:
https://prototypeverdict.com/guides/what-actually-broke/
Method and exceptions:
https://prototypeverdict.com/how-we-test/
Usage dataset:
https://prototypeverdict.com/usage-index/
Troubleshooting companion:
https://prototypeverdict.com/guides/fix-ai-app-builder-errors/
Storage semantics:
https://developer.mozilla.org/en-US/docs/Web/API/Window/localStorage
Database authorization concepts (not a SQL recipe):
https://supabase.com/docs/guides/database/postgres/row-level-security
Redirect reference (not a proven cause):
https://supabase.com/docs/guides/auth/redirect-urls
Softr record restrictions (current docs, not a retest):
https://docs.softr.io/user-groups-and-permissions/global-data-restrictions
Code/database restore distinction (not tested in our lab):
https://docs.lovable.dev/features/projects/history
https://support.bolt.new/managing-projects/rollback-backup
Review and redact before sharing. Nothing is sent by this worksheet.
Prelaunch kit text, version 2026-11-07.2
# The prelaunch test kit: prove what your small app actually does **PrototypeVerdict · Version 2026-11-07.2 · Prepared Nov 7, 2026 (CST, America/Chicago)** Companion: **TEST-EVIDENCE-WORKSHEET.csv** · Original lab observations: **Oct 29, 2026** A practical set of **24 proposed checks** for an app you own or are authorized to test. Each check gives prerequisites, steps, expected behavior, failure interpretation, retest instructions, and a stop condition. The CSV contains the same executable check descriptions—not just blank notes—and starts every result at **Not checked**. **These are suggested tests, not lab-tested repairs.** The original cases show why these checks matter, not that these new procedures have been run or that a builder is generally reliable. This kit is not a security audit, penetration test, compliance certificate, cost benchmark, or clearance to launch. No new app tests were performed in preparing it. ## Before you touch the app - Use only your own authorized app, preferably an isolated test copy with **synthetic data only**. If you cannot isolate real records or outbound actions, stop and ask the authorized maintainer to prepare a safe environment. Do not publish an unsafe app just to make a second-login test possible. - No destructive clears, deletion, schema resets, broad security changes, fake payment transactions, purchases, upgrade prompts, live migration, or blanket checkpoint approval. Nothing here authorizes a role change, a live cutover, publication, or sending messages to third parties. - Keep the original working browser/profile intact, especially if it might hold the only local copy. Do not clear cookies/site data/localStorage or close the sole private-browsing session to make a test “clean”. A second independent profile/session can test separation without destroying the first. - Read the exact app contract first: who may read/edit which records, which data should be shared, what should persist, and what counts as intentionally public. A public brochure page is different from a private client portal. **An undefined expectation is an open requirement, not a passing test.** - **Stop immediately** if real or unexpected private records appear. Do not inspect more, copy contents, test further actions, or paste exposed material into a prompt. Privately notify the owner with a redacted description and use their containment process. Restricting public access is an owner decision, not an action performed by this kit. - Keep passwords, keys, session tokens, auth links, signed download links, raw logs/HAR files, personal addresses, and private record contents out of the worksheet. Record route paths/labels and fake markers, not token-bearing URLs. All evidence stays local unless you independently choose to share a redacted copy. ## Fixture card: prepare once, reuse carefully | Label | What to prepare | Important distinction | |---|---|---| | Owner | Your explicitly authorized management context | An owner/editor may have wider access than an app admin; do not treat those as interchangeable | | Admin | A designated staff account, if your app has this role | The owner must approve initial setup through a trusted management path | | Client A | A separate account you control | Own fake marker `PV-A-001`, business `Sample Alpha`, note `Synthetic A record` | | Client B | Another distinct account you control | Own fake marker `PV-B-001`, business `Sample Beta`, note `Synthetic B record` | | Signed out | An independent unauthenticated session | An editor’s owner preview is not signed out | | Optional invoice fixture | Two fake lines: quantity 2 × rate 10 and quantity 1 × rate 5 | Expected synthetic total 25 before any explicitly declared tax/rounding rules; use no payment provider | | Optional files | Small generated inert PDF/image/text fixtures with obvious synthetic labels | Only file types your app declares; no real IDs, tax returns, receipts, malware, or stress-size files | Use your own controlled sign-in destinations inside the app’s trusted flow; this kit supplies no mailbox addresses. Do not place passwords or contact details in the CSV. If your system lacks roles or uploads, mark the relevant checks **Not applicable** with a reason after reviewing the requirement—not by assuming they passed. Write a small permission matrix before PV-09–PV-13: rows for Signed out, Client A, Client B, Admin; columns for public page, A read/edit, B read/edit, admin read/action, and private files. Define allow/deny for each. Staff-wide access may be legitimate; cross-client access must match the intended contract. A hidden link does not prove a deny, and an empty own-record view is not a successful cross-client test. ## A small, realistic run order 1. **Identity first:** PV-01. If identity/environment is uncertain, do not trust later results. 2. **Core save/content/experience:** PV-02, PV-04–PV-05 where a backend exists, PV-18–PV-19. 3. **Private apps:** PV-09–PV-11, with legitimate own-record allow cases before denial checks. PV-12–PV-13 need owner/maintainer involvement; do not improvise low-level requests. 4. **Choose only relevant branches:** sessions/devices PV-03/PV-06; confirmation/admin PV-07/PV-08; status PV-14; upload/files/export PV-15–PV-17. Observation-only PV-20/PV-21 do not require starting paid work or creating a dangerous checkpoint. 5. **Migration/restore:** PV-22/PV-23 are optional maintainer-assisted rehearsal/planning, never an instruction to cut over production. 6. **After any approved change:** PV-24 and the specific failed check, not all unrelated checks blindly. There is no promised completion time. Stop at a failure that makes later checks unsafe or uninterpretable. You do not need to run every optional check to obtain a percentage score; **this kit deliberately has no score**. ## How to use the evidence CSV Download/save a fresh local copy for each environment/version/run. The file is UTF-8 with a BOM for common spreadsheet apps, has 24 data rows, stable test IDs, quoted fields, and no macros or executable formulas. It is wide because each row contains its actual procedure. Freeze the identity/title columns, wrap text, and hide long procedure columns temporarily if useful; keep them in the file. - Fill `run_id`, `app_label`, `environment_label`, `app_version`, and `identity_label` with nonsecret labels. Record actual test identity separately from the recommended identity in the procedure. - `evidence_result` and `retest_result` begin **Not checked** on every row. Allowed user-recorded values: **Not checked, Passed, Failed, Blocked, Not applicable**. Supply a rationale for Blocked/Not applicable. A planned or reviewed procedure is not an executed pass. - Fill `observed_result_redacted`, `evidence_reference_redacted`, and `tested_at_with_timezone` only after running the check. Use a local evidence label, not a private sharing URL. Report timestamps with their timezone, for example `2026-11-07 11:10 PM CST`; no timestamps are prefilled as if tests ran. - Use `retest_run_id`, `retest_result`, and `retest_notes_redacted` for a linked follow-up; preserve the original failure. A separate copied sheet per material version is preferable to silently replacing history. - An AI’s “fixed” reply, intended behavior from documentation, or passing screenshot from the owner context cannot substitute for the stated test and actual result. - Before sharing any worksheet, inspect every cell and evidence attachment for secrets/private data. Treat imported spreadsheet content cautiously; never add formulas copied from untrusted material. This supplied template has no formulas. For a completed sheet, import evidence columns explicitly as text. Before opening or sharing it in a spreadsheet, inspect user-entered cells for formula-leading =, +, -, or @, including after leading whitespace/control characters. CSV quoting alone does not neutralize formulas; use a reviewed plain-text import/export workflow and do not enable external links or macros. ## The checks ### PV-01 · Prove which environment and identity you are testing **When:** Every app **Identity:** Owner; Client A; Client B; Signed out **Mode:** Start here · Proposed test, not a tested repair **Prerequisites:** You own the app or have explicit permission. Use a private test copy containing only fake records. Have three distinct controlled accounts if roles exist, and separate browser profiles or otherwise independent sessions. Never reuse the owner session as a client. **Steps:** 1. Record a nonsecret app/environment label and version. 2. In each separate session, read the visible signed-in identity and intended role. 3. Mark the signed-out session separately. 4. Verify all sessions target the same intended environment; note protocol/host/port privately without token-bearing URLs. 5. If using Preview as, record that method instead of claiming independent login. **Expected result:** Each context has an unambiguous identity, role, session, and environment. Owner/admin is not mistaken for Client A/B; preview simulation is labeled. **If it fails / what it does not prove:** Wrong account/environment or an owner session can create false passes and false missing-data reports. A hidden role label or preview switch is not proof of independent identity. **Retest:** Repeat after any account switch, domain/environment change, or role update. Do not reuse downstream evidence from a mismatched context. **Stop conditions:** Stop if you cannot establish authorization or see real records. Do not sign into accounts you do not control. **Why this is in the kit:** Softr checks used Preview as; Lovable/Base44 later used separate client accounts. These methods are not equivalent. **Sources:** L1; O1. See source register below. ### PV-02 · A completed save survives a reload in the same session **When:** Apps that save records **Identity:** Client A or intended single user **Mode:** Core data · Proposed test, not a tested repair **Prerequisites:** PV-01 complete. Pick a fake record type. Use marker PV-A-001 and note its expected field values locally. Keep any original browser-local records intact. **Steps:** 1. Create one fake record through the normal UI and wait for the completion state. 2. Open its detail view and note marker and key values. 3. Reload the same page in the same browser/session/origin. 4. Reopen the record and compare the values. 5. Record any visible error rather than repeatedly submitting. **Expected result:** Exactly the intended fake record appears with unchanged values after reload; the UI does not merely retain unsaved form text or show a duplicate. **If it fails / what it does not prove:** Missing or changed data could indicate a failed save, failed read, wrong environment/account, or storage behavior. A pass proves only same-session persistence, not a database or device sync. **Retest:** After a focused repair, use a new fake marker and repeat. Preserve the failing record and note both attempt IDs. **Stop conditions:** Stop on unexpected real data, duplicate submissions, or a write error. Do not clear site data or reseed the database. **Why this is in the kit:** Lovable invoice data passed refresh yet another-origin preview was empty; Base44 database records also passed refresh. **Sources:** L1; M1. See source register below. ### PV-03 · A later independent session can read the intended saved record **When:** Apps promising account-backed persistence between sessions **Identity:** Client A in a second controlled session **Mode:** Core data · Proposed test, not a tested repair **Prerequisites:** PV-02 complete. Shared/account persistence is part of your declared requirement. Keep the first working session open; do not close the only private/local-storage copy. **Steps:** 1. Open an independent controlled browser profile/session at the same intended environment. 2. Sign into Client A through the normal flow. 3. Locate PV-A-001 without recreating it. 4. Compare its saved fields. 5. Note whether the first session still sees it. Do not expect unrelated Client B to see a private A record. **Expected result:** The same authorized user reads the same record and values in the second session if that is the app contract. **If it fails / what it does not prove:** Failure does not prove deletion: compare environment/account first, then storage, read errors, and access rules. Browser-only storage may be an intentional design but does not meet account-shared requirements. **Retest:** Use the same saved marker after the maintainer explains the fix; repeat both original and independent-session reads. **Stop conditions:** Do not log out/clear/close the sole surviving original session to run this. Stop if sign-in is locked or repeated attempts fail. **Why this is in the kit:** Cross-session/device sharing was not tested in the Lovable storage case; this is a proposed test, not a reported lab pass. **Sources:** L1; M1. See source register below. ### PV-04 · Confirm the saved row exists in the intended backend **When:** Apps claiming database-backed saves **Identity:** Owner/authorized backend maintainer; Client A performs UI save **Mode:** Core data · Proposed test, not a tested repair **Prerequisites:** Read-only access to the intended data-management interface, and an agreed record/table mapping. If unavailable, record Blocked, not Passed. Never reveal backend keys or grant wider access to inspect. **Steps:** 1. Save one uniquely marked fake record through the normal app. 2. In the authorized read-only backend view, find that exact marker in the intended environment. 3. Compare key saved fields and intended ownership mapping. 4. Note only a nonsecret table label, marker, count, and outcome in the worksheet. Do not export user tables or auth records. **Expected result:** The corresponding backend row exists once, has the expected values, and is linked to the correct test owner. State clearly if the app intentionally has no backend. **If it fails / what it does not prove:** A UI success toast without a backend row is not evidence of a durable database write. Check storage design, request completion, wrong environment, or rejected write; do not assume a cause. **Retest:** With a new marker, repeat UI save plus backend confirmation and then PV-02/PV-03. Keep old evidence separate. **Stop conditions:** Stop if you would need to create a privileged token, change grants, expose secrets, or inspect real customer records. **Why this is in the kit:** Base44 invoice tables were observed; Lovable invoice notes suspected local-only storage. Backend existence must be checked per app, not inferred from builder branding. **Sources:** L1; M2. See source register below. ### PV-05 · Distinguish an empty dataset from a failed read **When:** Apps displaying record lists or dashboards **Identity:** Client A; optional authorized maintainer **Mode:** Core data · Proposed test, not a tested repair **Prerequisites:** A known fake backend row from PV-04, or an explicitly declared empty dataset. No need to invent a failure, disable access, disconnect a production backend, or edit credentials. **Steps:** 1. Open the known fake record/list normally. 2. If empty or stuck, note whether the UI says no records, loading, or error. 3. An authorized maintainer may inspect the existing request status and redacted error category using trusted logs/devtools; do not export HAR files, request headers, cookies, tokens, or response contents. 4. Compare backend row existence with the user-visible read. If no failure occurs, record only the successful-read behavior; error-state behavior remains untested. **Expected result:** Known permitted fake records load; if an existing read error occurs, the app distinguishes it from a genuinely empty result and exposes no sensitive debugging output. **If it fails / what it does not prove:** A present row plus an empty view may involve filtering, identity, permissions, environment, or a request failure. Empty results can also be the intended outcome of a deny policy; a 500 does not prove recursion. **Retest:** After repair, repeat the exact read with the same identity and known marker. A maintainer can plan a separate controlled failure-state test, but do not claim one occurred here. **Stop conditions:** Stop on unexpected access or sensitive debug output. Do not broaden RLS or use elevated credentials in a browser to make the list populate. **Why this is in the kit:** Bolt login failed alongside a recursive-policy 500; Lovable another-origin dashboard was empty. The symptom alone does not identify the failed layer. **Sources:** L1; M2; O1. See source register below. ### PV-06 · Compare intended shared data across origins or devices **When:** Only if cross-device or alternate-domain access is an intended feature **Identity:** Same Client A; independent controlled browser/device **Mode:** Conditional data · Proposed test, not a tested repair **Prerequisites:** PV-02/PV-04 if available; a documented expectation of shared storage; both target addresses belong to your authorized app. Preserve the original local-data session. Do not copy token-bearing preview links. **Steps:** 1. Record whether protocol/host/port changes between views. 2. Confirm the same app/account/environment. 3. In the second intended view, locate PV-A-001 without recreating it. 4. Compare fields and existing error states. 5. Recheck original view to separate visibility disagreement from loss. **Expected result:** The same permitted record is visible where sharing is explicitly promised. If sharing is required but the record is not available, record Failed (or Blocked if prerequisites prevented execution), even if local-only storage was intentional. Use Not applicable only when a reviewed requirement explicitly excludes cross-origin/device sharing; do not imply sync was tested. **If it fails / what it does not prove:** Origin-specific browser storage is one plausible reason for disagreement, not proof. Account, test/live environment, read filters, or request errors remain alternatives. **Retest:** Repeat with a new fake marker after an approved storage change; reconcile both views and PV-04. A migration itself requires PV-22. **Stop conditions:** Never clear browser storage or move real records to test this. Stop if one address is untrusted or the original view is the only remaining copy. **Why this is in the kit:** Lovable notes recorded another-origin emptiness; no second-device test or database follow-up ran. **Sources:** L1; M1. See source register below. ### PV-07 · Fresh account confirmation reaches a usable state **When:** Apps requiring email confirmation **Identity:** Fresh Client B account controlled by tester **Mode:** Conditional identity · Proposed test, not a tested repair **Prerequisites:** Authorized test environment; a controlled mailbox entered only into the app’s trusted signup flow. Confirm no invitations/messages will be sent to third parties. Know the app’s normal sign-in entry point. **Steps:** 1. Create one fresh test client using the normal signup flow. 2. Follow its confirmation once through the trusted app/mail flow. 3. Record only the destination route path and visible state, never the complete link. 4. If blank, try the established normal sign-in entry once and record that separately as a workaround. 5. Reload the intended authenticated page and check that it renders a useful signed-in or sign-in state. **Expected result:** A fresh valid confirmation completes the intended account-confirmation flow, then reaches the signed-in app or the app’s documented normal sign-in step. It must not remain blank. A displayed error is recorded as a failed or blocked confirmation attempt, not a successful confirmation; record the reason. Record normal sign-in behavior separately, without substituting it for confirmation success. **If it fails / what it does not prove:** A successful manual sign-in does not repair the confirmation flow. Session initialization, routing, and redirect settings are investigation areas; the test does not establish which is wrong. **Retest:** After a proposed fix, use another freshly controlled test account; retest confirmation, normal sign-in, and direct reload. Do not reuse an already-consumed link as if fresh. **Stop conditions:** Stop after a lockout/rate-limit message or identity uncertainty. Never disable verification or broadly wildcard production redirects to pass. **Why this is in the kit:** Lovable blank /portal was bypassed by signing in at /auth; root cause and permanent repair were not tested. **Sources:** L1; M3. See source register below. ### PV-08 · First admin is established only by an authorized owner **When:** Apps with an admin role and no existing authorized admin **Identity:** Owner; intended Admin; ordinary Client A **Mode:** Conditional identity · Proposed test, not a tested repair **Prerequisites:** Owner has explicitly chosen the initial admin identity through a trusted management path. No public role-promotion route or generic manual database-edit instructions. If an admin already exists, use established management instead. **Steps:** 1. Confirm ordinary client sign-in works and admin access is denied before setup. 2. Owner asks the maintainer for the app-specific, narrowly scoped bootstrap procedure and approves that exact action separately. 3. Only after approved setup, intended admin opens admin view. 4. Check Client A remains non-admin. 5. Have the maintainer confirm any temporary bootstrap is closed. This kit itself grants no role-change permission. **Expected result:** Only the explicitly designated admin receives access; ordinary users remain unable to promote themselves, and no reusable public bootstrap remains. **If it fails / what it does not prove:** A hidden Admin link or an owner preview pass does not prove server authorization. A correct deny for an ordinary user is not necessarily a login bug. **Retest:** Retest admin function plus PV-09 through PV-13 after any role/bootstrap change. Record Blocked with the missing prerequisite until the trusted owner process is complete. **Stop conditions:** Stop if asked to expose privileged keys, accept user-editable admin metadata, add public make-me-admin controls, or guess the intended admin identity. **Why this is in the kit:** Lovable needed one admin-assignment message; Bolt used a manual database step. Neither observation is a universal safe bootstrap recipe. **Sources:** L1; M2; O1. See source register below. ### PV-09 · Signed-out visitors cannot read private routes **When:** Apps with private client/admin pages **Identity:** Independent Signed out session **Mode:** Core permission · Proposed test, not a tested repair **Prerequisites:** PV-01 complete. Owner identifies a known protected route containing only fake test records. Public pages/records are separately listed as intentionally public. **Steps:** 1. Open a fresh signed-out session, not an owner preview. 2. Navigate directly to the owner-specified protected test route. 3. Observe whether private content appears before or after any redirect. 4. Record denial/sign-in/error state without copying content. 5. Do not scan or guess other paths. **Expected result:** Private content and private actions remain unavailable to signed-out visitors. A sign-in redirect, forbidden state, or safe not-found can be acceptable if no private content leaks. **If it fails / what it does not prove:** An invisible nav link alone is not a denial test. A later redirect does not excuse earlier visible private data. An intentionally public page is not automatically a failure. **Retest:** Repeat the exact protected route in a new signed-out session after a fix, plus the intended signed-in allow case. **Stop conditions:** Stop immediately if private or real content is exposed; notify the owner with a redacted description. Do not explore further. **Why this is in the kit:** Base44 guest was redirected from /admin; Softr signed-out redirect passed while its preview client access still failed. A guest pass does not prove client isolation. **Sources:** L1; O1. See source register below. ### PV-10 · A normal client cannot enter admin by direct route **When:** Apps with separate client and admin roles **Identity:** Client A, not Owner/Admin **Mode:** Core permission · Proposed test, not a tested repair **Prerequisites:** Independent Client A login verified by PV-01; owner-specified admin route with fake records. No need to enumerate routes or alter browser role fields. **Steps:** 1. Note whether client navigation shows admin controls. 2. Open the known admin route directly while still Client A. 3. Record whether the app denies access before showing admin data/actions. 4. Compare intended Admin allow behavior separately. Do not treat owner preview as the client check. **Expected result:** Client A cannot read admin-only data or execute admin actions by direct route; intended Admin can perform the permitted function. **If it fails / what it does not prove:** Hidden navigation plus an accessible page indicates a mismatch; rendering an admin page shell without private data may need clarification rather than proving full data exposure. **Retest:** After repair repeat Client A denial, Admin allow, and signed-out denial. Preserve method labels and exact route paths without private query strings. **Stop conditions:** If any unexpected admin data appears, stop and contain rather than testing further actions. **Why this is in the kit:** Softr preview-as client had hidden navigation but reachable sample admin pages; Lovable/Base44 later client checks denied direct access. **Sources:** L1; O1; S1. See source register below. ### PV-11 · Clients cannot read each other’s known fake records **When:** Apps whose records are private per client **Identity:** Client A and independent Client B **Mode:** Core permission · Proposed test, not a tested repair **Prerequisites:** Owner-approved test environment with PV-A-001 owned by A and PV-B-001 owned by B. Owner supplies only these known test detail routes. No arbitrary ID guessing, scanning, or real records. **Steps:** 1. Confirm A can read PV-A-001 and B can read PV-B-001. 2. In B’s independent session open the owner-supplied route for PV-A-001. 3. Repeat the inverse known test route as A. 4. Compare list and detail behavior; record only markers and allow/deny outcomes. 5. Stop at unexpected disclosure. **Expected result:** Each client can read its own permitted fake record and is denied the other’s private fake record, including direct details. A group-shared product may define a different explicit matrix. **If it fails / what it does not prove:** A blank own-record list is not a passing denial test. First prove the target record exists and its legitimate owner can read it. Preview simulation alone cannot establish independent account separation. **Retest:** Repeat both directions after a role, filter, sharing, or data-model change, and retest the legitimate own-record allow cases. **Stop conditions:** Stop on unexpected disclosure. Do not browse adjacent IDs or copy record contents. This is a bounded check of your own known fake records only. **Why this is in the kit:** Softr preview-as client could read all sample submissions; Bolt independent-client isolation remained untested. **Sources:** L1; O1; M2. See source register below. ### PV-12 · A denied cross-client write leaves the target unchanged **When:** Private apps with client edit actions; optional controlled denial test **Identity:** Owner preapproves; independent Client B; legitimate Client A rechecks **Mode:** Maintainer-assisted permission · Proposed test, not a tested repair **Prerequisites:** Only an isolated synthetic environment, a known A-owned fake record, and owner-approved reversible test marker. Define intended B denial. If no suitable safe normal UI exists, mark Blocked and ask a maintainer for a scoped allow/deny test; do not craft requests or modify IDs/permissions. **Steps:** 1. A records the target’s fake note value. 2. In B’s session open only the owner-specified known test record route. If the denied read itself reveals content, stop under PV-11. 3. If this app intentionally allows B to read but not edit, check edit controls and, only within the approved safe normal UI test, attempt the single reversible fake-note change. 4. A rereads the target and confirms the original note remains. No delete/bulk action is tested. **Expected result:** The denied action changes nothing. Hidden controls are a UI observation only; server-side write enforcement needs a maintainer’s appropriate test when the UI cannot exercise it safely. **If it fails / what it does not prove:** A success toast or zero-row response alone can mislead. Confirm target state. If B changes the note, authorization failed for this known test action; do not continue. **Retest:** After maintainer repair, repeat the same bounded write-denial plus A’s legitimate write allow and state verification. **Stop conditions:** No real records, guessed IDs, request crafting, deletes, or broad policy changes. Stop on unexpected read/write success; have the owner review the synthetic state. **Why this is in the kit:** Original portal checks did not establish complete cross-client write authorization. This is a proposed extension, not a reported lab pass. **Sources:** O1; M2. See source register below. ### PV-13 · Ordinary users have no self-promotion path **When:** Apps with privileged roles **Identity:** Client A; authorized maintainer **Mode:** Maintainer-assisted permission · Proposed test, not a tested repair **Prerequisites:** An agreed privilege matrix and independent ordinary-client session. Maintainer can inspect trusted role-enforcement logic without sharing code secrets or modifying production data. **Steps:** 1. In Client A’s normal profile/settings UI, note whether privileged role management is exposed. Do not submit role changes. 2. Ask maintainer to confirm privileges are enforced outside user-editable client state and untrusted profile metadata. 3. Ask for a specific automated/scoped denial test if available; record its method and evidence separately. No test evidence means Not checked/Blocked, not Passed. **Expected result:** Client A has no authorized self-promotion function, and maintainer-backed evidence establishes that client-editable metadata cannot grant admin. UI absence alone is only a partial observation. **If it fails / what it does not prove:** A hidden button cannot prove the backend ignores client role claims. This worksheet does not certify authorization or provide a role-escalation recipe. **Retest:** Repeat the review/test after signup, profile, role, or bootstrap changes, alongside legitimate owner-controlled management. **Stop conditions:** Stop if instructed to inject role claims, alter browser storage, expose keys, or weaken rules. Escalate to the authorized maintainer. **Why this is in the kit:** Missing first-admin paths in Lovable/Bolt must not be fixed with public or user-controlled promotion mechanisms. **Sources:** L1; O1; M2. See source register below. ### PV-14 · An authorized status change persists without expanding access **When:** Apps with staff-controlled record statuses **Identity:** Admin; Client A **Mode:** Conditional workflow · Proposed test, not a tested repair **Prerequisites:** Only fake record PV-A-001; a declared transition such as New to In review. Intended Admin and client roles already verified. **Steps:** 1. Admin opens PV-A-001 and changes its test status once through normal UI. 2. Reload Admin view and compare status. 3. Client A checks its own record if the contract permits status viewing. 4. Verify Client A has no staff-only status action. 5. If backend access exists, read-only verify the saved fake status. **Expected result:** The allowed transition saves and survives reload, appears consistently where intended, and does not grant a client a staff-only action. **If it fails / what it does not prove:** A read-only status tag is not a working admin edit control; a UI-only update is not a durable save. A client’s legitimate own-status read is not a privilege leak. **Retest:** Repeat one allowed fake transition after a repair; rerun relevant client/signed-out denial checks. **Stop conditions:** Stop if the transition sends third-party messages, bills, publishes, or affects real workflows. Use an isolated no-outbound test setup. **Why this is in the kit:** Lovable/Base44/Bolt logged persisted status changes; Softr preview admin statuses were read-only and persistence was untestable. **Sources:** L1; O1. See source register below. ### PV-15 · Declared file type/count limits behave predictably **When:** Only if the app accepts uploads **Identity:** Client A **Mode:** Conditional upload · Proposed test, not a tested repair **Prerequisites:** Owner-approved fake-data environment; app-specific allowed types, size, and count written down. Prepare small inert generated text/PDF/image fixtures with no personal data. No executables, malware, oversized stress uploads, or real tax/identity documents. **Steps:** 1. Upload one permitted inert fixture through normal UI. 2. Confirm its marker/name appears after reload. 3. If an explicit count limit exists, add permitted tiny fixtures up to it. 4. Try one additional tiny fixture once and observe the declared limit response. 5. If a harmless disallowed file type is part of your approved spec, check one only; otherwise leave type rejection untested. **Expected result:** Permitted fixtures save; declared limits have clear responses; rejected extra files do not appear as accepted records. Record only the limits actually exercised. **If it fails / what it does not prove:** A successful single upload does not prove count, size, type validation, or private file access. UI rejection alone does not validate every server-side restriction. **Retest:** After a scoped repair, repeat the same tiny fixtures and limit boundary; recheck existing fixture integrity and PV-16 if private. **Stop conditions:** Stop on unexpected charges, scanning alerts, real-document requests, or uncontrolled storage growth. Do not increase quotas or change bucket visibility to pass. **Why this is in the kit:** The prompt’s three-document limit was not tested in the Oct 29 lab. Softr uploads were blocked by its unreachable form. All limit checks here are proposed. **Sources:** L1; M4; O1. See source register below. ### PV-16 · Private file access matches the declared link model **When:** Only if uploaded files are intended private **Identity:** Client A; independent Client B; Signed out; maintainer **Mode:** Maintainer-assisted upload · Proposed test, not a tested repair **Prerequisites:** One inert A-owned fixture from PV-15. Owner/maintainer specifies whether access is authenticated or intentionally bearer/signed-link based, including intended lifetime. URLs/tokens remain private and never enter this worksheet. **Steps:** 1. Confirm A can open its own fixture. 2. Using only the owner-approved known fake fixture path in the controlled test environment, check B/signed-out behavior according to the declared access model. 3. Ask maintainer to distinguish listing, reading, and upload permissions. 4. Record outcomes and link model, not links. Do not alter token/query values or enumerate files. **Expected result:** Authenticated-only files deny unauthorized contexts. An intentionally issued bearer/signed link may work for its holder during its validity; assess that against the app’s declared privacy contract rather than calling it automatic bypass. **If it fails / what it does not prove:** Private database rows do not prove private file storage. A working signed link is not by itself proof of broken RLS, but unintended public/bearer exposure needs review. **Retest:** Repeat the same declared access cases after a storage-policy/link change; lifetime/revocation behavior needs a maintainer-planned test, not an invented observed result. **Stop conditions:** Stop if real file contents or unexpected public files appear. Never make a bucket public, share privileged keys, guess file paths, or paste signed links in reports. **Why this is in the kit:** Lab uploads/download links were checked narrowly; private-file cross-user authorization was not established. **Sources:** M4; O1; L1. See source register below. ### PV-17 · Export contains the intended fake rows and scope **When:** Only if data export exists or is required **Identity:** Intended export role; Client A/B if role-scoped **Mode:** Conditional export · Proposed test, not a tested repair **Prerequisites:** Only synthetic records with an expected marker/count/field list. Local trusted text or spreadsheet viewer. Decide whether export is owner-wide or client-only. Code ZIPs are not assumed to contain data. **Steps:** 1. Export once using the existing normal control. 2. Inspect headers and expected fake row count/markers locally. 3. Compare one numeric total and status to the app/backend fake data. 4. For client export, verify no other client marker is included. 5. Do not open untrusted formula-bearing real exports or publish the file. **Expected result:** The file downloads, contains the declared fields and intended rows, and respects the exporting identity’s scope. Missing export capability is a requirement gap, not a fabricated pass. **If it fails / what it does not prove:** A button or downloaded ZIP does not prove a usable data export. Empty downloads, wrong counts, or cross-client rows require investigation. No export timing/cost prediction follows. **Retest:** Repeat with a fresh expected marker/count after repair; keep before/after exports private and verify no duplicates or omitted rows. **Stop conditions:** Stop on any real/private cross-client content or an unexpected external destination. Do not send the export or enable broad permissions. **Why this is in the kit:** Lovable/Base44 CSV follow-ups downloaded data; Bolt CSV follow-up was not run after tokens were exhausted. **Sources:** L1; O1. See source register below. ### PV-18 · Every business claim has an approved factual source **When:** Apps with public-facing copy, prices, testimonials, or promises **Identity:** Owner/content reviewer **Mode:** Core content · Proposed test, not a tested repair **Prerequisites:** Approved source-of-truth copy, intended prices and toggle states, and clear fictional/sample labels. No assumption that generated text is factual. **Steps:** 1. Review every numeric claim, trust badge, testimonial, service promise, and price state. 2. Compare each with an approved fact or label it unsupported. 3. Check monthly/yearly toggles separately if present. 4. Confirm sample testimonials remain visibly labeled on phone width. 5. Propose a copy-only correction; do not publish or change destinations during this test. **Expected result:** All factual claims have an approved source; fictional examples are clearly labeled; no unrequested discount or unsupported promise remains. **If it fails / what it does not prove:** A visually polished page and working form do not validate claims. Invented statistics are content failures, not evidence of database unreliability. **Retest:** After approved copy changes, reread all affected states and verify navigation/forms still behave without destination changes. **Stop conditions:** Do not invent replacement proof, testimonials, client names, or contact details. Stop if testing would publish unsupported claims. **Why this is in the kit:** Softr’s fictional landing page invented a trust badge, yearly prices, and promises. No copy repair was tested. **Sources:** L1. See source register below. ### PV-19 · Phone layout and visible loading/error states remain usable **When:** Every app; exercise only relevant screens **Identity:** Owner or permitted test client **Mode:** Core experience · Proposed test, not a tested repair **Prerequisites:** Known fake-data screens; ordinary browser viewport tools or a controlled phone; no external screenshot service or real-account capture. **Steps:** 1. At approximately 375px width, navigate the main task using normal controls. 2. Check no essential button/field/error is clipped or hidden behind navigation. 3. Use keyboard navigation on desktop and verify focus remains visible. 4. Reload a key screen once; note whether loading/blank/error state resolves understandably. 5. Read punctuation and sample labels, including literal escape sequences. **Expected result:** The core task is usable at phone width; meaningful text/errors and focus are visible; a brief loading state is distinguished from a persistent blank. This is a practical spot check, not WCAG certification. **If it fails / what it does not prove:** An editor’s wide preview is not a phone test. A transient blank is not automatically an outage. Cosmetic escape text is distinct from failed logic. **Retest:** Repeat the same viewport/task after layout or copy changes; record actual width and any unresolved limitation. **Stop conditions:** Stop before repeated submissions or unintended external actions. Crop/redact any screenshots privately before sharing. **Why this is in the kit:** Lovable initial mobile captures were too wide and later retaken; brief blank admin reload was logged. Bolt printed a literal escape sequence. **Sources:** L1. See source register below. ### PV-20 · Record usage without treating units as comparable cost **When:** Builders displaying usage; optional before/after an already-approved normal task **Identity:** Owner of builder workspace **Mode:** Observation only · Proposed test, not a tested repair **Prerequisites:** Access to the current allowance display. Choose an action already planned and authorized independently; this kit does not ask you to buy, upgrade, or generate extra AI work just to measure it. **Steps:** 1. Record displayed unit, remaining/used amount, denominator, and whether daily/monthly/project-specific before the planned task. 2. Observe the one authorized task; do not repeat to force a measurement. 3. Record the same panel after it and any completion/gate state. 4. Note grants/bonus pools and disagreement between counters as unknowns. 5. If a limit appears, pause new requests and read the actual display. **Expected result:** A bounded observation identifies what that display measured for that task, with source/date and uncertainty. No cross-builder cost winner, reset promise, or unobserved completion is inferred. **If it fails / what it does not prove:** Different units, prices, daily/monthly denominators, grants, and metering prevent controlled cost comparison. A blocked request is not a failed implemented feature test. **Retest:** Only if another normal task is independently planned, repeat the same observation method. Keep each run separate; never average unlike pools or turn one observation into a forecast. **Stop conditions:** Stop on quota gates or unexpected paid-work/purchase prompts. Do not upgrade, purchase credits, schedule retries, or keep pressing fix to complete this test. **Why this is in the kit:** Bolt exhausted its observed allowance before CSV; Lovable counters disagreed. Neither establishes a general cost ranking or tested reset remedy. **Sources:** L1. See source register below. ### PV-21 · A checkpoint explains consequential scope before approval **When:** Only if a build presents an approval checkpoint **Identity:** Owner/authorized maintainer **Mode:** Observation only · Proposed test, not a tested repair **Prerequisites:** An existing checkpoint, not a deliberately created destructive task. No live changes are authorized by this worksheet. **Steps:** 1. Read the proposed action and identify affected environment/records. 2. Ask whether it changes data, access, publication, or spending. 3. Record whether a non-destructive alternative and preservation plan are explained. 4. Stop before approval unless the owner independently authorizes the exact action. 5. If it is only clarification, answer within the existing scope and record the next state separately. **Expected result:** The scope and consequence are clear before any approval. A pending or declined destructive action can be the correct outcome; do not grade automatic continuation as success. **If it fails / what it does not prove:** A permission checkpoint can be intentional. “Continue” is not a repair, safety proof, or permission for arbitrary deletion. **Retest:** If the maintainer revises the proposed action, review its exact new scope; do not carry forward a generic earlier approval. **Stop conditions:** Stop on unclear deletion, public launch, access change, or new spending. Never approve merely to finish the test. **Why this is in the kit:** Base44 paused before deleting its own test record; logged continue was redundant. No causal repair was demonstrated. **Sources:** L1; M5; M6. See source register below. ### PV-22 · Rehearse a data move without production cutover **When:** Only when moving data/storage is an explicit requirement **Identity:** Owner; authorized maintainer; Client A/B **Mode:** Maintainer-assisted migration · Proposed test, not a tested repair **Prerequisites:** A separate target test environment, source preserved, synthetic-only fixture set, explicit field/ownership mapping, expected counts/totals/files, and a written cutover/rollback plan. No live writes or destination changes by this kit. **Steps:** 1. Inventory the synthetic source dataset and separately inventory files and identity dependencies. 2. Have the maintainer prepare a scoped nonproduction rehearsal through supported tooling; do not assume code import moves records/passwords/files. 3. Compare target counts, marker fields, totals, ownership, and attachment availability. 4. Run PV-02/PV-04/PV-09/PV-11 and relevant file tests in the target. 5. Review how new writes, duplicate IDs, reconciliation, fallback, and final owner approval would be handled before any eventual cutover. Do not perform production cutover here. **Expected result:** A documented rehearsal reconciles the fake dataset and access matrix, with source intact and unresolved differences listed. Passing rehearsal does not authorize live cutover or guarantee recovery. **If it fails / what it does not prove:** A running UI with an empty target database is not a completed migration. Code import/restore does not prove data, credentials, or files moved. Correct counts alone do not prove ownership or permissions. **Retest:** After a mapping or process fix, rerun the same fixture rehearsal in a reviewed nonproduction target; preserve both evidence sets and avoid uncontrolled duplicate writes. **Stop conditions:** Stop on any proposed production move, overwrite, source deletion, credential export, unexpected real data, or automatic switch of live form/database destinations. **Why this is in the kit:** Lovable’s offered database follow-up was not run. No storage migration, cutover, or recovery was tested in this lab. **Sources:** L1; M1; M2; M4; M5; M6. See source register below. ### PV-23 · Keep project restore separate from database recovery **When:** Apps whose troubleshooting plan involves rollback or restore **Identity:** Owner; authorized maintainer **Mode:** Planning only · Proposed test, not a tested repair **Prerequisites:** Read-only access to relevant project/version documentation and an inventory of data dependencies. No restore is required by this test. **Steps:** 1. Identify what a proposed restore actually covers: code, functions, data, migrations, files, or identity. 2. Read the platform’s current warning. 3. List data changes that would remain after code restore. 4. Ask for a separate tested data-preservation/recovery plan before any future destructive step. 5. Mark recovery behavior Not checked unless an independently authorized safe rehearsal actually ran. **Expected result:** The owner understands that a code/version restore may leave current database data unchanged; no untested recovery guarantee appears in the plan. **If it fails / what it does not prove:** A version-history button or code ZIP is not proof of a restorable database backup. A code rollback can also become incompatible with current schema/data. **Retest:** Review this distinction again after schema/storage changes. Any real recovery exercise needs its own scoped environment, approval, and evidence. **Stop conditions:** Do not click restore, delete records, reset schema, clear storage, or test recovery on the only surviving source merely to mark a row passed. **Why this is in the kit:** No rollback was used in the Oct 29 tests. Lovable/Bolt official docs explicitly separate project restore from database data. **Sources:** M5; M6; L1. See source register below. ### PV-24 · Retest the repair and adjacent allow/deny behavior **When:** Any app after an independently approved repair **Identity:** Same identities as failing case; authorized reviewer **Mode:** Finish each scoped change · Proposed test, not a tested repair **Prerequisites:** A redacted failing attempt, a new identifiable version/change label, and the same synthetic fixtures or a documented equivalent. Original evidence must remain intact. **Steps:** 1. Duplicate the CSV for the new run rather than overwriting old evidence. 2. Repeat the exact failing steps with the same role/environment. 3. Repeat the nearest legitimate allow case and protected denial case. 4. Record actual outcomes, remaining errors, method, and timestamp with timezone. 5. If a test is blocked or not relevant, state why; do not convert it into Passed. 6. Review unresolved failures/unknowns with the owner before any launch decision. **Expected result:** The specific regression is demonstrably resolved or still open; adjacent function/access behavior is recorded separately. Results are self-recorded checks, not certification or permission to launch. **If it fails / what it does not prove:** An AI success message, partial screenshot, or unrelated passing path cannot close the original failure. Repeated failure suggests stopping the repair loop and reviewing evidence, not spending indefinitely. **Retest:** Use a new run ID for every material repair/version. Preserve before/after evidence and re-open failures after relevant changes. **Stop conditions:** Stop on exposure, data loss, unclear authorization, quota gate, or new consequential scope. This kit makes no publication/payment decision. **Why this is in the kit:** Softr claimed groups were reactivated, but settings contradicted it after reload. Observed behavior must control the result. **Sources:** L1; O1. See source register below. ## Reading your results without overstating them - **Failed:** the observed behavior disagrees with the declared expectation for this specific test. Preserve the redacted evidence; do not guess a root cause or repeatedly press fix. - **Blocked:** prerequisites or a safe execution method are absent. Examples: no backend read access; preview cannot independently sign clients in; quota gate prevents a planned export. Blocked is not Passed or proof that the feature is impossible. - **Passed:** this particular action/identity/environment behaved as expected at this time. A same-session save pass does not establish a backend, cross-device sync, private files, or authorization. A read deny does not establish write denial. A narrow UI pass is not a full server-side review. - **Not applicable:** an explicitly reviewed requirement makes this test irrelevant. Do not hide a required but unavailable feature here. - **Not checked:** no usable execution evidence. Documentation, plans, fixture preparation, and a generated fix message leave runtime behavior Not checked. For the original Oct 29 evidence: Lovable’s invoice storage follow-up was **suggested, not run**; another-device persistence and migration were untested. Bolt’s independent-client isolation and later CSV export were not run. Softr’s client result used **Preview as**, not independent login. A blank confirmation destination was bypassed with normal sign-in, not permanently repaired. Base44’s “continue” was redundant, not a demonstrated repair. These limits remain even if a reader later records their own successful tests. ## Source register and provenance Official references were opened during Nov 7, 2026 preparation. They support testing principles and platform concepts, **not a claim that these 24 proposed procedures or any repair was lab-tested**. Cross-platform apps need their own current official documentation; do not paste vendor-specific SQL or configuration into another builder. - **L1 — PrototypeVerdict: original Oct 29 failure records and method** - https://prototypeverdict.com/guides/what-actually-broke/ - https://prototypeverdict.com/how-we-test/ - **M1 — MDN: localStorage origin and persistence** - https://developer.mozilla.org/en-US/docs/Web/API/Window/localStorage - **M2 — Supabase: RLS, grants, read/write denial and policy tests** - https://supabase.com/docs/guides/database/postgres/row-level-security - **M3 — Supabase: intended redirects and authentication error handling** - https://supabase.com/docs/guides/auth/redirect-urls - **M4 — Supabase: storage operation access controls** - https://supabase.com/docs/guides/storage/security/access-control - **M5 — Lovable: version history does not restore database data** - https://docs.lovable.dev/features/projects/history - **M6 — Bolt: project restore does not change current databases** - https://support.bolt.new/managing-projects/rollback-backup - **O1 — OWASP Authorization Cheat Sheet: least privilege, every-request checks, tests, safe denial** - https://cheatsheetseries.owasp.org/cheatsheets/Authorization_Cheat_Sheet.html - **S1 — Softr: page/action/record permission layers** - https://docs.softr.io/user-groups-and-permissions - https://docs.softr.io/user-groups-and-permissions/global-data-restrictions Key boundaries: OWASP calls for least privilege, authorization on every request, safe denial/error handling, and dedicated allow/deny tests; it does not certify this kit or endorse its completion as a security audit. Supabase storage docs distinguish operations and warn privileged keys bypass controls; this kit provides no bypass/SQL recipe. MDN explains origin-scoped browser storage, not recoverability. Lovable/Bolt history docs distinguish project restore from database recovery. Softr current docs distinguish visibility/restrictions; no higher-plan cure is claimed. Original-lab provenance: sanitized cases were prepared from the contemporaneous Oct 29 Lovable Pro, Bolt Free, Base44 Builder, and Softr Basic test notes. Public case catalog and method above are reader-accessible companions; private raw notes and account/project identifiers are not part of this download. ## Before an owner decides what happens next Review unresolved failures, blocked prerequisites, untested private-data boundaries, and data-preservation needs with the authorized maintainer. Record what was actually checked and what remains unknown. Any publication, consequential role/data change, migration/cutover, third-party communication, or purchase is a separate owner-controlled decision—not authorized by passing a worksheet. **No score, certificate, launch clearance, reliability ranking, measured cost comparison, or guarantee is produced.** The useful output is an honest record of specific checks and a focused retest plan.
Evidence worksheet CSV. Every result starts Not checked.
"kit_version","prepared_date","original_lab_date","test_basis","test_id","title","group","applicability","recommended_identity","prerequisites","steps","expected_result","failure_interpretation","retest_procedure","stop_conditions","lab_connection","source_ids","source_urls","scope_warning","run_id","app_label","environment_label","app_version","identity_label","evidence_result","observed_result_redacted","evidence_reference_redacted","tested_at_with_timezone","retest_run_id","retest_result","retest_notes_redacted" "2026-11-07.2","2026-11-07","2026-10-29","Proposed test; not a lab-tested repair; no execution claimed","PV-01","Prove which environment and identity you are testing","Start here","Every app","Owner; Client A; Client B; Signed out","You own the app or have explicit permission. Use a private test copy containing only fake records. Have three distinct controlled accounts if roles exist, and separate browser profiles or otherwise independent sessions. Never reuse the owner session as a client.","1. Record a nonsecret app/environment label and version. 2. In each separate session, read the visible signed-in identity and intended role. 3. Mark the signed-out session separately. 4. Verify all sessions target the same intended environment; note protocol/host/port privately without token-bearing URLs. 5. If using Preview as, record that method instead of claiming independent login.","Each context has an unambiguous identity, role, session, and environment. Owner/admin is not mistaken for Client A/B; preview simulation is labeled.","Wrong account/environment or an owner session can create false passes and false missing-data reports. A hidden role label or preview switch is not proof of independent identity.","Repeat after any account switch, domain/environment change, or role update. Do not reuse downstream evidence from a mismatched context.","Stop if you cannot establish authorization or see real records. Do not sign into accounts you do not control.","Softr checks used Preview as; Lovable/Base44 later used separate client accounts. These methods are not equivalent.","L1; O1","https://prototypeverdict.com/guides/what-actually-broke/ | https://prototypeverdict.com/how-we-test/ | https://cheatsheetseries.owasp.org/cheatsheets/Authorization_Cheat_Sheet.html","Only own authorized apps and synthetic data. Stop on exposure or consequential scope. No destructive clears, security bypass, purchases, production cutover, certification, or launch clearance.","","","","","","Not checked","","","","","Not checked","" "2026-11-07.2","2026-11-07","2026-10-29","Proposed test; not a lab-tested repair; no execution claimed","PV-02","A completed save survives a reload in the same session","Core data","Apps that save records","Client A or intended single user","PV-01 complete. Pick a fake record type. Use marker PV-A-001 and note its expected field values locally. Keep any original browser-local records intact.","1. Create one fake record through the normal UI and wait for the completion state. 2. Open its detail view and note marker and key values. 3. Reload the same page in the same browser/session/origin. 4. Reopen the record and compare the values. 5. Record any visible error rather than repeatedly submitting.","Exactly the intended fake record appears with unchanged values after reload; the UI does not merely retain unsaved form text or show a duplicate.","Missing or changed data could indicate a failed save, failed read, wrong environment/account, or storage behavior. A pass proves only same-session persistence, not a database or device sync.","After a focused repair, use a new fake marker and repeat. Preserve the failing record and note both attempt IDs.","Stop on unexpected real data, duplicate submissions, or a write error. Do not clear site data or reseed the database.","Lovable invoice data passed refresh yet another-origin preview was empty; Base44 database records also passed refresh.","L1; M1","https://prototypeverdict.com/guides/what-actually-broke/ | https://prototypeverdict.com/how-we-test/ | https://developer.mozilla.org/en-US/docs/Web/API/Window/localStorage","Only own authorized apps and synthetic data. Stop on exposure or consequential scope. No destructive clears, security bypass, purchases, production cutover, certification, or launch clearance.","","","","","","Not checked","","","","","Not checked","" "2026-11-07.2","2026-11-07","2026-10-29","Proposed test; not a lab-tested repair; no execution claimed","PV-03","A later independent session can read the intended saved record","Core data","Apps promising account-backed persistence between sessions","Client A in a second controlled session","PV-02 complete. Shared/account persistence is part of your declared requirement. Keep the first working session open; do not close the only private/local-storage copy.","1. Open an independent controlled browser profile/session at the same intended environment. 2. Sign into Client A through the normal flow. 3. Locate PV-A-001 without recreating it. 4. Compare its saved fields. 5. Note whether the first session still sees it. Do not expect unrelated Client B to see a private A record.","The same authorized user reads the same record and values in the second session if that is the app contract.","Failure does not prove deletion: compare environment/account first, then storage, read errors, and access rules. Browser-only storage may be an intentional design but does not meet account-shared requirements.","Use the same saved marker after the maintainer explains the fix; repeat both original and independent-session reads.","Do not log out/clear/close the sole surviving original session to run this. Stop if sign-in is locked or repeated attempts fail.","Cross-session/device sharing was not tested in the Lovable storage case; this is a proposed test, not a reported lab pass.","L1; M1","https://prototypeverdict.com/guides/what-actually-broke/ | https://prototypeverdict.com/how-we-test/ | https://developer.mozilla.org/en-US/docs/Web/API/Window/localStorage","Only own authorized apps and synthetic data. Stop on exposure or consequential scope. No destructive clears, security bypass, purchases, production cutover, certification, or launch clearance.","","","","","","Not checked","","","","","Not checked","" "2026-11-07.2","2026-11-07","2026-10-29","Proposed test; not a lab-tested repair; no execution claimed","PV-04","Confirm the saved row exists in the intended backend","Core data","Apps claiming database-backed saves","Owner/authorized backend maintainer; Client A performs UI save","Read-only access to the intended data-management interface, and an agreed record/table mapping. If unavailable, record Blocked, not Passed. Never reveal backend keys or grant wider access to inspect.","1. Save one uniquely marked fake record through the normal app. 2. In the authorized read-only backend view, find that exact marker in the intended environment. 3. Compare key saved fields and intended ownership mapping. 4. Note only a nonsecret table label, marker, count, and outcome in the worksheet. Do not export user tables or auth records.","The corresponding backend row exists once, has the expected values, and is linked to the correct test owner. State clearly if the app intentionally has no backend.","A UI success toast without a backend row is not evidence of a durable database write. Check storage design, request completion, wrong environment, or rejected write; do not assume a cause.","With a new marker, repeat UI save plus backend confirmation and then PV-02/PV-03. Keep old evidence separate.","Stop if you would need to create a privileged token, change grants, expose secrets, or inspect real customer records.","Base44 invoice tables were observed; Lovable invoice notes suspected local-only storage. Backend existence must be checked per app, not inferred from builder branding.","L1; M2","https://prototypeverdict.com/guides/what-actually-broke/ | https://prototypeverdict.com/how-we-test/ | https://supabase.com/docs/guides/database/postgres/row-level-security","Only own authorized apps and synthetic data. Stop on exposure or consequential scope. No destructive clears, security bypass, purchases, production cutover, certification, or launch clearance.","","","","","","Not checked","","","","","Not checked","" "2026-11-07.2","2026-11-07","2026-10-29","Proposed test; not a lab-tested repair; no execution claimed","PV-05","Distinguish an empty dataset from a failed read","Core data","Apps displaying record lists or dashboards","Client A; optional authorized maintainer","A known fake backend row from PV-04, or an explicitly declared empty dataset. No need to invent a failure, disable access, disconnect a production backend, or edit credentials.","1. Open the known fake record/list normally. 2. If empty or stuck, note whether the UI says no records, loading, or error. 3. An authorized maintainer may inspect the existing request status and redacted error category using trusted logs/devtools; do not export HAR files, request headers, cookies, tokens, or response contents. 4. Compare backend row existence with the user-visible read. If no failure occurs, record only the successful-read behavior; error-state behavior remains untested.","Known permitted fake records load; if an existing read error occurs, the app distinguishes it from a genuinely empty result and exposes no sensitive debugging output.","A present row plus an empty view may involve filtering, identity, permissions, environment, or a request failure. Empty results can also be the intended outcome of a deny policy; a 500 does not prove recursion.","After repair, repeat the exact read with the same identity and known marker. A maintainer can plan a separate controlled failure-state test, but do not claim one occurred here.","Stop on unexpected access or sensitive debug output. Do not broaden RLS or use elevated credentials in a browser to make the list populate.","Bolt login failed alongside a recursive-policy 500; Lovable another-origin dashboard was empty. The symptom alone does not identify the failed layer.","L1; M2; O1","https://prototypeverdict.com/guides/what-actually-broke/ | https://prototypeverdict.com/how-we-test/ | https://supabase.com/docs/guides/database/postgres/row-level-security | https://cheatsheetseries.owasp.org/cheatsheets/Authorization_Cheat_Sheet.html","Only own authorized apps and synthetic data. Stop on exposure or consequential scope. No destructive clears, security bypass, purchases, production cutover, certification, or launch clearance.","","","","","","Not checked","","","","","Not checked","" "2026-11-07.2","2026-11-07","2026-10-29","Proposed test; not a lab-tested repair; no execution claimed","PV-06","Compare intended shared data across origins or devices","Conditional data","Only if cross-device or alternate-domain access is an intended feature","Same Client A; independent controlled browser/device","PV-02/PV-04 if available; a documented expectation of shared storage; both target addresses belong to your authorized app. Preserve the original local-data session. Do not copy token-bearing preview links.","1. Record whether protocol/host/port changes between views. 2. Confirm the same app/account/environment. 3. In the second intended view, locate PV-A-001 without recreating it. 4. Compare fields and existing error states. 5. Recheck original view to separate visibility disagreement from loss.","The same permitted record is visible where sharing is explicitly promised. If sharing is required but the record is not available, record Failed (or Blocked if prerequisites prevented execution), even if local-only storage was intentional. Use Not applicable only when a reviewed requirement explicitly excludes cross-origin/device sharing; do not imply sync was tested.","Origin-specific browser storage is one plausible reason for disagreement, not proof. Account, test/live environment, read filters, or request errors remain alternatives.","Repeat with a new fake marker after an approved storage change; reconcile both views and PV-04. A migration itself requires PV-22.","Never clear browser storage or move real records to test this. Stop if one address is untrusted or the original view is the only remaining copy.","Lovable notes recorded another-origin emptiness; no second-device test or database follow-up ran.","L1; M1","https://prototypeverdict.com/guides/what-actually-broke/ | https://prototypeverdict.com/how-we-test/ | https://developer.mozilla.org/en-US/docs/Web/API/Window/localStorage","Only own authorized apps and synthetic data. Stop on exposure or consequential scope. No destructive clears, security bypass, purchases, production cutover, certification, or launch clearance.","","","","","","Not checked","","","","","Not checked","" "2026-11-07.2","2026-11-07","2026-10-29","Proposed test; not a lab-tested repair; no execution claimed","PV-07","Fresh account confirmation reaches a usable state","Conditional identity","Apps requiring email confirmation","Fresh Client B account controlled by tester","Authorized test environment; a controlled mailbox entered only into the app’s trusted signup flow. Confirm no invitations/messages will be sent to third parties. Know the app’s normal sign-in entry point.","1. Create one fresh test client using the normal signup flow. 2. Follow its confirmation once through the trusted app/mail flow. 3. Record only the destination route path and visible state, never the complete link. 4. If blank, try the established normal sign-in entry once and record that separately as a workaround. 5. Reload the intended authenticated page and check that it renders a useful signed-in or sign-in state.","A fresh valid confirmation completes the intended account-confirmation flow, then reaches the signed-in app or the app’s documented normal sign-in step. It must not remain blank. A displayed error is recorded as a failed or blocked confirmation attempt, not a successful confirmation; record the reason. Record normal sign-in behavior separately, without substituting it for confirmation success.","A successful manual sign-in does not repair the confirmation flow. Session initialization, routing, and redirect settings are investigation areas; the test does not establish which is wrong.","After a proposed fix, use another freshly controlled test account; retest confirmation, normal sign-in, and direct reload. Do not reuse an already-consumed link as if fresh.","Stop after a lockout/rate-limit message or identity uncertainty. Never disable verification or broadly wildcard production redirects to pass.","Lovable blank /portal was bypassed by signing in at /auth; root cause and permanent repair were not tested.","L1; M3","https://prototypeverdict.com/guides/what-actually-broke/ | https://prototypeverdict.com/how-we-test/ | https://supabase.com/docs/guides/auth/redirect-urls","Only own authorized apps and synthetic data. Stop on exposure or consequential scope. No destructive clears, security bypass, purchases, production cutover, certification, or launch clearance.","","","","","","Not checked","","","","","Not checked","" "2026-11-07.2","2026-11-07","2026-10-29","Proposed test; not a lab-tested repair; no execution claimed","PV-08","First admin is established only by an authorized owner","Conditional identity","Apps with an admin role and no existing authorized admin","Owner; intended Admin; ordinary Client A","Owner has explicitly chosen the initial admin identity through a trusted management path. No public role-promotion route or generic manual database-edit instructions. If an admin already exists, use established management instead.","1. Confirm ordinary client sign-in works and admin access is denied before setup. 2. Owner asks the maintainer for the app-specific, narrowly scoped bootstrap procedure and approves that exact action separately. 3. Only after approved setup, intended admin opens admin view. 4. Check Client A remains non-admin. 5. Have the maintainer confirm any temporary bootstrap is closed. This kit itself grants no role-change permission.","Only the explicitly designated admin receives access; ordinary users remain unable to promote themselves, and no reusable public bootstrap remains.","A hidden Admin link or an owner preview pass does not prove server authorization. A correct deny for an ordinary user is not necessarily a login bug.","Retest admin function plus PV-09 through PV-13 after any role/bootstrap change. Record Blocked with the missing prerequisite until the trusted owner process is complete.","Stop if asked to expose privileged keys, accept user-editable admin metadata, add public make-me-admin controls, or guess the intended admin identity.","Lovable needed one admin-assignment message; Bolt used a manual database step. Neither observation is a universal safe bootstrap recipe.","L1; M2; O1","https://prototypeverdict.com/guides/what-actually-broke/ | https://prototypeverdict.com/how-we-test/ | https://supabase.com/docs/guides/database/postgres/row-level-security | https://cheatsheetseries.owasp.org/cheatsheets/Authorization_Cheat_Sheet.html","Only own authorized apps and synthetic data. Stop on exposure or consequential scope. No destructive clears, security bypass, purchases, production cutover, certification, or launch clearance.","","","","","","Not checked","","","","","Not checked","" "2026-11-07.2","2026-11-07","2026-10-29","Proposed test; not a lab-tested repair; no execution claimed","PV-09","Signed-out visitors cannot read private routes","Core permission","Apps with private client/admin pages","Independent Signed out session","PV-01 complete. Owner identifies a known protected route containing only fake test records. Public pages/records are separately listed as intentionally public.","1. Open a fresh signed-out session, not an owner preview. 2. Navigate directly to the owner-specified protected test route. 3. Observe whether private content appears before or after any redirect. 4. Record denial/sign-in/error state without copying content. 5. Do not scan or guess other paths.","Private content and private actions remain unavailable to signed-out visitors. A sign-in redirect, forbidden state, or safe not-found can be acceptable if no private content leaks.","An invisible nav link alone is not a denial test. A later redirect does not excuse earlier visible private data. An intentionally public page is not automatically a failure.","Repeat the exact protected route in a new signed-out session after a fix, plus the intended signed-in allow case.","Stop immediately if private or real content is exposed; notify the owner with a redacted description. Do not explore further.","Base44 guest was redirected from /admin; Softr signed-out redirect passed while its preview client access still failed. A guest pass does not prove client isolation.","L1; O1","https://prototypeverdict.com/guides/what-actually-broke/ | https://prototypeverdict.com/how-we-test/ | https://cheatsheetseries.owasp.org/cheatsheets/Authorization_Cheat_Sheet.html","Only own authorized apps and synthetic data. Stop on exposure or consequential scope. No destructive clears, security bypass, purchases, production cutover, certification, or launch clearance.","","","","","","Not checked","","","","","Not checked","" "2026-11-07.2","2026-11-07","2026-10-29","Proposed test; not a lab-tested repair; no execution claimed","PV-10","A normal client cannot enter admin by direct route","Core permission","Apps with separate client and admin roles","Client A, not Owner/Admin","Independent Client A login verified by PV-01; owner-specified admin route with fake records. No need to enumerate routes or alter browser role fields.","1. Note whether client navigation shows admin controls. 2. Open the known admin route directly while still Client A. 3. Record whether the app denies access before showing admin data/actions. 4. Compare intended Admin allow behavior separately. Do not treat owner preview as the client check.","Client A cannot read admin-only data or execute admin actions by direct route; intended Admin can perform the permitted function.","Hidden navigation plus an accessible page indicates a mismatch; rendering an admin page shell without private data may need clarification rather than proving full data exposure.","After repair repeat Client A denial, Admin allow, and signed-out denial. Preserve method labels and exact route paths without private query strings.","If any unexpected admin data appears, stop and contain rather than testing further actions.","Softr preview-as client had hidden navigation but reachable sample admin pages; Lovable/Base44 later client checks denied direct access.","L1; O1; S1","https://prototypeverdict.com/guides/what-actually-broke/ | https://prototypeverdict.com/how-we-test/ | https://cheatsheetseries.owasp.org/cheatsheets/Authorization_Cheat_Sheet.html | https://docs.softr.io/user-groups-and-permissions | https://docs.softr.io/user-groups-and-permissions/global-data-restrictions","Only own authorized apps and synthetic data. Stop on exposure or consequential scope. No destructive clears, security bypass, purchases, production cutover, certification, or launch clearance.","","","","","","Not checked","","","","","Not checked","" "2026-11-07.2","2026-11-07","2026-10-29","Proposed test; not a lab-tested repair; no execution claimed","PV-11","Clients cannot read each other’s known fake records","Core permission","Apps whose records are private per client","Client A and independent Client B","Owner-approved test environment with PV-A-001 owned by A and PV-B-001 owned by B. Owner supplies only these known test detail routes. No arbitrary ID guessing, scanning, or real records.","1. Confirm A can read PV-A-001 and B can read PV-B-001. 2. In B’s independent session open the owner-supplied route for PV-A-001. 3. Repeat the inverse known test route as A. 4. Compare list and detail behavior; record only markers and allow/deny outcomes. 5. Stop at unexpected disclosure.","Each client can read its own permitted fake record and is denied the other’s private fake record, including direct details. A group-shared product may define a different explicit matrix.","A blank own-record list is not a passing denial test. First prove the target record exists and its legitimate owner can read it. Preview simulation alone cannot establish independent account separation.","Repeat both directions after a role, filter, sharing, or data-model change, and retest the legitimate own-record allow cases.","Stop on unexpected disclosure. Do not browse adjacent IDs or copy record contents. This is a bounded check of your own known fake records only.","Softr preview-as client could read all sample submissions; Bolt independent-client isolation remained untested.","L1; O1; M2","https://prototypeverdict.com/guides/what-actually-broke/ | https://prototypeverdict.com/how-we-test/ | https://cheatsheetseries.owasp.org/cheatsheets/Authorization_Cheat_Sheet.html | https://supabase.com/docs/guides/database/postgres/row-level-security","Only own authorized apps and synthetic data. Stop on exposure or consequential scope. No destructive clears, security bypass, purchases, production cutover, certification, or launch clearance.","","","","","","Not checked","","","","","Not checked","" "2026-11-07.2","2026-11-07","2026-10-29","Proposed test; not a lab-tested repair; no execution claimed","PV-12","A denied cross-client write leaves the target unchanged","Maintainer-assisted permission","Private apps with client edit actions; optional controlled denial test","Owner preapproves; independent Client B; legitimate Client A rechecks","Only an isolated synthetic environment, a known A-owned fake record, and owner-approved reversible test marker. Define intended B denial. If no suitable safe normal UI exists, mark Blocked and ask a maintainer for a scoped allow/deny test; do not craft requests or modify IDs/permissions.","1. A records the target’s fake note value. 2. In B’s session open only the owner-specified known test record route. If the denied read itself reveals content, stop under PV-11. 3. If this app intentionally allows B to read but not edit, check edit controls and, only within the approved safe normal UI test, attempt the single reversible fake-note change. 4. A rereads the target and confirms the original note remains. No delete/bulk action is tested.","The denied action changes nothing. Hidden controls are a UI observation only; server-side write enforcement needs a maintainer’s appropriate test when the UI cannot exercise it safely.","A success toast or zero-row response alone can mislead. Confirm target state. If B changes the note, authorization failed for this known test action; do not continue.","After maintainer repair, repeat the same bounded write-denial plus A’s legitimate write allow and state verification.","No real records, guessed IDs, request crafting, deletes, or broad policy changes. Stop on unexpected read/write success; have the owner review the synthetic state.","Original portal checks did not establish complete cross-client write authorization. This is a proposed extension, not a reported lab pass.","O1; M2","https://cheatsheetseries.owasp.org/cheatsheets/Authorization_Cheat_Sheet.html | https://supabase.com/docs/guides/database/postgres/row-level-security","Only own authorized apps and synthetic data. Stop on exposure or consequential scope. No destructive clears, security bypass, purchases, production cutover, certification, or launch clearance.","","","","","","Not checked","","","","","Not checked","" "2026-11-07.2","2026-11-07","2026-10-29","Proposed test; not a lab-tested repair; no execution claimed","PV-13","Ordinary users have no self-promotion path","Maintainer-assisted permission","Apps with privileged roles","Client A; authorized maintainer","An agreed privilege matrix and independent ordinary-client session. Maintainer can inspect trusted role-enforcement logic without sharing code secrets or modifying production data.","1. In Client A’s normal profile/settings UI, note whether privileged role management is exposed. Do not submit role changes. 2. Ask maintainer to confirm privileges are enforced outside user-editable client state and untrusted profile metadata. 3. Ask for a specific automated/scoped denial test if available; record its method and evidence separately. No test evidence means Not checked/Blocked, not Passed.","Client A has no authorized self-promotion function, and maintainer-backed evidence establishes that client-editable metadata cannot grant admin. UI absence alone is only a partial observation.","A hidden button cannot prove the backend ignores client role claims. This worksheet does not certify authorization or provide a role-escalation recipe.","Repeat the review/test after signup, profile, role, or bootstrap changes, alongside legitimate owner-controlled management.","Stop if instructed to inject role claims, alter browser storage, expose keys, or weaken rules. Escalate to the authorized maintainer.","Missing first-admin paths in Lovable/Bolt must not be fixed with public or user-controlled promotion mechanisms.","L1; O1; M2","https://prototypeverdict.com/guides/what-actually-broke/ | https://prototypeverdict.com/how-we-test/ | https://cheatsheetseries.owasp.org/cheatsheets/Authorization_Cheat_Sheet.html | https://supabase.com/docs/guides/database/postgres/row-level-security","Only own authorized apps and synthetic data. Stop on exposure or consequential scope. No destructive clears, security bypass, purchases, production cutover, certification, or launch clearance.","","","","","","Not checked","","","","","Not checked","" "2026-11-07.2","2026-11-07","2026-10-29","Proposed test; not a lab-tested repair; no execution claimed","PV-14","An authorized status change persists without expanding access","Conditional workflow","Apps with staff-controlled record statuses","Admin; Client A","Only fake record PV-A-001; a declared transition such as New to In review. Intended Admin and client roles already verified.","1. Admin opens PV-A-001 and changes its test status once through normal UI. 2. Reload Admin view and compare status. 3. Client A checks its own record if the contract permits status viewing. 4. Verify Client A has no staff-only status action. 5. If backend access exists, read-only verify the saved fake status.","The allowed transition saves and survives reload, appears consistently where intended, and does not grant a client a staff-only action.","A read-only status tag is not a working admin edit control; a UI-only update is not a durable save. A client’s legitimate own-status read is not a privilege leak.","Repeat one allowed fake transition after a repair; rerun relevant client/signed-out denial checks.","Stop if the transition sends third-party messages, bills, publishes, or affects real workflows. Use an isolated no-outbound test setup.","Lovable/Base44/Bolt logged persisted status changes; Softr preview admin statuses were read-only and persistence was untestable.","L1; O1","https://prototypeverdict.com/guides/what-actually-broke/ | https://prototypeverdict.com/how-we-test/ | https://cheatsheetseries.owasp.org/cheatsheets/Authorization_Cheat_Sheet.html","Only own authorized apps and synthetic data. Stop on exposure or consequential scope. No destructive clears, security bypass, purchases, production cutover, certification, or launch clearance.","","","","","","Not checked","","","","","Not checked","" "2026-11-07.2","2026-11-07","2026-10-29","Proposed test; not a lab-tested repair; no execution claimed","PV-15","Declared file type/count limits behave predictably","Conditional upload","Only if the app accepts uploads","Client A","Owner-approved fake-data environment; app-specific allowed types, size, and count written down. Prepare small inert generated text/PDF/image fixtures with no personal data. No executables, malware, oversized stress uploads, or real tax/identity documents.","1. Upload one permitted inert fixture through normal UI. 2. Confirm its marker/name appears after reload. 3. If an explicit count limit exists, add permitted tiny fixtures up to it. 4. Try one additional tiny fixture once and observe the declared limit response. 5. If a harmless disallowed file type is part of your approved spec, check one only; otherwise leave type rejection untested.","Permitted fixtures save; declared limits have clear responses; rejected extra files do not appear as accepted records. Record only the limits actually exercised.","A successful single upload does not prove count, size, type validation, or private file access. UI rejection alone does not validate every server-side restriction.","After a scoped repair, repeat the same tiny fixtures and limit boundary; recheck existing fixture integrity and PV-16 if private.","Stop on unexpected charges, scanning alerts, real-document requests, or uncontrolled storage growth. Do not increase quotas or change bucket visibility to pass.","The prompt’s three-document limit was not tested in the Oct 29 lab. Softr uploads were blocked by its unreachable form. All limit checks here are proposed.","L1; M4; O1","https://prototypeverdict.com/guides/what-actually-broke/ | https://prototypeverdict.com/how-we-test/ | https://supabase.com/docs/guides/storage/security/access-control | https://cheatsheetseries.owasp.org/cheatsheets/Authorization_Cheat_Sheet.html","Only own authorized apps and synthetic data. Stop on exposure or consequential scope. No destructive clears, security bypass, purchases, production cutover, certification, or launch clearance.","","","","","","Not checked","","","","","Not checked","" "2026-11-07.2","2026-11-07","2026-10-29","Proposed test; not a lab-tested repair; no execution claimed","PV-16","Private file access matches the declared link model","Maintainer-assisted upload","Only if uploaded files are intended private","Client A; independent Client B; Signed out; maintainer","One inert A-owned fixture from PV-15. Owner/maintainer specifies whether access is authenticated or intentionally bearer/signed-link based, including intended lifetime. URLs/tokens remain private and never enter this worksheet.","1. Confirm A can open its own fixture. 2. Using only the owner-approved known fake fixture path in the controlled test environment, check B/signed-out behavior according to the declared access model. 3. Ask maintainer to distinguish listing, reading, and upload permissions. 4. Record outcomes and link model, not links. Do not alter token/query values or enumerate files.","Authenticated-only files deny unauthorized contexts. An intentionally issued bearer/signed link may work for its holder during its validity; assess that against the app’s declared privacy contract rather than calling it automatic bypass.","Private database rows do not prove private file storage. A working signed link is not by itself proof of broken RLS, but unintended public/bearer exposure needs review.","Repeat the same declared access cases after a storage-policy/link change; lifetime/revocation behavior needs a maintainer-planned test, not an invented observed result.","Stop if real file contents or unexpected public files appear. Never make a bucket public, share privileged keys, guess file paths, or paste signed links in reports.","Lab uploads/download links were checked narrowly; private-file cross-user authorization was not established.","M4; O1; L1","https://supabase.com/docs/guides/storage/security/access-control | https://cheatsheetseries.owasp.org/cheatsheets/Authorization_Cheat_Sheet.html | https://prototypeverdict.com/guides/what-actually-broke/ | https://prototypeverdict.com/how-we-test/","Only own authorized apps and synthetic data. Stop on exposure or consequential scope. No destructive clears, security bypass, purchases, production cutover, certification, or launch clearance.","","","","","","Not checked","","","","","Not checked","" "2026-11-07.2","2026-11-07","2026-10-29","Proposed test; not a lab-tested repair; no execution claimed","PV-17","Export contains the intended fake rows and scope","Conditional export","Only if data export exists or is required","Intended export role; Client A/B if role-scoped","Only synthetic records with an expected marker/count/field list. Local trusted text or spreadsheet viewer. Decide whether export is owner-wide or client-only. Code ZIPs are not assumed to contain data.","1. Export once using the existing normal control. 2. Inspect headers and expected fake row count/markers locally. 3. Compare one numeric total and status to the app/backend fake data. 4. For client export, verify no other client marker is included. 5. Do not open untrusted formula-bearing real exports or publish the file.","The file downloads, contains the declared fields and intended rows, and respects the exporting identity’s scope. Missing export capability is a requirement gap, not a fabricated pass.","A button or downloaded ZIP does not prove a usable data export. Empty downloads, wrong counts, or cross-client rows require investigation. No export timing/cost prediction follows.","Repeat with a fresh expected marker/count after repair; keep before/after exports private and verify no duplicates or omitted rows.","Stop on any real/private cross-client content or an unexpected external destination. Do not send the export or enable broad permissions.","Lovable/Base44 CSV follow-ups downloaded data; Bolt CSV follow-up was not run after tokens were exhausted.","L1; O1","https://prototypeverdict.com/guides/what-actually-broke/ | https://prototypeverdict.com/how-we-test/ | https://cheatsheetseries.owasp.org/cheatsheets/Authorization_Cheat_Sheet.html","Only own authorized apps and synthetic data. Stop on exposure or consequential scope. No destructive clears, security bypass, purchases, production cutover, certification, or launch clearance.","","","","","","Not checked","","","","","Not checked","" "2026-11-07.2","2026-11-07","2026-10-29","Proposed test; not a lab-tested repair; no execution claimed","PV-18","Every business claim has an approved factual source","Core content","Apps with public-facing copy, prices, testimonials, or promises","Owner/content reviewer","Approved source-of-truth copy, intended prices and toggle states, and clear fictional/sample labels. No assumption that generated text is factual.","1. Review every numeric claim, trust badge, testimonial, service promise, and price state. 2. Compare each with an approved fact or label it unsupported. 3. Check monthly/yearly toggles separately if present. 4. Confirm sample testimonials remain visibly labeled on phone width. 5. Propose a copy-only correction; do not publish or change destinations during this test.","All factual claims have an approved source; fictional examples are clearly labeled; no unrequested discount or unsupported promise remains.","A visually polished page and working form do not validate claims. Invented statistics are content failures, not evidence of database unreliability.","After approved copy changes, reread all affected states and verify navigation/forms still behave without destination changes.","Do not invent replacement proof, testimonials, client names, or contact details. Stop if testing would publish unsupported claims.","Softr’s fictional landing page invented a trust badge, yearly prices, and promises. No copy repair was tested.","L1","https://prototypeverdict.com/guides/what-actually-broke/ | https://prototypeverdict.com/how-we-test/","Only own authorized apps and synthetic data. Stop on exposure or consequential scope. No destructive clears, security bypass, purchases, production cutover, certification, or launch clearance.","","","","","","Not checked","","","","","Not checked","" "2026-11-07.2","2026-11-07","2026-10-29","Proposed test; not a lab-tested repair; no execution claimed","PV-19","Phone layout and visible loading/error states remain usable","Core experience","Every app; exercise only relevant screens","Owner or permitted test client","Known fake-data screens; ordinary browser viewport tools or a controlled phone; no external screenshot service or real-account capture.","1. At approximately 375px width, navigate the main task using normal controls. 2. Check no essential button/field/error is clipped or hidden behind navigation. 3. Use keyboard navigation on desktop and verify focus remains visible. 4. Reload a key screen once; note whether loading/blank/error state resolves understandably. 5. Read punctuation and sample labels, including literal escape sequences.","The core task is usable at phone width; meaningful text/errors and focus are visible; a brief loading state is distinguished from a persistent blank. This is a practical spot check, not WCAG certification.","An editor’s wide preview is not a phone test. A transient blank is not automatically an outage. Cosmetic escape text is distinct from failed logic.","Repeat the same viewport/task after layout or copy changes; record actual width and any unresolved limitation.","Stop before repeated submissions or unintended external actions. Crop/redact any screenshots privately before sharing.","Lovable initial mobile captures were too wide and later retaken; brief blank admin reload was logged. Bolt printed a literal escape sequence.","L1","https://prototypeverdict.com/guides/what-actually-broke/ | https://prototypeverdict.com/how-we-test/","Only own authorized apps and synthetic data. Stop on exposure or consequential scope. No destructive clears, security bypass, purchases, production cutover, certification, or launch clearance.","","","","","","Not checked","","","","","Not checked","" "2026-11-07.2","2026-11-07","2026-10-29","Proposed test; not a lab-tested repair; no execution claimed","PV-20","Record usage without treating units as comparable cost","Observation only","Builders displaying usage; optional before/after an already-approved normal task","Owner of builder workspace","Access to the current allowance display. Choose an action already planned and authorized independently; this kit does not ask you to buy, upgrade, or generate extra AI work just to measure it.","1. Record displayed unit, remaining/used amount, denominator, and whether daily/monthly/project-specific before the planned task. 2. Observe the one authorized task; do not repeat to force a measurement. 3. Record the same panel after it and any completion/gate state. 4. Note grants/bonus pools and disagreement between counters as unknowns. 5. If a limit appears, pause new requests and read the actual display.","A bounded observation identifies what that display measured for that task, with source/date and uncertainty. No cross-builder cost winner, reset promise, or unobserved completion is inferred.","Different units, prices, daily/monthly denominators, grants, and metering prevent controlled cost comparison. A blocked request is not a failed implemented feature test.","Only if another normal task is independently planned, repeat the same observation method. Keep each run separate; never average unlike pools or turn one observation into a forecast.","Stop on quota gates or unexpected paid-work/purchase prompts. Do not upgrade, purchase credits, schedule retries, or keep pressing fix to complete this test.","Bolt exhausted its observed allowance before CSV; Lovable counters disagreed. Neither establishes a general cost ranking or tested reset remedy.","L1","https://prototypeverdict.com/guides/what-actually-broke/ | https://prototypeverdict.com/how-we-test/","Only own authorized apps and synthetic data. Stop on exposure or consequential scope. No destructive clears, security bypass, purchases, production cutover, certification, or launch clearance.","","","","","","Not checked","","","","","Not checked","" "2026-11-07.2","2026-11-07","2026-10-29","Proposed test; not a lab-tested repair; no execution claimed","PV-21","A checkpoint explains consequential scope before approval","Observation only","Only if a build presents an approval checkpoint","Owner/authorized maintainer","An existing checkpoint, not a deliberately created destructive task. No live changes are authorized by this worksheet.","1. Read the proposed action and identify affected environment/records. 2. Ask whether it changes data, access, publication, or spending. 3. Record whether a non-destructive alternative and preservation plan are explained. 4. Stop before approval unless the owner independently authorizes the exact action. 5. If it is only clarification, answer within the existing scope and record the next state separately.","The scope and consequence are clear before any approval. A pending or declined destructive action can be the correct outcome; do not grade automatic continuation as success.","A permission checkpoint can be intentional. “Continue” is not a repair, safety proof, or permission for arbitrary deletion.","If the maintainer revises the proposed action, review its exact new scope; do not carry forward a generic earlier approval.","Stop on unclear deletion, public launch, access change, or new spending. Never approve merely to finish the test.","Base44 paused before deleting its own test record; logged continue was redundant. No causal repair was demonstrated.","L1; M5; M6","https://prototypeverdict.com/guides/what-actually-broke/ | https://prototypeverdict.com/how-we-test/ | https://docs.lovable.dev/features/projects/history | https://support.bolt.new/managing-projects/rollback-backup","Only own authorized apps and synthetic data. Stop on exposure or consequential scope. No destructive clears, security bypass, purchases, production cutover, certification, or launch clearance.","","","","","","Not checked","","","","","Not checked","" "2026-11-07.2","2026-11-07","2026-10-29","Proposed test; not a lab-tested repair; no execution claimed","PV-22","Rehearse a data move without production cutover","Maintainer-assisted migration","Only when moving data/storage is an explicit requirement","Owner; authorized maintainer; Client A/B","A separate target test environment, source preserved, synthetic-only fixture set, explicit field/ownership mapping, expected counts/totals/files, and a written cutover/rollback plan. No live writes or destination changes by this kit.","1. Inventory the synthetic source dataset and separately inventory files and identity dependencies. 2. Have the maintainer prepare a scoped nonproduction rehearsal through supported tooling; do not assume code import moves records/passwords/files. 3. Compare target counts, marker fields, totals, ownership, and attachment availability. 4. Run PV-02/PV-04/PV-09/PV-11 and relevant file tests in the target. 5. Review how new writes, duplicate IDs, reconciliation, fallback, and final owner approval would be handled before any eventual cutover. Do not perform production cutover here.","A documented rehearsal reconciles the fake dataset and access matrix, with source intact and unresolved differences listed. Passing rehearsal does not authorize live cutover or guarantee recovery.","A running UI with an empty target database is not a completed migration. Code import/restore does not prove data, credentials, or files moved. Correct counts alone do not prove ownership or permissions.","After a mapping or process fix, rerun the same fixture rehearsal in a reviewed nonproduction target; preserve both evidence sets and avoid uncontrolled duplicate writes.","Stop on any proposed production move, overwrite, source deletion, credential export, unexpected real data, or automatic switch of live form/database destinations.","Lovable’s offered database follow-up was not run. No storage migration, cutover, or recovery was tested in this lab.","L1; M1; M2; M4; M5; M6","https://prototypeverdict.com/guides/what-actually-broke/ | https://prototypeverdict.com/how-we-test/ | https://developer.mozilla.org/en-US/docs/Web/API/Window/localStorage | https://supabase.com/docs/guides/database/postgres/row-level-security | https://supabase.com/docs/guides/storage/security/access-control | https://docs.lovable.dev/features/projects/history | https://support.bolt.new/managing-projects/rollback-backup","Only own authorized apps and synthetic data. Stop on exposure or consequential scope. No destructive clears, security bypass, purchases, production cutover, certification, or launch clearance.","","","","","","Not checked","","","","","Not checked","" "2026-11-07.2","2026-11-07","2026-10-29","Proposed test; not a lab-tested repair; no execution claimed","PV-23","Keep project restore separate from database recovery","Planning only","Apps whose troubleshooting plan involves rollback or restore","Owner; authorized maintainer","Read-only access to relevant project/version documentation and an inventory of data dependencies. No restore is required by this test.","1. Identify what a proposed restore actually covers: code, functions, data, migrations, files, or identity. 2. Read the platform’s current warning. 3. List data changes that would remain after code restore. 4. Ask for a separate tested data-preservation/recovery plan before any future destructive step. 5. Mark recovery behavior Not checked unless an independently authorized safe rehearsal actually ran.","The owner understands that a code/version restore may leave current database data unchanged; no untested recovery guarantee appears in the plan.","A version-history button or code ZIP is not proof of a restorable database backup. A code rollback can also become incompatible with current schema/data.","Review this distinction again after schema/storage changes. Any real recovery exercise needs its own scoped environment, approval, and evidence.","Do not click restore, delete records, reset schema, clear storage, or test recovery on the only surviving source merely to mark a row passed.","No rollback was used in the Oct 29 tests. Lovable/Bolt official docs explicitly separate project restore from database data.","M5; M6; L1","https://docs.lovable.dev/features/projects/history | https://support.bolt.new/managing-projects/rollback-backup | https://prototypeverdict.com/guides/what-actually-broke/ | https://prototypeverdict.com/how-we-test/","Only own authorized apps and synthetic data. Stop on exposure or consequential scope. No destructive clears, security bypass, purchases, production cutover, certification, or launch clearance.","","","","","","Not checked","","","","","Not checked","" "2026-11-07.2","2026-11-07","2026-10-29","Proposed test; not a lab-tested repair; no execution claimed","PV-24","Retest the repair and adjacent allow/deny behavior","Finish each scoped change","Any app after an independently approved repair","Same identities as failing case; authorized reviewer","A redacted failing attempt, a new identifiable version/change label, and the same synthetic fixtures or a documented equivalent. Original evidence must remain intact.","1. Duplicate the CSV for the new run rather than overwriting old evidence. 2. Repeat the exact failing steps with the same role/environment. 3. Repeat the nearest legitimate allow case and protected denial case. 4. Record actual outcomes, remaining errors, method, and timestamp with timezone. 5. If a test is blocked or not relevant, state why; do not convert it into Passed. 6. Review unresolved failures/unknowns with the owner before any launch decision.","The specific regression is demonstrably resolved or still open; adjacent function/access behavior is recorded separately. Results are self-recorded checks, not certification or permission to launch.","An AI success message, partial screenshot, or unrelated passing path cannot close the original failure. Repeated failure suggests stopping the repair loop and reviewing evidence, not spending indefinitely.","Use a new run ID for every material repair/version. Preserve before/after evidence and re-open failures after relevant changes.","Stop on exposure, data loss, unclear authorization, quota gate, or new consequential scope. This kit makes no publication/payment decision.","Softr claimed groups were reactivated, but settings contradicted it after reload. Observed behavior must control the result.","L1; O1","https://prototypeverdict.com/guides/what-actually-broke/ | https://prototypeverdict.com/how-we-test/ | https://cheatsheetseries.owasp.org/cheatsheets/Authorization_Cheat_Sheet.html","Only own authorized apps and synthetic data. Stop on exposure or consequential scope. No destructive clears, security bypass, purchases, production cutover, certification, or launch clearance.","","","","","","Not checked","","","","","Not checked",""
Public sources
- What Actually Broke in Our AI App Builder Lab
- How We Test AI App Builders
- Fix AI App Builder Errors
- PrototypeVerdict Usage Index — allowance pressure, not a cost ranking.
- MDN: Window.localStorage — official reference, not our test.
- Supabase: Row Level Security — concepts only, not a SQL recipe.
- Supabase: Redirect URLs — not a proven cause.
- Softr: Global data restrictions — current docs, not a retest.
- Lovable: Version history
- Bolt: Backups, restore, and version history
By the PrototypeVerdict Test Team
Answers stay in this page’s memory when the guided flow is running. This summary does not collect them either. No sign-in, upload, app connection, or remote AI.
PrototypeVerdict · Lab-backed troubleshooting
Your next safe step, before another fix.
Your app looks broken. Before another fix attempt, narrow the symptom, protect your data, and leave with a focused next message and a checklist. For the Lovable builds behind those lab notes, see our Lovable review.
Small unpublished apps · Oct 29 lab · Nov 12 fix-loop note · Suggested checks, not a diagnosis
Answers stay in this page’s memory. This widget does not transmit them. No sign-in, upload, app connection, or remote AI. The host page may still use its own analytics. Copy or download your notes before leaving; reloading starts over.
Lab update · November 12, 2026
First-try fix-loop results (single run)
This is a separate Nov 12, 2026 test, not a re-run of the 24 proposed prelaunch checks in the Rescue Desk. Official first-try rates apply only to measurable seeded bugs. They are not 10-bug pass rates. Items that did not seed have no pass or fail mark.
| Tool and plan | First-try rate | Credit use |
|---|---|---|
| Lovable Pro | 4/4 (100%) of measurable seeded bugs (6 not seeded) | Daily build credits: −2.5 across the 4 fixes, plus −0.8 to seed. Monthly credits stayed at 62.5. |
| Base44 Builder | 4/5 (80%) of measurable seeded bugs (5 not seeded) | +8.91 app message credits and +8 integration credits across the 5 official fix messages; seed +4.13 message and +8 integration. |
| Softr Basic | 0/0 (undefined / incomplete) | Seed prompt refused; Basic code editor read-only; all 10 defects not seeded. Displayed AI credits stayed 37.76. |
| Bolt Free | Not run | No T3 first-try rate or cost from this test. |
The one Base44 failure (guest access to a previously public fabricated test file) still failed after the first-try message. That observation rests on the browser navigation log. We did not test a 390 px phone width (Lovable was measured at 393 CSS px; Base44 at a 373 CSS px preview). Lovable’s admin-status fix has no mobile capture. Per-bug costs and the full wording limits are in the fix guide. Lovable credit totals are also in what it costs to fix a Lovable app.
The Rescue Desk checks remain suggested tests, not a diagnosis, security audit, or launch clearance. This update does not transmit answers and does not change the guided flow.
Lab observations: Oct 29, 2026. Fix-loop run: Nov 12, 2026. Sources reviewed: Nov 7, 2026. The 24 prelaunch checks were not re-executed. Published: Nov 8, 2026. Updated: Nov 16, 2026.