Add Compass as site 23 with source-backed data and verified tasks

[jackjin1997]

Reviewer takeover of original #25, contributed by sarendis56 (Peichun Hua). Original commits and attribution are retained.

Result

This revision integrates Compass as WebHarbor's 23rd site on container port 40022. It preserves Rotten Tomatoes on 40021 and incorporates current main at 5e279ad1f925987e51f878a461e876064ea77635.

• Adds a source-backed Compass mirror with 505 listings, 255 agents, 19 open houses, 24 regional guides, four benchmark users, and a versioned deterministic SQLite seed.
• Adds 19 task/verifier contracts: Compass--0 through --7 and --10 through --20. Unsupported task numbers 8 and 9 remain unused without renumbering existing tasks.
• Derives every verifier target from the supplied initial database. Exact rankings fail closed on ties. Verifiers enforce task identity, local origin, ordered page workflows, fully decoded PNG evidence, answer semantics, exact schema, and bounded before/after state.
• Adds request/session/input limits, CSRF-protected POST mutations, ownership checks, safe local redirects, SQLite foreign keys, uniqueness/constraint handling, one-time seller confirmations, and atomic/versioned seed construction.
• Adds responsive and accessibility repairs for 320/390/815/1440 px layouts, dialogs, Sell carousel progressive enhancement, keyboard semantics, focus handling, reduced motion, and local media fallback.
• Removes committed answer-bearing facts and 242 solution/evidence artifacts. The audit report remains; complete task trajectories and screenshots stay outside the agent-visible repository.
• Preserves current-main OSU and Rotten Tomatoes build behavior while extending 23-site startup, health, restart and reset behavior.

Independent review

Sixteen independent review contexts covered security/state, application/data, task ground truth, adversarial verifier behavior, UI/responsive behavior, current-main integration, test evidence, provenance/repository hygiene, four first-fix re-reviews, and four completion audits. Every finding, verification step and disposition is recorded in PR-84-FINAL-AUDIT.md.

Asset provenance

• sites/compass/asset_inventory.json covers exactly 2,907 runtime files totaling 209,945,248 bytes and records source URL, byte length and SHA-256 for every file.
• Normal host and Docker builds fail on missing, extra, stale, malformed or hash-mismatched inventoried assets.
• Archive extraction validates site roots, managed paths and member types, stages content, replaces managed roots, and rolls back failed replacement.
• Pinned HF revision: 070123d74c01a8b29808201be85462fd7d0ec3c4 with 23 site archives.
• Deterministic Compass media archive SHA-256: 4af8396785c35db1a5cf57ed5b77e00c8a7b8c484b250707ba64d1936691d8d5.
• HF #59 and HF #60 are merged; obsolete HF #53 is closed.

Validation

• Compass application/seed/archive/asset/task/verifier suite: 229 passed.
• Cross-site regression: Rotten Tomatoes 66 passed / 4,143 subtests; OSU 25 passed / 264 subtests; TED 27 passed / 291 subtests.
• Clean build with Compass, OSU and Rotten Tomatoes host seed directories absent: PASS.
• Release image: sha256:cf6fe05b6600e92727816aff018f90d19f18277e8276f6e54a46aaecc34c79af (4,681,221,472 bytes).
• Packaged Playwright execution plus deterministic verification: 19/19 tasks, 209 browser steps, PASS.
• Packaged responsive run: 140/140 route/viewport checks, zero broken images and zero page overflow.
• 23/23 sites HTTP-ready; all roots return HTTP 200.
• /reset-all: 23/23 ready and every runtime seed tree byte-identical to instance_seed.
• Ten consecutive Compass resets: PASS; control-process file descriptors remained 4 before and after.
• Compass image seed/runtime SHA-256: a21bc6d0f4a1063d1807b4d4b6c2ff4851e99d777cedac6d7070c2f5aa95f257.

The deterministic verifier validates the submitted trajectory, state, answer and decoded screenshots. It does not cryptographically authenticate the recorder or bind screenshot pixels to a URL. The independent packaged Playwright execution is retained as execution evidence for that boundary.

[jackjin1997]

Independent blind first-pass review (Compass, 16 tasks)

Reviewer model: Claude Fable 5.1 (claude-fable-5-1), first pass restricted to the frozen execution packet. No verifier code/results, expected answers, PR contents, source, or live DB were read before this verdict was frozen.

Scope

• Reviewed code: b66016dd16dfc7eb776f58d5490fc31caecac68c (current PR head at posting time). Image code actually run: 3cdc0b66d72b8a789b89583176c668081de36852 (wh-review025:3cdc0b6).
• HF asset revision: 47232eba972567d138bbac478a6f4af9775a1d90.
• Input manifest SHA-256: 3125bf726ae7e8faa3020c070d5f08037b619f0474ab3c2c552a272e25be41fd (618 listed files, all sizes/hashes verified, no extra files).
• Frozen verdict SHA-256: 342633f8e1a3f0a31cf5983912c196573a4c7336591d66d3c315938f1035f662.
• Executor: Codex GPT-6, guided/source-aware run. These are judgments of whether those runs completed the tasks, not a measure of independent exploration capability.

Result: 16/16 PASS, 0 FAIL (Compass--0 to 7, 10 to 17). No failed task IDs.

Method

• Every before.db equals the runtime seed; all SQLite read read-only.
• Each read-only task's target was recomputed independently from the seed DB using unrounded price/area ratios and compared against the final answer and the property-details DOM.
• Row-level diff of all tables before/after for every task: write tasks (4, 5, 6, 7, 10, 14, 15) changed exactly the requested rows and nothing else; read tasks left the DB byte-identical to the seed. The new account's password hash was bcrypt-verified.
• Trajectories and DOM observations checked for genuine local UI evidence on the decisive pages (details, Saved Homes, Tours, collection and share page, Saved Searches reopen, Inquiries); 4 screenshots sampled visually.
• Compass--3 contains one retained failed link lookup (selector without the "Pending" badge prefix), recovered on the next step; not counted against the task.

Interpretive assumption worth cross-checking: listings with status = for-sale and a "Pending" badge were treated as for-sale eligible, matching the snapshot status field and the app's own "For sale" filter. This affects the selections in Compass--3, --11 and --17; if pending homes are meant to be excluded, those three would need re-evaluation. " search area" was mapped to market_city; using city gives identical results.

Not performed / not certified

• Did not re-run the app or image; did not verify harness, archive or image hashes, or raw JPEG-to-PNG fidelity.
• Only 4 of 258 screenshots inspected visually; the rest relied on complete DOM observations.
• Compass--15 message body verified from the after DB only (the inquiries list shows subject, not message).
• Human experience, upstream source fidelity, visual quality and verifier implementation are out of scope and not certified by this review.

Structured per-task rationale and evidence are in the frozen verdict file held by the reviewer.

[jackjin1997]

Independent-review reconciliation

Rechecked the frozen verdict hash and all 618 input-file hashes. All 16 task verdicts agree with deterministic scoring. The original Claude verdict remains unchanged.

• Pending: the existing local For sale filter includes Pending / Contract Signed sale listings; their source status is preserved in property details. Tasks 3, 11 and 17 do not ask for active-only listings. Recomputing their selections supports the recorded PASS results.
• Search area correction: city and market_city are not interchangeable for New York. Literal city='New York' yields no qualifying Co-ops; the actual UI search area (city OR market_city) includes seven, including Brooklyn/Manhattan addresses. Rechecking that query preserves task 3's target and PASS.
• Screenshot correction: the immutable manifest contains 268, not 258, screenshots. Claude visually sampled four; the review does not establish comprehensive visual acceptance.
• Task 15: independently confirmed exactly one inquiry with the requested subject/message and correct user/listing/agent, with all other tables unchanged. The task requires persistence, not a subsequent message-body display.

Updated report and evidence boundaries. This follow-up changes documentation only and reuses the frozen review candidate b66016dd16dfc7eb776f58d5490fc31caecac68c. Human experience remains pending; the PR stays Draft.

[jackjin1997]

Addressed the human-reported missing Sell section in e55721b5eb9bd814910dcb3895a665558d730276 (application checkpoint a5e25af): desktop/mobile navigation and the homepage entry now reach /sell/ instead of a missing page. The local introduction uses the original hero image and measured desktop/mobile typography/layout; its contact CTA opens the existing agent directory. The source seller lead form and full seller-service workflow are not implemented.

Four original / before / after comparison groups and remaining differences · Checks and image/input hashes

Sell, 1440×900	Original	Before	After
Actual browser capture

Validation: 42 application/source tests pass, existing Buy/Rent and login/navigation work in the real browser, all 20 homepages return 200, and Compass reset is byte-identical. The incremental Docker build and owner runtime have the same six changed file hashes; updating the owner environment preserved its database and left the other 19 processes untouched.

HF #53 now pins 421ca6a529b88bdd214ecf6308d124798ab1e20b: one added Sell image, all 2,412 previous archive files including seed unchanged, other 19 site archives unchanged. The PR body/report distinguish this update from the frozen task runs and Claude verdict. Task/rubric/verifier inputs are unchanged; original task evidence is reused by impact with fresh shared-entry checks. Claude has not reviewed the new Sell page. Human experience is still pending; this PR remains Draft.

[jackjin1997]

Homepage/navigation correction pushed in 9741284 (evidence/report 5859a3c). The earlier Sell-only patch left substantial structure missing. Original / before / after comparisons, scope limits and validation now cover the desktop/mobile homepage, all dropdown groups, drawer, services, community banners and footer; 32 verified viewport screenshots are included.

Current checks: 168 tests PASS, 91 local homepage URLs/assets, 20/20 site homepages, matching 29-file incremental image, and byte-identical Compass reset. Compass--17 was newly recorded at the final application commit: 8 steps / 16 screenshots, deterministic PASS and unchanged DB. Its refreshed independent review is pending; the older 16-task review is not relabeled as coverage of this UI.

HF #53 now pins e06307429dca76a8a7240d507ab69be9e5b84060 (16 source media added; all 2,413 prior files and 19 other archives unchanged). PR remains Draft for human visual acceptance, confirmation of the disclosed deep-page limits, and the refreshed independent verdict.

[jackjin1997]

Follow-up: the actual 815px viewport exposed remaining structural defects in 9741284. Commit 82b5915 fixes the tablet hero, two-column cards/promotion and neighborhood layout, stacked Concierge source imagery, directory folding and footer columns.

The updated source/before/after report adds 20 verified viewport images (52 retained total), final responsive and runtime checks, asset pin 5b66b75, and explicit evidence-reuse scope. The existing frozen task 17 execution stays bound to 9741284; it was not relabeled or overwritten. Human acceptance and its independent verdict remain pending; this PR remains Draft.

[jackjin1997]

Concierge and Neighborhood Guides reconstruction is now in 25b8b3375aeb5a3b2c5823f8ed6d52a05eddeccf (report fd30342795f477165adb3c794a5e8cbe5205a3f2). The former one-panel Concierge and six-tile directory now use source-derived structure, original media and responsive layouts, including all eight Concierge case states and 24 directory regions.

26 source / before / after screenshots, measured sections and remaining boundaries · Validation and hashes

Validation: 168 tests passed during the repair; final browser checks exercise the case controls and local CTA. The final incremental image matches 47 runtime files; all 20 homepages and 69 local routes/media URLs return 200, and Compass reset restores the seed byte-for-byte. HF #53 now pins b09ad95ac9428d5f98e1dc6206f1fa5f87aec671 with 38 added public assets; all previous asset bytes and seed are preserved.

Human visual/scope acceptance and the refreshed Compass--17 independent verdict remain pending. Existing frozen executions keep their original versions; these new page captures are UI QA evidence. PR remains Draft.

[jackjin1997]

Sell and regional-guide reconstruction is now in 7f6dc008603182c6ace79ad52e52f2f8b804ae4b (report 4a80f3819f0c3b443a11663c5ec79a8862c3a234). The previous short Sell page and regional property-list substitutes were material fidelity gaps. The candidate now includes13Sell sections,3/6-state carousels, local inquiry persistence, and24source-derived regional landings with401community cards. A final source comparison also corrected the nearby-home groups from three columns to the original stacked layout.

36 source / before / after and interaction captures, measured geometry, validation and explicit scope limits.

178testsPASS;20homepages+470localURLs200; realUI inquiry/reset and23/23reset-all seed equality. Final455payload hashes match the prepared image and owner runtime; original benchmark rows are preserved. HF53 now pins d86ec0bfcbb98f92efbe6b6b4440f9dcc6cceb69 with442publicmedia additions and an additive seed migration; other19archives are unchanged.

Frozen task trajectories and the earlier Claude verdict retain their original identities. Human visual/scope acceptance and the refreshed task17 independent verdict remain pending. PR stays Draft.

[jackjin1997]

Marked Ready for maintainer review at the owner's explicit request, following acceptance of the current human experience and disclosed scope. Documentation-only handoff: 2579570e11e8d8220232187d7b02bd3a52a80472; application and HF assets remain 7f6dc00 / d86ec0b.

Maintainer handoff and remaining actions · Fresh transition checks.

Rechecked current455-file runtime payload,20 site homepages,16 grading contracts,932 original execution files,39 refreshed task17 inputs and36 public image hashes. The existing178-test and realUI/reset evidence applies unchanged. HF53 is reachable, open and conflict-free; all archive hashes match.

The original16-task independent result is complete and reconciled. The refreshed task17 independent verdict has not returned and is not marked PASS. It remains disclosed for resolution alongside HF merge/pin coordination and final maintainer review. No GitHub or HF merge was performed.

[jackjin1997]

Integrated upstream main 7269134 after TED/OSU merged; the PR remains Ready. Compass is now on 40021 in the22-site registry. Its task origins were updated without changing prompts, grading or UI source.

230 tests pass,22 homepages/470 Compass URLs200. Actual UI inquiry/reset passes: Compass0.74s, reset-all3.57s,26/26 seed files equal. OSU WebP and stable seed-build issues found by integration tests are fixed. HF #53 now pins 8ade6b3a80d376c801f88e2409b72e8909444927 with22 verified archives; Compass and the19 existing archives are unchanged.

Current integration evidence · Maintainer handoff. Old frozen runs and verdicts remain unchanged. Refreshed task17 independent verdict is still pending; no new independent PASS is claimed. Maintainers coordinate final public merges.

[jackjin1997]

Added 3 tasks, bringing Compass to 19 while preserving all16 existing contracts and retired IDs8/9:

• 18: select the latest existing tour, read property details, cancel only that tour.
• 19: compare price/area inside an existing collection and remove only the selected member, preserving unknown-area members and the share link.
• 20: read account contact details, submit one local Sell inquiry, report its generated receipt.

Also fixed a missed verifier compatibility issue: the current seed has11 tables, while the old snapshot loader required9. Historical snapshots remain supported; guide and seller-inquiry data are now protected from unrelated writes.

**Validation:**222 Compass tests pass, including44 new regressions. Three new actual guided runs pass with36 steps/72 screenshots and full before/after state. All16 historical native exports plus the task17 follow-up regrade PASS without changing frozen inputs. New runs used documented native Flask startup and pinned assets because the shared Docker host rejected new containers with a read-only-filesystem error; no new Docker run is claimed.

Task quality matrix, eight actual screenshots, exact versions and reproduction. The new frozen blind packet covers task17–20, with88 screenshots; its independent verdicts are pending. Public Ready and HF asset revision remain unchanged. Maintainers retain final review/merge authority.

[jackjin1997]

Independent blind review — Compass--17, 18, 19, 20 (blind-expansion-4542b35 bundle)

Reviewer: Claude Fable 5.1 (claude-fable-5-1, Claude Code), 2026-09-08.
Scope: frozen executions only, judged from the handoff bundle (task + rubric, full trajectory, DOM observations, screenshots, final answer, before/after SQLite opened read-only). The verdict file was written and hashed before any verifier code, expected answers, scorer output, prior review or this PR's discussion was read.

Inputs / outputs

• Bundle manifest.json SHA-256: 02a89fbb0442ca1abd93d9c305e014eeea299d2a5f2475e7a448bb691d8b94f8 — 273 entries, every file hash and size verified, no unlisted files; screenshot-export.json raw/native/decoded-pixel hashes verified for all 72 JPG/PNG pairs in 18–20.
• verdict.json SHA-256: 7944248a06963a8d85c944da96435ac475234802bb2fb86e07c520bd1bca4f97.
• Recorded versions: Compass--17 is the unchanged earlier prepared execution at code 9741284, asset e063074, 390×844 mobile viewport, Docker image. Compass--18/19/20 were recorded 2026-09-08 at code 4542b35, app checkpoint 4574f24, asset 8ade6b3, native Flask process, 1280×720, all three starting from the identical seed DB. These are not one version rerun. All four runs are guided / source-aware; nothing here measures independent exploration difficulty.

Per-task verdicts

Task	Verdict	Persistent change (row-level before→after diff)	Answer vs snapshot
Compass--17	PASS	none — after.db byte-identical to before.db	150 Charles Street, Unit M10 / Holly Parker / 2013 is the top for-sale Condo ≥ $5M ($21.5M; next is 1425 Brickell Ave 42F at $11.75M). It carries a Pending / "Contract Signed" badge while its snapshot status is for-sale; the final answer discloses this.
Compass--18	PASS	only tours.id=5 status confirmed → cancelled, row retained	Carol's uniquely latest non-cancelled tour (2026-06-16) → 170 Northwest 44th Street, Miami, FL 33127, year built 2026; Tours page shows "Tour cancelled." and the other tour still requested.
Compass--19	PASS	only collections.id=2 listing_ids_json [16,18,9] → [16,18]; share token, saved_homes, all other tables unchanged	$1,495,000 / 1,200 sqft (≈1,246) > $1,125,000 / 1,217 sqft (≈924); unknown-area member excluded from comparison and kept. Removed 130 Prospect Place, Unit 1, Brooklyn, NY 11217, built 1930; collection page confirms 2 remaining members with the same share token.
Compass--20	PASS	only seller_inquiries +1 row (ref 8f0950b7ea87798c27896315, Alice Johnson, alice.j@test.com, (415) 555-0144, 94107, created 13:28:53 matching the submit step); users unchanged	Values match the account overview; receipt shows the same reference before and after reload; exactly one row.

Evidence locators (relative to the bundle)

• 17: observations/001-after.txt (luxury page lists the listing), 005-after.txt (For sale + Condo + min $5M, price desc → 4 results, this one first), 007-after.txt (agent link, Status Contract Signed, Year Built 2013); screenshots/step_011.png, step_013.png, step_015.png.
• 18: observations/004-after.txt (tours before), 006-after.txt (Year Built 2026), 009-after.txt (tours after); screenshots/step_009.png, step_013.png, step_019.png.
• 19: observations/005-after.txt/006-after.txt (members, share token), 008-after.txt (Year Built 1930), 010-after.txt (after removal); screenshots/step_013.png, step_017.png, step_021.png.
• 20: observations/003-after.txt (account overview), 010-after.txt and 011-after.txt (receipt before/after reload); screenshots/step_007.png, step_019.png, step_021.png, step_023.png.
• DOM observation chains (after N == before N+1) are unbroken in all four tasks.

Findings / caveats (no verdict changed)

1. Compass--17 qualifying-status ambiguity. The chosen listing is in contract (Pending / Contract Signed). Under the fixed snapshot its status field is for-sale and the UI "For sale" filter includes it, so it qualifies as worded. If "for-sale" is meant to exclude in-contract listings, the expected answer becomes 1425 Brickell Avenue, Unit 42F (Audrey Ross, 2003). Worth stating explicitly in the task or rubric.
2. Compass--17 screenshot gap. The agent name is only in the DOM observation; no captured 390×844 viewport screenshot contains it. Also this task has no raw JPGs / screenshot export, and its trajectory.json does not link the observation files (they exist and chain consistently).
3. Pixel-chain breaks in 19 (steps 5→6, 6→7) and 20 (1→2, 4→5, 6→7, 7→8, 9→10, 10→11): after-screenshot ≠ next before-screenshot by pixel hash while the DOM is identical — consistent with cursor blink / lazy image loading, not with state changes.
4. Full-page screenshots were not captured for 18–20 (declared in the export).

Screenshot coverage

Viewed individually at full size: 17 step_003/011/013/015; 18 step_009/013/019; 19 step_013/017/021; 20 step_007/019/021/023. All 88 native PNGs were viewed as labelled thumbnails to confirm the page sequence. Raw JPGs were hash-verified, not viewed.

Not executed

No replay or environment launch; no verifier / code review; no source-fidelity check against the upstream site or the asset revision; no verification of the documented commit / asset / archive hashes against the repository during the blind pass (after freezing I only confirmed that 9741284 and 4542b35 are commits on this PR). This comment judges the frozen executions only and does not certify source fidelity or the verifier implementation. No merge or Ready-state change is requested.

[Raibows]

Final 16-agent review remediation is pushed at a78ee91d2cd2526ed9d0d25ccc3451b53ca37c30.

• Current main 5e279ad1f925987e51f878a461e876064ea77635 is integrated; the PR is now MERGEABLE / CLEAN.
• Compass is site 23 on 40022; Rotten Tomatoes remains on 40021.
• All 16 independent agent findings, verification evidence and dispositions are in the final audit.
• Compass suite: 229 passed. Cross-site regression: Rotten Tomatoes 66/4,143 subtests, OSU 25/264, TED 27/291.
• Exact packaged image: sha256:cf6fe05b6600e92727816aff018f90d19f18277e8276f6e54a46aaecc34c79af; 19/19 task runs with 209 browser steps passed, and 140/140 responsive checks passed.
• 23/23 sites were HTTP-ready; /reset-all restored every runtime seed tree byte-for-byte; ten Compass resets left the control-process FD count at 4.
• Runtime assets are pinned at HF revision 070123d74c01a8b29808201be85462fd7d0ec3c4; the deterministic Compass archive SHA-256 is 4af8396785c35db1a5cf57ed5b77e00c8a7b8c484b250707ba64d1936691d8d5.

[Raibows]

Thanks for your contribution! @jackjin1997 @sarendis56