feat: add Bell Integrator outside-in audit v0.1

[safal207]

Summary

Introduces an independent, evidence-first Bell Integrator public audit across three lenses:

• QA and content integrity;
• system and content-model consistency;
• business and candidate-conversion impact.

The branch starts from exact main SHA 19d1cd72472666df21350c6f2b2e54591eb3f462.

What changed

• added a bounded machine-readable audit contract;
• added a sequential public-content probe using Python standard library only;
• added a desktop/mobile rendered matrix with screenshots, accessibility state, keyboard trace, and console/network summaries;
• added fail-closed regression tests for origin and authority boundaries;
• added read-only GitHub Actions workflows with exact-attempt artifacts and checksum verification;
• documented source-to-rendered adjudication;
• added a concise collaboration brief suitable for a first conversation.

Confirmed candidates

1. BELL-001 — the QA service is published with the ABS modernization description — HIGH;
2. BELL-002 — corporate age and office footprint claims are inconsistent across the current domain — MEDIUM, pending business-definition review;
3. BELL-003 — the careers page contains a visible subject-verb agreement defect — LOW;
4. BELL-004 — current engineering vacancies contain visible mixed-script, inflection, and spelling defects — LOW.

All four were reproduced in settled desktop and mobile profiles on exact source head e764606889ea9871ea1f0ccd65a17599f374a42a.

Source-to-rendered adjudication

The first raw pass reproduced four source-level signals. The first browser pass confirmed BELL-001 and BELL-002, while exposing two over-broad assertions:

• older employee-story text existed in the document source but was not visible in settled desktop or mobile innerText;
• a split с огласованных source sequence rendered as the normal word согласованных and was therefore a parser-boundary artifact, not a user-visible defect.

The contract was narrowed to visible evidence: Мы предоставляет, mixed-script QА, особенностей управление памятью, and инстуменами отладки.

Decision model

search or cached observation
→ NEEDS_EVIDENCE

current public response marker reproduced
→ PRODUCT_SIGNAL

settled desktop + mobile reproduction
→ CONFIRMED_PRODUCT_DEFECT_CANDIDATE

human semantic and impact review
→ final severity / collaboration decision

No public marker is treated as proof of internal root cause or measured commercial loss.

Safety boundary

Public allowlisted HTTPS pages and natural sequential GET navigation only.

No authentication, form submission, direct application API testing, enumeration, fuzzing, load testing, active security testing, vulnerability claim, external contact, remediation, deployment, delivery, or merge is performed or authorized.

Exact validation

On source head e764606889ea9871ea1f0ccd65a17599f374a42a:

• raw audit: success, 6/6 routes, 4/4 PRODUCT_SIGNAL;
• rendered audit: success, 12/12 desktop/mobile observations, 4/4 CONFIRMED_PRODUCT_DEFECT_CANDIDATE;
• all final origins remained bounded;
• no anti-bot variants, console errors, failed requests, or HTTP 4xx/5xx were observed;
• authority-boundary and checksum verification passed;
• repository CI and Security Audit passed.

This PR remains draft pending the human collaboration and reporting decision.

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on this repository. Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro Plus

Run ID: 8a0db3a4-d763-4747-8be1-b7083ccc24b8

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch agent/bell-integrator-outside-in-audit-v0-1

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[safal207]

Source-to-rendered adjudication

The first exact raw run completed successfully with 4/4 source-level PRODUCT_SIGNAL findings across six bounded public routes.

The first rendered run completed successfully with 12/12 desktop/mobile observations, bounded final origins, and no anti-bot, console-error, failed-request, or HTTP 4xx/5xx variants. It promoted BELL-001 and BELL-002 and correctly withheld two over-broad claims.

What browser evidence changed

• Older employee-story text exists in the document source but is not visible in settled desktop or mobile innerText; it is not promoted as a current user-visible defect.
• The raw с огласованных sequence renders as согласованных; it is a parser-boundary artifact, not a UI typo.
• Visible Мы предоставляет, mixed-script QА, особенностей управление памятью, and инстуменами отладки remain valid rendered signals.

The contract and both documents were narrowed accordingly. Corrected raw and rendered runs are now the promotion gate.

root_cause: HYPOTHESIS_ONLY
business_impact: PLAUSIBLE_NOT_MEASURED
security_claim: NOT_MADE
external_contact: NOT_AUTHORIZED
merge: NOT_AUTHORIZED

[safal207]

Final exact evidence · source head e764606889ea9871ea1f0ccd65a17599f374a42a

Both corrected evidence layers and repository checks completed successfully.

Raw public layer

• run: 29897326377
• coverage: 6/6
• decision: PRODUCT_SIGNAL
• findings: 4/4
• artifact digest: sha256:c937354babb85ad0d15ad6e9962d08213f9abbff7a5b15ec0168c0baae9e4fda

Desktop/mobile rendered layer

• run: 29897326392
• coverage: 12/12
• decision: RENDERED_PRODUCT_DEFECT_CANDIDATES
• candidates: 4/4
• artifact digest: sha256:244a8cccf060e6b523e0a6326ec7ba64647b8b823ee4b163a0b4a58c1ec4f6cf

All routes returned HTTP 200 in both profiles. Final origins remained bounded. No anti-bot variant, console error, failed request, or HTTP 4xx/5xx was observed.

Repository CI and Security Audit also completed successfully.

Judgment boundary

root_cause: HYPOTHESIS_ONLY
business_impact: PLAUSIBLE_NOT_MEASURED
accessibility_signals: OBSERVED_NOT_PROMOTED_IN_THIS_CONTRACT
external_contact: BLOCK_UNTIL_HUMAN_REVIEW
merge: NOT_AUTHORIZED

The PR remains draft. The technical evidence pack and the external collaboration brief are ready for human review.

[safal207]

Codex exact-head review

Reviewed all 9 changed files on exact head e764606889ea9871ea1f0ccd65a17599f374a42a without relying on CodeRabbit.

Confirmed

• raw audit workflow: success (29897326377);
• rendered audit workflow: success (29897326392);
• repository CI: success (29897326448);
• Security Audit: success (29897326425);
• contract is sequential, unauthenticated, evidence-only, and keeps external submission / deployment / merge authority disabled;
• source-to-rendered adjudication correctly removed the two over-broad claims.

Blocking boundary gap

Both probes enforce the canonical origin only after navigation completes:

• Python uses the default urllib redirect handler, then compares response.geturl();
• Puppeteer runs page.goto(), then compares page.url().

If an allowlisted Bell Integrator URL later redirects off-origin, the external document is fetched before the run is marked unbounded. That conflicts with the stated public allowlisted pages only / bounded-origin contract.

Required correction

Fail closed before following a top-level off-origin redirect:

1. Python: use a custom HTTPRedirectHandler that rejects any redirect whose canonical origin differs from https://bellintegrator.ru, with a regression test.
2. Rendered probe: intercept main-frame navigation requests and abort an off-origin redirect before loading it; record the blocked URL and force NEEDS_EVIDENCE.
3. Re-run raw, rendered, CI, and Security Audit on the new exact head.

Codex verdict: CHANGES_REQUIRED for the audit mechanism. The four documented product findings remain supported by the current exact-head evidence; this blocker concerns future rerun containment, not the findings themselves.

[safal207]

Codex exact-head follow-up

The redirect-containment blocker from the prior review is resolved on exact head 56db55f3aa54748f27bd16f174916bc1eaf02fab.

Implemented

• Python now uses a bounded HTTPRedirectHandler and rejects cross-origin or HTTPS→HTTP redirects before following them.
• Regression tests cover same-origin allow, cross-origin block, and downgrade block.
• Puppeteer now intercepts main-frame navigation requests and aborts off-origin destinations before document content loads.
• Blocked navigation is recorded and forces the rendered result to remain unconfirmed / NEEDS_EVIDENCE.

Exact-head validation

• Bell Integrator Outside-In Audit v0.1 — success (29928835464)
• Bell Integrator Rendered Audit v0.2 — success (29928835323)
• CI — success (29928832696)
• Security Audit — success (29928832858)

Codex follow-up verdict: the mechanism blocker is closed. The four documented product findings remain supported. PR may remain draft only for the human collaboration/reporting decision; no CodeRabbit review is required.