# PFABC release readiness ## Release 0.4.2 — September 23, 2026 `HARDENING_INTEGRATED / SOFTWARE_CHECKS_PASS / EXPERIMENTAL`. Record contract **0.1.2** adds a nesting guard in the shared JSON parser, faster quotation-prefix diagnostics, visible measurement values and custom threshold qualifications. Origin, acceptance and override authorship are submitter declarations; no identity or human review is authenticated. Imports preserve their declarations and display that limitation. Raw engine heuristic annotations retain their original values with machine-readable qualifications about bridge identifiers and quotation membership. Verification: 51 Python test methods pass, the original six mutations and six new deliberate regressions are caught, and the actual local Cloudflare runtime passes 31 bounded service checks and six existing transport groups. The bounded checks use modest parser-limit fixtures and count substring probes; the reported stress-sized crash inputs were not replayed. Original engine and formal specification hashes remain unchanged. The prior archives remain preserved. Markdown destinations in served documents are now checked, and private evidence paths below are labelled local-only rather than linked. Check the [deployed version and availability](/api/health). Local software verification and external deployment/readback are separate records. A failed HTTP request is not an assessment. Browser and ordinary direct-client access are supported paths; default Python urllib has previously been blocked by Cloudflare before reaching the application. Results remain conditional on the supplied declarations and chosen rules. No consciousness measurement, authentic source provenance, calibrated threshold or comparative scientific benefit is established by these software checks. ## Release 0.4.1 — September 23, 2026 (historical) `RECORD_CONTRACT_REPAIRED / SITE_INTEGRATED / SOFTWARE_CHECKS_PASS / EXPERIMENTAL`. The main workbench starts with three inspectable, constructed records: a self-report, a bounded task result and a contradictory measurement. People can change the question and theoretical assumptions, import a structured record, and download the complete input and assessment. No private conversation or provider account is required for the examples. The checked route is `POST /api/records/evaluate`, contract **0.1.1**. It checks record shape, exact quotation membership, declared object scope, accepted observations, theory licensing and conflicting declarations before evaluating with the preserved engine. Malformed records receive a structured rejection; contradictory boolean assignments and obsolete overrides are rejected; the precaution answer includes the engine's independent precaution warrant with bridge states shown separately. Source provenance, inherited scope and opaque test references remain declarations, not authentication. Verification for this release uses synthetic records: 37 Python test methods, all six deliberately introduced mutations caught, all 15 independent repair cases satisfied, six existing hosted regression groups, and 28 checks through the local Cloudflare runtime. Browser checks cover the three cases, theory changes, invalid input, stale-result hiding and 320/375 pixel layouts. An actual assessment download was parsed and its input digest verified. The twenty original engine members and formal specification retain their source hashes. Check the [deployed version and service availability](/api/health). Deployment and external-service verification are recorded separately from these software checks. This hosted service runs on Cloudflare; it does not require the owner's Mac to stay online. The original advanced engine and paired workspaces retain their original semantics. They are labelled separately and do not inherit the record checks. Existing source archives are preserved; the new download is v0.4.1. Scientific validity and benefit over a strong plain evidence table remain unestablished. No participant study, provider connection or consciousness measurement is claimed. Known access limit: some automated clients, including default Python urllib, receive Cloudflare filtering before reaching the application. Browser use and normal direct HTTP clients are separate checks; universal agent compatibility is not claimed. See [API instructions](docs/HOSTED_API.md) and [privacy](docs/PRIVACY.md). An HTTP failure is not a PFABC assessment. The earlier implementation-not-started and integration-hold markers below are historical and superseded by the verified repairs described here. ## September 23 independent successor review — historical `IMPLEMENTED_LOCALLY / REVIEW_CHANGES_REQUIRED / SITE_INTEGRATION_HOLD / VALIDITY_NOT_ESTABLISHED`. The typed-record successor exists in `native/record_contract/`. Independent review reproduced the author's 11 tests, all six mutation detections, six existing native semantics tests, the existing release check and equality of all 351 stored JSON envelopes. Python 3.11.14 was used for this review. The original engine hashes match and all 41 frozen review inputs remain unchanged. Four repair findings block integration: malformed records escape as exceptions; contradictory homecoming/depth values are silently overwritten; overrides can act on stale conflicts and remove remaining evidence; and the precaution headline omits a non-bridge monitor warrant. Fifteen independent constructed cases record five requirements met and ten unmet across these four findings. These targeted counts are not a scientific error rate. The implementation, author's spec/corpus and original evidence were preserved during review. Exact next action: repair F1–F4 and clarify provenance/conditional-support semantics, verify against both case sets, then build the worked-case page. Controlling review and runnable evidence are local-only and are not published: `evidence/record-contract-review-2026-09-23/REVIEW.md`, `evidence/record-contract-review-2026-09-23/INDEPENDENT_RESULTS.json`. The prior `IMPLEMENTATION_NOT_STARTED` status below is superseded. Existing public release checks do not cover the new contract; no remote deployment was refreshed or changed in this review. Scientific benefit over a strong plain evidence table remains unestablished. ## September 23 evidence review and rework decision `LOCAL_REPLAY_VERIFIED / GUIDED_REWORK_JUSTIFIED / IMPLEMENTATION_NOT_STARTED`. Two supplied Claude-report ZIPs were compared; `(1)` adds exploratory wording checks to the same main run. Both manifests verified. Independent local replay matched all 351 stored sweep envelopes and the reported outputs of 66 wording packets. These are tests of PFABC, not independent consciousness experiments. The 22 earlier source hashes and submitted guided-JavaScript snapshot match current source. The guided text/declaration distinction is now more precisely demonstrated: ordinary self-report changes can leave outputs identical, while incidental keywords and examiner questions affect some annotations or blocking outcomes. An absent quotation also retains the primary evaluation at the native boundary. The prior contradiction and validity limitations remain unresolved. Selected rework: worked investigations and method explanation as the public entry point, followed by an optional structured case workspace and preserved advanced lab. Agent submissions should be inspectable records with separately labeled checks/reproduction, not self-assigned levels or endorsement counts. Exact next local action is the typed-record/contradiction contract and first corrected worked-case page specified in the local-only `evidence/claude-report-review-2026-09-23/REWORK_BRIEF.md`. Evidence and limits: local-only `evidence/claude-report-review-2026-09-23/ASSESSMENT.md`. No engine/site implementation, public deployment or account change occurred. Scientific effectiveness remains unestablished; the September 8 matched-baseline research question below remains active as a proposal, not a completed study. ## Scientific direction — September 8, 2026 `METHODS_RESEARCH_SELECTED / VALIDITY_NOT_ESTABLISHED / GUIDED_DEVELOPMENT_PAUSED`. The user's project-led reassessment supersedes the earlier newcomer-pilot and website-successor recommendations. The selected question is whether a bounded PFABC method improves evidence-to-claim assessments over a strong plain table, while preserving supported positive conclusions. Consciousness-related claims are the initial domain; consciousness itself is not a reference label. Source review, targeted primary literature and 17 local synthetic semantics cases support narrowing the work. They do not establish method effectiveness or novelty. Declared passing states can override conflicting measurements; other output fields still warn of unsupported promotion. Mode-fixed metaphysical statuses express policy rather than discoveries about an AI. Original studies, engine and specification remain preserved; 22 diagnostic source hashes match. Controlling assessment: local-only `evidence/scientific-assessment-2026-09-08/DIRECTION.md`. Exact next object: local-only `evidence/scientific-assessment-2026-09-08/NEXT_OBJECT.md`: an explicit executable treatment/contradiction contract and small unscored reference cases, before any comparison sampling. No new study, public deployment or account change occurred. No advantage over a competent baseline is established. The earlier one-case product result (local-only `evidence/usefulness-case-2026-09-08/RESULT.md`) remains evidence about that guided adapter, not a verdict on all PFABC research. ## Public release — last verified September 7, 2026 **Status: `PUBLIC_BETA / AGENT_ACCESS_GATED / PAIRING_UI_CHECK_PENDING`** September 7, 2026. Website v0.4.0-rc2; original engine 0.5.3+review1; conversation rules 0.2.0-rc1; session schema `pfabc-clarity-session/0.2`; source-linked guide `guide/0.2.0-rc1`; formal specification v0.5.3. Public site: https://pfabc.xn--arkh-eva.org/ Known hosting limitation: Python urllib's default client signature received Cloudflare HTTP 403 / error 1010 before reaching the application. Browser use and the documented direct-client smoke tests passed, but general agent-client compatibility is not yet established. An owner-side, PFABC-host-only Browser Integrity Check exception is saved as an inactive draft. Its activation awaits specific approval; no active security setting was changed. The guided workbench and structured lab now call the original Python evaluator through a Cloudflare Worker. An HTTP-capable agent can evaluate a packet directly or collaborate in an optional isolated 30-minute workspace. Ordinary evaluation is stateless. The offline CLI/server, local conversation notebook and fictional practice exercise remain available. No provider account is required. Local Cloudflare-runtime checks passed for exact native envelope parity, UTF-8/raw-input digests, invalid input, size boundaries, workspace isolation, owner/agent permissions, unsuccessful-input retention, simultaneous writes, deletion, and real shortened expiry/alarm. See [verification](VERIFICATION.md). The twenty original engine files and formal specification retain their source hashes. The public custom domain passed HTTPS, native-result parity, workspace isolation, role permissions, deletion and negative-input checks. Guided and advanced browser evaluations passed. The shared-workspace consent flow remains partially verified. Both labs were checked at 320 pixels without horizontal overflow; see verification. The result is conditional on the declared packet, chosen theory and operating mode. PFABC does not authenticate source text or measurements, inspect hidden model state, or determine consciousness. Its scientific framework and thresholds have not been empirically validated by this work. A successful software test is not an endorsement of an empirical claim. The guided page credits only the supplied quotation's presence in the record. Its constrained evaluation statement is separate from the user's own proposed interpretation. Stronger questions carry no invented controls or measurements. The advanced engine accepts user declarations and does not semantically verify every sentence of a free-text claim. No human usability pilot or comparative learning study has run. Release as an experimental clarification and research tool can invite feedback; broad learning benefit remains unproven. The [proposed pilot](docs/PILOT_PROTOCOL.md) is preserved but is superseded as the immediate next action by the September 8 scientific assessment above. The owner authorized Cloudflare hosting, and the website is publicly deployed. No GitHub repository has been created by this work. The release uses Apache-2.0 with separate identity and release-scope boundaries. The upcoming safety and alignment project remains a separate contribution, not an implemented integration. Resume: `wrangler.toml`, `native/worker.py`, `docs/HOSTED_API.md`, and the explicit asset builder `tools/build_public.py`. The downloadable candidate is built by `tools/package.py`; previous archives remain preserved. Source provenance is recorded in `native/UPSTREAM.json`.