Ask OKF service changelog

0.7.0 public observation — 25 September 2026

• Bind the service guide to the retained hosting and public SDK receipts: 12 source-bound cases and 135 bounded HTTP requests. Keep the older 0.6.0 receipt and separate public transport from client and legal acceptance.

0.7.0 — Evidence Connect source admission

• Pin the immutable Evidence Connect corpus and descriptor from DWP commit 7eeded763042ddd0070f4fed834c6074149e8e2f, and a compatible v3 engine from Explorer commit d6930bbcddaab616deec002d9e6efff6e3aae953.
• Make Evidence Connect the default source while retaining all earlier source versions, frozen engine pairs and engine-specific replay identities.
• Make new questions on the review page use the assembler approved for the selected source version, including historical sources.
• Bind discovery and relationship assets to the manifest allow-list and shared per-request resource limits. Verify actual local source text and cursor rejection without making model calls.
• Update the landing page, tool version description and public verifier for six sources, three engines and ten permitted pairs. Deployment and connected client acceptance require separate observations.

0.6.1 — complete questions in connected clients

• Anchor the shared question pattern so both substring and whole-string regex matching accept the same non-blank text. The original \S matched a single non-whitespace character; a client applying a whole-string match rejected ordinary questions before contacting the service. A one-character control succeeded. The fix keeps types, lengths, budgets and evidence selection intact.
• Cover all three tools, Unicode, multiline questions, blank and overlong inputs, and equivalence with the original accepted set. Preserve all 0.6.0 receipt bytes and validate their own historical build and runner.
• An installed connection must refresh its tool metadata after publication. A new session needs the plugin selected. See the connection guide and the shared publication record below; code and local tests do not establish client access.

<!-- ask-okf-publication:start -->

Recorded public deployment: 0.7.0

For the latest recorded deployment and verification, use the shared DWP service publication status. This dated observation is not a live health check.

On 25 September 2026, the hosting record records service 0.7.0 as deployed. The separate public SDK observation passed 12 evidence cases and 135 requests, reconstructing complete packages from bounded reads. It recorded 11,974,541 received bytes, no automatic retries and no model calls.

The hosting record and public health report have different scopes: health does not independently attest hosted Worker bytes. Delivery checks do not establish complete legal evidence, specialist acceptance, answer quality or compatibility with a particular ChatGPT, Data agent or Voice client. This observation includes no new public browser journey. Earlier failures and observations retain their own scope. <!-- ask-okf-publication:end -->

0.6.0 — partner and qualification preparation

• Pin final combined source 723bcc5b015ab38a026625c2148edbd784edf7c7 by exact manifest bytes. Retain 0.5.0's four sources and both archived engines; allow the new source only with the current c4f engine, giving nine pairs.
• Preserve historical packages and prior observations. Add ten actual local adapter/SDK cases, including the care-home question and an unknown-term control at 512 KiB. Four earlier complete-package hashes reproduce exactly.
• Add explicit current question/budget identities to the successor compact live verifier. It remains opt-in and bounded. The later public observation above does not establish model acceptance. Keep all 203 staff obligations open.

Earlier engine-pinned replay preparation

• Retain exact c4f and b9 assembler modules with build-verified manifests and explicit compatibility for the four existing approved source revisions.
• Add engine identity to replay recipes, catalogue/read envelopes and health. Preserve historical v1 package bytes and the expected context-ID guard.
• Reconstruct old engine-unspecified links using at most two sequential bounded attempts. Compare complete package bytes; keep originating-engine uncertainty visible and fail closed for unavailable or ambiguous replay.
• Share file, transfer, decoded-work and deadline bounds across attempts. Keep read-only tools, inert source data, privacy and CSP boundaries.
• Clear stale evidence when a replay fragment changes in the same tab. Preserve engine/context/package identity through catalogue and evidence pagination.
• Retain separate local two-engine/four-source and Chrome observations. Earlier observations remain unchanged; package version and public deployment are not promoted by this candidate.

0.5.0 — household and statutory evidence candidate

• Pin the default combined corpus to 3ef0e786e9a18e76fa17c7d925ff509d6d6c9f84: 901 semantic records, 1,427 assertions and 20 selected statutory units. Keep machine extraction, legal applicability and specialist acceptance distinct; all 203 obligations remain open and the demonstrated household context remains insufficient.
• Preserve the preceding 9de52acf1db84b27f8933d80480eaa850e74fa33 staff corpus, earlier discovery corpus and original custody profile as three explicit historical versions. Add a separate verified staff manifest and version-routing, manifest-swap and package/lock/build identity controls.
• Verify all four versions using the official SDK 2 and SDK 1 clients against exact local immutable DWP Git blobs: 108 files hash-verified. Retain a new integration receipt and the earlier observations unchanged. Preserve the original custody context identity; each other receipt binds its actual engine and version rather than promising blanket cross-engine byte compatibility.
• Exercise compact replay for the current self-funded care-home question at 256 KiB, prior-staff Child DLA/PIP, discovery-corpus abroad and historical custody. The current care-home package includes 35 records and 50 relationships. Keep the existing catalogue/reader schemas and primary receipt field names.
• Extend the remote verifier to seven full and four compact cases. Pace service HTTP requests by at least 750 ms, record counts and spacing, and disable automatic retries without changing production rate limits or dependencies.
• Link the landing page to the public beginner learning path. Retain all existing read-only tools, request/source bounds, portable builds, CSP and no-transform headers. Deployment, browser, ChatGPT and Voice checks remain separate gates.

0.4.0 — combined staff evidence candidate

• Make build output and receipts independent of real or symlinked locked dependency installations; bind the build script and test both layouts after relocation. Preserve the original local observation and rerun integration against the corrected build without weakening receipt checks.

• Prepare the combined DMG and ADM corpus as the default, including the proposed staff-task concepts, evidence requirements and explicit unresolved obligations. Pin the source to 9de52acf1db84b27f8933d80480eaa850e74fa33; the release build rejects an unpinned candidate or altered manifest.

• Preserve the earlier bf50ef8d91b9f1ccc2cbdb354198eae74c9ed752 full-source discovery corpus and efb05c66616a9cd4328a86cf412780fe7bc7cf0b custody profile as explicit versions with separate identities and integrity checks. Add manifest-swap, altered-byte and version-routing regressions, a real three-version loader test and a retained hash-bound local corpus replay receipt. Keep the architecture notes aligned with all three read-only tools.

• Extend the live verifier with a 256 KiB staff Child DLA/PIP context and compact replay across current, earlier corpus and original custody versions. Preserve the primary browser-receipt fields and existing historical observations.

• Send Cache-Control: no-store, no-transform on HTML only while preserving CSP. Other responses keep no-store. This follows the documented Cloudflare JavaScript Detections behaviour. Add route-level header regressions. Live transformation and cookie checks remain separate; no hosting issue is declared resolved by the candidate.

• Retain earlier release receipts. Source tests and local builds do not establish deployment, ChatGPT acceptance, Voice access or specialist approval.

0.3.1 — replay-link correction

• Remove a displayed replay link when the question or source version changes, or a new replay starts. A link created for an earlier context could otherwise reappear beneath a later context. Regenerate the link only when requested for the current verified context.
• Add a browser regression that recreates two different questions and checks the second link's question, version and context identity, plus repeated replay and version-change invalidation.
• Preserve 0.3.0 observations. Deployment and acceptance of 0.3.1 require their own build, hosted-service and browser receipts.

0.3.0 — compact evidence delivery

• Add bounded evidence catalogues, exact evidence reads and an inspectable browser replay page alongside the existing full-package tool. Context assembly and evidence authority remain separate from delivery and AI answers.