Independent publication review: discovery ranking probes

Reviewed on 22 September 2026 by the dependency-review agent, read-only, in /Users/crpage/tmp/okf-dwp-discovery-cards-pilot, against source 7412d1d02e023794ebb1d5a5f385f08f59be7ee2. Reviewed HEAD was 401a77a38a11ff2f24224dff3f07003ac78517a0 plus the six documentation/backlog changes present during review. This is a bounded publication/confinement review; the separate retained arithmetic reviews are not represented as a new duplicate evaluation.

Result

No blocking finding for publication as research. No approval of runtime adoption, relevance, answer quality, legal completeness, affordability or summary-card effectiveness is implied. One minor reproduction-documentation clarification is recommended below.

Checked boundaries

• The 18 changed paths consist of two probe scripts, new protocols/results/reviews and documentation/backlog updates. No changes from the source baseline occur under frozen source, knowledge, domain profiles, logical units/context, combined context, pilot/full-DMG outputs, prior logical evaluations, staff questions or semantic-coverage records.
• Both protocol files still exactly match their pre-implementation Git commits (f00d299a and e1654a4f). Both retained runner copies exactly match the corresponding script. The length runner SHA-256 also matches the independently reviewed identity.
• The runners use local source reads with relative-path validation, realpath equality, file-size bounds, exact length/hash bindings and bounded gzip decompression. Imported engine modules are verified against the four admitted module hashes and their Git blobs before import. Those modules have no unchecked runtime module dependency beyond that set. No fetch/provider call is introduced: the alias baseline receives a confined in-memory local fetch adapter.
• Output is a fixed additive research leaf, with refusal if it already exists; the commands do not replace production indexes, source units or existing observations. The alias runner checks output existence late, so an accidental repeat can perform unnecessary work before failing, but it still preserves the prior output. These are local research runners, not an untrusted-input production service.
• Public prose accurately separates candidate ranking from graph expansion, dependency closure, final budget selection and AI answering. The failed alias variant, unchanged cases, unresolved boundaries, citizenship false positives and historical-source ambiguity remain visible. The 40 cases are explicitly known development cases, with a duplicate identified; no held-out generalisation claim appears.
• All three public ranking counts have the correct described grain: earlier location overlap and candidate-text bytes. The material explicitly rejects treating these as answerability, relevance or preserved-qualification scores. Arithmetic itself was previously independently recomputed in the retained coverage-agent reviews; this review checked their stated scope and identities rather than rerunning them.
• Changelog, public experiment note, work log, Markdown backlog and machine backlog agree: discovery-ranking-probes is completed research, while discovery-cards remains unimplemented. Summary-card research and live ranking adoption are not marked complete. Separate protected publication remains required.
• git diff --check 7412d1d0 passed. No provider/acquisition/evaluation calls or repository changes were made during this review.

Minor reproduction clarification

The published commands accept only --explorer-root and write to fixed evaluation/discovery-probe/run or length-run directories. Those directories are tracked in a fresh checkout of this research increment. Therefore “use a fresh output leaf” or “clean checkout” alone will deliberately fail the preservation guard. State that reproduction takes place in an isolated disposable checkout with the existing relevant run directory moved to a separately named retained-history location before invocation; for the length probe preserve run/comparison.json at its original path because it is a bound input. Do not advise deleting or overwriting an existing receipt. Alternatively, a separately reviewed output argument could be added later, but changing the already executed runner is unnecessary for publishing this research.

Review scope limitations

This review is not a new arithmetic replay, full security audit, source-PDF layout review, protected-main CI result, public Pages observation or model trial. It endorses honest publication of the observed diagnostics and their limits. Production retrieval and summaries need the separate planned experiment and acceptance evidence.

Reviewed file identities

• CHANGELOG.md: 71f3f95d35a5a747bac83ae78f2281d7212abf7ff7113655110139f11665e10e
• docs/backlog-work-packages.md: 7cad9bd5b2f5da2d7c0a0d6f0cf418828d49671f2ac07de2f1ba5c1aa91c0c68
• docs/backlog.md: 15c652087e9e600e1d83c4b8b3ba9e7d282225aee45738f345894aba72de3490
• docs/discovery-cards-experiment.md: e426b3fe7874593719ec1d66552f1eae5ad6c37597218c6806efe0e08325aac5
• docs/logical-units-work-log.md: bdadedca6aeb1a599672508b1ee2bc7dfaddf5bc70155d324d0d6908960143b0
• evaluation/backlog.json: b70513e190f06c990d9d9f4d494d5ac290c55a92c3e30bdfe1f223187a3bc203
• evaluation/discovery-probe/README.md: 4d8cc70e66f19f10291b7ec40f18cfad1dcbafddf49bc40a4e6ab71c381b739c
• evaluation/discovery-probe/independent-review.md: 3a55c7aa78b85f3c0ac1c9d687321cbc79011d5dfddec6318bb798ed9aeb194f
• evaluation/discovery-probe/length-independent-review.md: 9537fb92826012bca66959aae865a65a7fcff7173a5291ac771b889a4ceb11c6
• evaluation/discovery-probe/length-protocol.json: 9fa98082efbf93596f5c3eab55f41a4ad9bbd10c9fb98163522cb33fda09c95d
• evaluation/discovery-probe/length-ranking.md: ce715335dd97c964712f18a4d5effd81435d757199c6e5a41d74c138fa8ee596
• evaluation/discovery-probe/length-run/runner.mjs: b24442457c64926d3662ebaaa49b1e32714f75e6acc667a003965a497d8ad76f
• evaluation/discovery-probe/protocol.json: f4dab58f1a0559a1e0a7063242f2b12f32fd5bf33d5e4761e3faa50a097897b3
• evaluation/discovery-probe/run/runner.mjs: ebb976a1f398692c308bd34c23639d2e80a80d34aff4e8985aaa1eab06262ece
• scripts/probe_discovery_length.mjs: b24442457c64926d3662ebaaa49b1e32714f75e6acc667a003965a497d8ad76f
• scripts/probe_discovery_terminology.mjs: ebb976a1f398692c308bd34c23639d2e80a80d34aff4e8985aaa1eab06262ece

Follow-up: reproduction clarification resolved

Reviewed commit 6acb36d181200c0708c52e1cb10adf4bd69a6a85. Both reproduction sections now explicitly require isolated scratch checkouts and move only the relevant existing output directory to a unique mktemp archive. The length run retains its hash-bound alias baseline in place. No script, protocol or result changed. The minor reproduction-documentation recommendation is resolved; no publication blocker remains within this review scope.