Bounded evidence packages: implementation status

Assessment date: 26 September 2026. Code baseline: Explorer c80c2c696aa554a4879a6c85fc2ffcfa0c5112b7; DWP 5d2f083202e60edac5ec395dc27fc6e7af285e5b.

The useful core is implemented; the complete research proposal is not. A bounded evidence package (BEP) is a limited collection of evidence, reasons for selection, source identities and visible gaps. It is not an AI answer. Publishing the BEP research does not adopt its proposed schema or make its experiments production acceptance tests.

What is implemented

Capability	Repository evidence	Limit
Shared, deterministic context assembly	Context engine and types	Uses `okf-governed-context.v1`, not the reconstructed research envelope.
Provenance, authority, directed paths, declared requirements and explicit gaps	Context types, A–H evaluation method	Missing evidence stays missing. A requirement can only be checked when a producer has declared it adequately.
Bounded selection and evidence status after trimming	Context engine, regression tests	Retrieval and assembly limits are separate; either can prevent a complete result.
Read-only remote access using the same engine	MCP service	Client support and observed deployment are separate from the existence of code.
Compact catalogue, exact evidence slices and browser replay link	`ask_okf_manifest` and `read_okf_evidence` in the service	Follow continuations with the same source, engine, budget and context identity. Verify the digest; a partial slice may omit qualifications. A replay link is not a stored audit log or an answer.
Human inspection and source-linked reading help	Evidence workbench, reading help	Displaying source evidence does not establish legal answerability. Reading help is a separate producer capability.

These are code and contract findings. Tests cited above are evidence of intended coverage, not a new whole-corpus or live-client trial. For dated deployment and client receipts use the service publication record and its shared DWP reference. Do not replace that receipt-backed record with an unqualified claim that a service is currently deployed or callable in every host.

What has not been adopted or demonstrated

Research schema migration: the reconstructed bep-core.v0.1-reconstructed.1 profile and its package_id are a separate, explicitly incompatible proposal. Production uses context_id. No reviewed, versioned adapter or information-preserving migration is recorded. See the migration proposal.
Cross-runtime canonicalisation: production has deterministic serialisation for its current identity domain. That is not proof of conformance to a new canonical JSON standard or proof that independent implementations generate the same identity. Research test vectors do not supply that production proof.
Alternative storage architecture: the research discusses database and workflow options, including PostgreSQL and DBOS. Comparative storage and serving benchmarks remain proposed. Adopting those products is conditional, not a missing prerequisite for today’s file-backed service.
Restricted-data assurance: this exemplar is designed for public evidence. It does not establish a reviewed access-control or retention system for confidential claimant data.
Complete domain evidence and answer quality: source capture, package reconstruction and transport success do not discharge unresolved legal or question-scope obligations. The dated DWP source-led results still record insufficient packages and open obligations. Those historical results are not silently upgraded by later reader improvements.

What to do next

For broader DMG and ADM reading help, improve source-bound passages, local reference linking and scoped explanations first; a wholesale BEP rewrite is not required. Use the existing governed context and delivery contracts while testing new producer annotations independently.

If the research profile is worth adopting, first write a small contract-mapping matrix and freeze migration fixtures. Add an adapter alongside the old contract, prove that provenance and limitations survive round trips, then evaluate clients at fixed evidence and equal budgets. Do not rename the current format and claim research conformance. Keep storage experiments and legal acceptance as separate work items. Jev-Mem is deferred until after the presentation.