Inspect discovery and delivery separately
This independent experimental candidate addresses the reported truncation in four layers: question routing, bounded retrieval, whole-item assembly and host delivery. It supplies evidence for engineering review, not benefits advice, an award calculation or an individual entitlement decision.
Measured source-retrieval result
The additive graph candidate passes the agreed declared-source gate. The retained evaluation records 263/263 persona target occurrences, 35/35 original custody targets and 4/4 UC targets, including their required directed source paths. Every selected package arrives intact through responses whose complete duplicated MCP/JSON-RPC envelope is at most 8,192 bytes. There are no assembly omissions in this run.
The persona coverage ledger
covers all eight authored personas and all 21 declared interest areas. The
75-case protocol contains 62 positive retrieval cases, ten boundary controls,
two UI checks and one integrity check. All original 40 staff question occurrences
remain positive retrieval cases with unchanged wording and targets. Five broad
supplementary stage-2 prompts retain their initial wording as boundary controls.
Four have separately documented benefit-framed positive pairs; stage2-005
remains boundary-only. Their
declared source targets have not been dropped. UI and specialist checks retain
their own pending status rather than being counted as successful retrieval.
The first failed run is retained. After freezing the graph, runtime and client, a separate agent ran 24 withheld wording cases: 45/45 targets in 12 positives and 12/12 negative controls passed. The receipt binds the protocol and frozen bytes. This is a bounded withheld wording trial, not an independent human blind benchmark or proof of complete legal reasoning.
Independent review then tightened concept-conflict admission, made resolver
budget exhaustion explicitly incomplete and increased the collector's bounded
session from three to ten minutes. The 30-second call and 8 KiB response limits
remain unchanged. The largest measured package needs at least 331 calls; a
virtual-clock regression verifies complete delivery at 750 ms per call, beyond
the old three-minute ceiling. A separate
post-review replay
passes the same 45 targets and 12 controls under the corrected identities.
It is a regression replay, not a fresh blind trial. The original trial remains
byte-for-byte replayable from producer commit
5516e6c604793a5c98c57a3ddb9c3c13f46afe64 through the disposable-tree harness.
How graph navigation helps
domain-profile/retrieval-graph/*.yamlld compiles into the separate
retrieval-graph/graph.json. It preserves frozen evidence and requirements,
adds 28 source-bound directed navigation repairs, 11 guarded compositional
concept bridges and three topic navigation profiles. These additions are
explicitly model-derived, unreviewed navigation hypotheses. They do not turn
mentions or graph reachability into legal applicability.
The runtime receives the research question, never the evaluation's target IDs. It resolves concepts, applies compound-word and ambiguity guards, follows scoped directed dependencies and checks each requirement from its own active root. An unrelated route cannot satisfy another requirement. The graph, resolver, runtime and client are hash-bound. Missing paths and sources remain explicit; hard work, byte and deadline ceilings cannot be raised by caller input.
plan_okf_evidence returns a bounded catalogue and diagnostics;
read_okf_planned_evidence supplies contiguous verified parts. The collector
checks every continuation identity, the complete package hash and each source
record's text hash before reporting complete delivery. The prepared service
exposes these two tools at /mcp; the historical three-tool connection is
separate at /okf/mcp. Deployment and actual host invocation require their own
receipts.
The diagnosis and strategy comparison explain why a larger package alone misses the required source. The candidate preserves that baseline, optionally assembles a separate compound-alias projection and makes at most three source-supported navigation queries. These are unreviewed research proposals. Each result keeps its own source identity, query, passages and omissions; they are never merged into a sufficient answer.
Run the inspector
Use Node 26.7.0 and the separately pinned Explorer checkout
40e659408485a7dfa0e8cb51ab27c1b0f890bc3c. No installation or model key is
needed for this Node workflow. For repository Python checks, use
uv sync --locked and uv run --locked python; do not upgrade dependencies.
node scripts/run_context_review.mjs --explorer-root /path/to/okf-explorer
node scripts/run_context_review.mjs --explorer-root /path/to/okf-explorer --check
Open evaluation/truncation/review/index.html in Chrome or Edge. Compare the
graph-guided plan, baseline and each recovery context, search for a passage, inspect its official
PDF reference and read the retained diagnostics. The HTML is generated: change
the producer and regenerate it rather than editing it. The report runs offline
with no external scripts or network requests. Following an explicit source link
opens that reference separately.
For a different general research question, use a separate output directory:
node scripts/run_context_review.mjs --explorer-root /path/to/okf-explorer \
--question 'Which public guidance discusses Pension Credit abroad?' \
--output-dir /tmp/okf-dwp-abroad-review
Use public research questions only. Do not enter names, claim details or other personal claimant data. A question that activates no recovery route still produces a baseline report. The bounded policy is designed around observed question dimensions; it is not a general legal reasoner.
Delivery and report contract
The runner calls the graph runtime and pinned legacy engine through local manifest/read adapters. It does not call the deployed MCP service. The progressive client collects every catalogue page and contiguous UTF-16 package slice, then checks the complete context digest, record text hashes, source selection and diagnostics. Resource and deadline limits have immutable ceilings. An individual JSON value and a duplicated MCP envelope have different sizes; local response measurements do not establish the host's clipping threshold.
report.json uses okf-dwp-context-review.v1. Its root records question,
status, implementation and source identity, recovery_plan, contexts and
limitations. Each context records:
id,label,queryandcontext_id, preserving its independent identity;evidence_status,assembly_truncated,retrieval_truncated,selected_records,relationshipsandpackage_bytesfrom the engine;delivery_complete,package_complete,package_sha256,deliveryandreceiptfrom successful progressive collection;- exact
recordswith authority, review status, provenance, selection reasons and paths, plus complete retaineddiagnostics.
The per-context JSON collections retain the original selection wrappers,
relationships, catalogue and delivery receipts. The report binds the policy,
runner, recovery adapter, progressive client, renderer and engine bytes.
--check recomputes expected artefacts and compares their bytes without
rewriting them. Completion flags are derived from verified collection, never
from the presence of a receipt. The HTML is an inspector for this trusted local
output; it is not a verifier for arbitrary third-party reports.
Builds require a fresh or empty output directory and preserve earlier reports.
Use a new --output-dir when implementation inputs change; keep the preceding
result as historical evidence. --check also rejects unexpected files and
symlinks in the declared output tree.
Complete delivery means that the selected package arrived intact. It does not
mean that retrieval selected every required source or that a passage applies.
All candidate contexts retain insufficient and ai_answer: null. All 40
original staff tasks and 203 unmet obligations retain their historical status.
Why evidence remains insufficient
The inspector groups reasons while preserving every original code, message, record and requirement identifier. Missing evidence in the working index does not by itself establish that the source is absent from the acquired corpus.
| Category | Lack or uncertainty recorded |
|---|---|
| Source coverage | Required passage or dependency has not been established in the declared evidence set. |
| Retrieval | An available record was not selected or a bounded search was incomplete. |
| Graph connections | A required directed dependency or supported source route is missing. |
| Question scope | Benefit, period, jurisdiction or other research dimensions are unclear. |
| Legal applicability and version | Relevant law, territorial scope, commencement, amendments or case law are unresolved. |
| Source interpretation | Extraction, passage boundaries, references or authored meanings need verification. |
| Specialist review | No qualified human has accepted the interpretation and dependencies. |
| Assembly and delivery | A selected item was omitted, clipped or not reconstructed intact. |
| Access | A source could not be accessed under the applicable boundary. |
The preserved 203 obligations comprise 43 evidence-closure gaps and 40 each for applicability, legal version, specialist review and question scope. Successful source retrieval does not silently close them. The 43 evidence-closure gaps record unverified completeness of the required evidence chain; they do not establish that 43 source passages are missing from the corpus. The statutory catalogue crosswalk is a separate source-identity aid; metadata alone supplies no provision body, current-law assurance or case law.
Repeat the evaluation
The statutory crosswalk
matches 20 work identities and inventories 53 provision targets: 13 existing DWP
text projections, 38 metadata-only observations and two targets with only work
metadata. It binds the July catalogue snapshot at legislation commit
9830fde2d32052f10713298f16b8ab1e321b1928 and unchanged DWP input blobs at
baseline 77ef16047415ee48dcb93b80adf1ba057ea00d9b. All case-law bodies remain
excluded; existing case-law mentions and unresolved citation notes are retained.
There are no new source acquisitions or closed obligations. Selected work
identity and catalogue version metadata reduce uncertainty without establishing
provision completeness, commencement, extent or temporal applicability.
For crosswalk checks, use a separate read-only legislation checkout at that exact commit. CI checks out only its bound catalogue and fetches the DWP input baseline. The check does not use a moving branch or treat a catalogue work as a provision body:
uv run --locked python scripts/build_legislation_crosswalk.py \
--legislation-root /path/to/okf-uk-legislation --check
uv run --locked python scripts/test_legislation_crosswalk.py \
--legislation-root /path/to/okf-uk-legislation
OKF_EXPLORER_ROOT=/path/to/okf-explorer node --test \
scripts/test_context_recovery.mjs scripts/test_progressive_evidence_client.mjs \
scripts/test_context_review_html.mjs scripts/test_run_context_review.mjs \
scripts/test_current_explorer_compatibility.mjs
node scripts/check_current_explorer_compatibility.mjs --explorer-root /path/to/okf-explorer
node scripts/evaluate_context_strategies.mjs --explorer-root /path/to/okf-explorer \
--output evaluation/truncation/strategy-results.json --check
node scripts/evaluate_context_recovery.mjs --explorer-root /path/to/okf-explorer \
--output evaluation/truncation/recovery-results.json --check
uv run --locked python scripts/build_evidence_route_graph.py --check
uv run --locked python scripts/build_persona_evaluation.py --check
uv run --locked python -m unittest discover -s scripts -p test_evidence_route_graph.py
uv run --locked python -m unittest discover -s scripts -p test_persona_evaluation.py
OKF_EXPLORER_ROOT=/path/to/okf-explorer node --test \
scripts/test_graph_evidence_runtime.mjs scripts/test_graph_evidence_client.mjs \
scripts/test_insufficiency_categories.mjs
node scripts/evaluate_graph_retrieval.mjs --explorer-root /path/to/okf-explorer \
--output evaluation/retrieval-gate/graph-evaluation.json --check
node scripts/replay_frozen_wording.mjs --explorer-root /path/to/okf-explorer
OKF_EXPLORER_ROOT=/path/to/okf-explorer node evaluation/retrieval-gate/withheld/attempt-02/run.mjs
git diff --exit-code -- evaluation/retrieval-gate/withheld/attempt-02/receipt.json \
evaluation/retrieval-gate/withheld/attempt-02/run-summary.json
The strategy matrix covers 17 known development cases. The recovery comparison
covers eight known positive and negative cases; it also checks recovery-plan
inactivity for all 40 original staff questions. That plan check is not a fresh
40-case runtime replay or independent answer-quality evaluation. The additive
current_explorer_validation CI job runs these checks alongside the three
historical prerequisite jobs. The protected aggregate remains validate.
Publication and remaining limits
The capital-source persona observation
adds a dedicated automated check for stage2-006 and
unhinted-accessible-reading. In both installed browsers, keyboard search and
Enter expansion reach exact source record page/84/0035, retain paragraph
84351, expose the official PDF's page 35 locator and show the independent,
unreviewed and extraction labels in text. Its
capital inspector
uses the retained benefit-framed research query. Human assistive-technology
acceptance remains not run. The
persona mutation controls
also exercise altered PDF, literal, source-context and provenance bindings for
the integrity area. These supplemental checks do not silently regrade the
earlier retrieval report's separately pending UI and integrity fields.
The local browser observation binds the generated report and HTML. Installed Chrome 157.0.8103.51 and Edge 157.0.4341.46 passed five-context selection, exact F1120 search, source expansion, keyboard radio navigation, empty search state and 390-pixel layout checks with no console warnings or errors. Retained screenshots show desktop and mobile layouts. These headless observations do not certify comprehensive accessibility, a public deployment or the intended connected client.
The historical reading-help release binds the two September contracts. Its
checker now accepts an explicit --frozen-contracts-commit solely for those
two files, validates their original byte counts and hashes against Git and
continues to validate every other source and receipt binding strictly. Current
contracts have their own mandatory check. This preserves the historical
release manifest when a later candidate changes its contracts; no historical
receipt or source projection is rewritten.
The existing tag demo-2026-09-30-reading-help-v2 already identifies baseline
77ef16047415ee48dcb93b80adf1ba057ea00d9b. It has not been moved or duplicated.
The publication review records the baseline
findings and preserved source boundaries.
The selected service observation retains its exact source and engine bindings. Local 0.8.0 preparation does not itself establish public deployment or host compatibility. Wider unseen subject coverage, benefits-specialist review, current legal applicability and answer accuracy remain separate acceptance tasks. No new source capture or model answer call is required by this workflow.