Context retrieval, allocation and loading experiments

Current qualification allocation result

The joint source/engine comparison records 320 deterministic assemblies and verified replay across two source versions, two engines, 40 tasks and two budgets. DWP 7f9feb9634e3d94004853b838462aca132c505a5 and Explorer c4f2de0a99b7bc2f8b8c8a06a3c715fb56b66d8e retain 177 of 177 candidate occurrences and 433 of 433 activated declared paths at 256 KiB and 512 KiB. All seven household pages remain inspectable for Staff 012 and 013. All contexts remain insufficient; 203 named obligations remain open. These candidates await protected publication and separate public service/browser checks.

The comparison separates source modelling from allocation. With the earlier source at 512 KiB, the new allocator displaces six incidentally selected household pages because those pages were not declared required. Adding explicit qualification paths restores all seven at both budgets. This is a useful measured trade-off, not a guarantee that unmodelled qualifications will survive. The larger-budget permanent-care-home package still exposes a missing optional temporary-residence dependency. Current budget controls also retain a legitimate zero-record metadata refusal for Staff 012 at 64 KiB.

The joint receipt's timings are single local observations. They do not establish a speed gain. The separate controlled loading experiment below is historical and retains its original inputs, counts and interpretation.

Historical ambiguity and loading experiment

The retained candidate Ask OKF assembler shows evidence for each possible meaning of an ambiguous term and loads up to four corpus files at a time. This improves inspection and reduces serial waiting. It does not choose the meaning or establish a correct answer.

What was held fixed

The experiment uses the previously published 40 staff questions and semantic index, before the additional household relationships and statutory bodies. Both engines read the same frozen inputs. Every corpus file must match its recorded length and SHA-256 digest: a digest is a content fingerprint used to detect a changed file.

The retained experiment includes its exact harness, both engine versions, input index, question registry, corpus manifest, full comparison and file digests. It makes no web or model calls. The retained question set is a development set that informed the work, not an independent test of unfamiliar questions.

Retrieval result

Thirty-eight packages are byte-identical. Only staff-008 and staff-009 change: both contain an ambiguous abbreviation. The assembler now retains labelled alternative branches while leaving the ambiguity unresolved. It does not merge the alternatives or activate their evidence requirements as if a meaning had been selected.

The other six missing candidate pages had no directed path from the resolved concepts in this frozen index. Increasing the number of selected records cannot create a missing relationship. Their later source-grounded authoring belongs to the household evidence expansion, which must be evaluated separately.

These are page-retrieval counts, not legal accuracy, claim-level correctness or proof that the evidence answers a question completely.

Loading result and its limits

The harness adds an artificial ten-millisecond delay to each read from cached local bytes. Three repetitions alternate engine order. The table shows the median, or middle, duration for each case from the retained portable run.

The peak number of concurrent reads was four. This isolates scheduling under a controlled delay. It is not a public-network or service benchmark, a promise about Monday's room network, or a measure of AI answer speed. Hardware load, browser behaviour, caches and real network latency change elapsed time.

The semantic index may be up to 8 MiB; ordinary corpus manifests, search files and source-record files remain capped at 4 MiB each. Each assembly retains its 16 MiB transfer and 32 MiB decompressed limits, with reservations made before parallel reads. Output remains bounded at 524,288 bytes. These are per-operation data limits, not a guarantee about total process memory. The remote service's existing 8 MiB cache stays bounded; a large base index can cause eviction and refetching.

Reproduce without replacing the observation

From this repository, use the Node version supported by the Explorer consumer:

node validation/context-performance/2026-09-21/verify.mjs
OKF_CONTEXT_COMPARISON_OUTPUT=/tmp/okf-context-comparison-new.json node --experimental-strip-types validation/context-performance/2026-09-21/compare.mjs

Choose an output filename that does not exist. The harness refuses to overwrite a previous observation. OKF_DWP_CHECKOUT optionally points to another checkout containing the same frozen corpus shards; their digests still have to match.

An earlier harness warning is preserved separately. Its exact pre-fix runner was not retained, so that attempt is not presented as independently replayable. The final run has its own executed harness digest and unchanged earlier observations.