Saving and searching 256 documents works. Savings aren't measured yet.

Intake: 256 documents, with cached replay

A selected collection contained 256 documents. Initial intake, including worker startup, took 101.8 seconds. An unchanged replay took 6.1 seconds with 256 cache hits and zero reparses. Fourteen previously excluded large documents were handled through an explicit bounded profile; the existing defaults stayed unchanged.

Correction (October 7, 2026): an earlier public line listed a 240 PDF / 14 Markdown / 2 HTML mix, plus 899 MB and 7,555 PDF pages. Those figures are not backed by a public artifact in this site repository, so the page now reports only 256 documents. The recorded timings have not changed.

This is accepted mechanical intake and source-preservation evidence, not a full-Library or semantic-fidelity benchmark. Preparation, engineering, review and energy are excluded from those timings. Indexing pages does not establish faithful prose, equations or tables, or improved agent completion.

PDF pages: improved prose order, remaining layout limits

Six fixed pages from two papers, ModernBERT and Latent Diffusion Models, exposed mixing between columns under coordinate-sorted extraction. On 17 valid prose checks, coordinate sorting passed one; creator order passed 16. An eighteenth frozen criterion had a transcription error: the original rubric, outputs and 18-check scores remain preserved, with the corrected denominator disclosed.

An optional, read-only creator-order page reader was separately accepted. It extracts from verified original bytes, keeps source identity and warnings, and checks the complete response against its token budget. One real page response took 0.315 seconds and 1,873 tokens including qualifications and warnings. This is one measured page, not a latency guarantee. The default parser and index were not replaced.

Footnotes and captions can still interrupt prose. Tables, equations and diagrams remain unqualified; inspect rendered originals for such claims. The page-reader revision recorded 248 passing tests plus recovery. A later required-range revision recorded 258 passing tests, zero skips and recovery. These are retained revision-specific checks, not tests rerun for this website update or evidence of agent efficiency.

Grounded answers: keep the lexical baseline

A completed comparison retained 36 responses across 18 questions per method: 12 source-answerable questions and six unsupported questions. Lexical retrieval produced seven of 12 grounded complete answers; reciprocal-rank fusion (RRF) produced six of 12. Lexical justified six of six unsupported refusals, RRF three of six. Correctness and actual cited support were graded separately.

Lexical remains the default. An earlier retrieval-support score favoring RRF did not translate into better answers. One exposed question cohort, model and seed cannot establish a universal winner, and engineering, retrieval preparation and review costs remain separate.

The isolated MiniCheck audit reproduced three false acceptances among 72 negative cases under submitted labels. Shared source summaries and development/held-out overlap limit generalization. Simulated reviewer savings are not operational evidence; this candidate remains advisory rather than an admission or truth gate.

Source-bound delivery and useful assisted maintenance

A roadmap's active section was reduced from 4,644 to 796 text tokens and delivered under a 2,000-token cap. A separate consumer used prepared evidence to repair stale route paragraphs; the result was checked in a separate AI review within the research effort. That consumer retained prior context, so this was useful assisted maintenance, not a fresh-agent efficiency comparison.

Explicit required text ranges now select complete passages directly from a hash-bound original UTF-8 source, avoiding a copied section brief. Callers still choose the range. Delivery preserves source roles, labels, warnings and locators, and refuses missing coverage, stale sources or insufficient budgets. PDF and extracted HTML ranges are unsupported; their page and passage tools remain separate.

An inspected source change invalidated the old packet. Explicit re-ingestion and reselection produced a checked successor. This is a recorded refresh workflow, not automatic updating, an atomic multi-file snapshot or future freshness assurance.

Prepared proposals: a negative comparison preserved

A later frozen route-repair comparison accepted zero of three ordinary proposals and zero of three prepared proposals. All six submitted, but stale claims, copied relative links and missing usable routes failed the criteria. Prepared context used 23,638 reported model tokens versus ordinary 13,012, 81.7% more. Preparation and engineering/review add further cost.

Three subsequent assisted routing corrections were separately accepted. They do not repair the six failed scores. The current goal remains useful completion with lower total effort, including preparation, retrieval, execution, correction and review. No new retrieval method or larger corpus is justified by this negative result alone.

Retained local records and public reproducibility

This update was checked against the live research plan, retained outputs, nonauthor review records and process validations on September 30, 2026. Nonauthor review within the research effort is not external reproduction or authenticated reviewer identity. Private papers, raw traces, runtime code and receipts are not published here. Readers cannot reproduce these private results from the public ledger or synthetic fixture.

The separate released synthetic fixture remains runnable. The pinned decision ledger verifies its public process records and integrity. Neither reproduces private study outcomes. Source identity checks do not establish semantic truth, general autonomy, complete Library integration or overall economics.

Research record, product and availability, evidence index.