Reading More, Finding Less: A Pre-Registered Anatomy of Progressive Disclosure for AI Agents

This paper measures how well a one-line-digest index routes an AI agent to the right information on a prose document corpus, under the access pattern known as progressive disclosure. Gold-reach instruments exist across the neighboring literatures: retrieval evaluation scores a retriever's returned set (recall@k), the agentic-retrieval literature scores trajectory-level gold coverage of an agent's own tool-call reads on raw corpora, context-retrieval benchmarks score explored context against annotated gold, and the skill-routing literature points routing telemetry at digest layers at scale. What this study contributes on that occupied map is the document-corpus instantiation with an attribution its neighbors do not carry: a one-line-digest index (model-written under a fixed operator prompt, as the paper states plainly) as the curated policy's only way to locate documents, per-question read-level telemetry on that arm, a decomposition of failure, and a difficulty-matched oracle control.

The measured result is that decomposition. The index routed the agent to the correct document on 52.0% of 102 questions; associative search (grep) over the same corpus located it on 80.4%. When the curated agent answered incorrectly without the correct document in context, it had read a wrong file in 43 of 47 cases and read nothing in 4. When the index did route correctly, accuracy was 86.8% (76.0% in the replication), and a difficulty-matched control detects no answerability difference in the questions the index failed to route (Fisher p = 0.80; the paper states the interval and what it can and cannot carry). The deficit of the curated policy is reaching the right document, not comprehending it, and the mechanism is measured: at 140 characters per file, 44% of the replication's questions share no content word with their own gold digest, and questions whose digest carried at least one query term were located far more often (71.6% versus 50.9%). The routing signal did not survive selection into the digest. The direction is convergent with the skill-routing literature, where hiding a digest's body costs 37-44 points of routing accuracy over ~80K skills; the paper's related-work section maps that neighborhood, the direct-corpus-interaction literature (with its opposite attribution explained by design: reach binds when a mis-route is terminal, post-reach resolution binds when the agent can iterate), and the opposite-signed results for derived structural indexes, which bound the staked prediction to its substrate.

A pre-registered public replication over 141 releasable documents and 120 frozen questions reproduced the direction under stricter criteria: search led by 12.5 percentage points on accuracy (63.3% versus 50.8%), the curated arm wrong-stopped on 34.2% of questions against search's 18.3% under a symmetric rule, and localization was 75.0% versus 62.5%. The anatomy replicates in public: of the curated arm's 45 non-located questions, 42 had read a wrong file, and 91.1% of the curated arm's non-hydrated answers were wrong (84.0% pooled across both arms under the frozen rule). The replication's formal verdict is "revised" because one hardened prediction about token headroom failed; the paper reports that failure at the same prominence as the passing predictions. In the same measurement program, files promoted into a durable memory directory were later read in 2 of 157 eligible cases, and the corresponding conjecture is refuted.

All predictions were frozen before data collection, analyzers were committed before their first output, and adjudication is mechanical against the frozen criteria. The deposit contains the paper (PDF, HTML, and canonical markdown), both figures, the reference list, and the complete replication study: the corpus snapshot, the frozen questions with gold answers, the index, the two frozen prompt strings that produced questions and digests, the registration chain with an exported commit-order proof, the harness, the adjudicator, all run records, and the C29 public demonstrator kit. The committed effect table re-derives byte-for-byte through the shipped adjudicator, and an externally supplied reproduction recomputed all 18 headline replication numbers from the raw run records before deposit (the paper's method section carries the standing caveat on reviewer independence). The corpus snapshot ships every one of its 141 measured documents; files such as CLAUDE.md and AGENTS.md inside it are measured corpus content, not configuration.