Reading More, Finding Less: release bundle

Reading More, Finding Less: A Pre-Registered Anatomy of Progressive Disclosure for AI Agents. Rashid Azarang, Independent Researcher. Preprint, 2026-08-16. DOI 10.5281/zenodo.21960138; the concept DOI 10.5281/zenodo.21938412 resolves to the current version.

Companion study (2026-08-16)

Finding More, Fusing Less: A Pre-Registered Locator Bake-off for AI Agents. DOI 10.5281/zenodo.21969901, mirrored here under c36-locator-bakeoff/. It tests the constructive converse of this paper: whether the search arm's own miss ceiling is reducible by ranked retrieval. Verdict: revised. BM25 beat exact search by +21.7 pp on localization, but the pre-registered fusion prediction failed (-7.8 pp vs BM25-alone) and the shipped RRF default was retired by rule. Full harness, gated question set, run records, mechanical adjudicator, and the independent validation sessions ship in that directory.

What is here

File	What it is
agent-memory-allocation.pdf	the preprint, typeset (30 pp)
paper.html	the same text as a single self-contained page
paper.md	the canonical markdown the renders are built from
fig1_misrouting.png, fig2_c29_policies.png	the two figures
references.bib	full BibTeX for the reference list
c34-study/	the public replication in full: corpus snapshot, frozen questions, index, harness, adjudicator, run records, results, tests, registration chain, and ORDER-PROOF.md (the commit chain exported for bundle-only readers)
repro-kit/	the C29 public demonstrator: harness, 17 operator-owned English methodology documents, frozen public question set, reference result (see the paper's §10 and §8 Limits (b′) for what it does and does not establish)

One note on the corpus snapshot

c34-study/corpus-snapshot/ contains the 141 documents the replication measured, snapshotted at the byte and pinned in corpus-manifest.json. Files inside it with configuration-sounding names (CLAUDE.md, AGENTS.md, and similar) are measured corpus documents, not configuration of this bundle: the frozen questions are about them, the gold answers are substrings of them, and the adjudicator recomputes generation-input hashes from them. Removing any of them breaks the manifest, the questions, and the replay.

Where to start

c34-study/ is a complete empirical study you can re-run: 141 releasable documents, 120 frozen confirmatory questions with gold answers, the one-line-digest index (model-written under the committed prompt, as the paper's §3 states), all 390 run records with per-attempt logs, the full registration chain, and the harness, adjudicator, and test suite. The paper's §4 anatomy derives from a withheld production corpus and re-derives descriptively from this bundle's records (§5); §5 re-derives from c34-study/ entirely.

Re-running the adjudicator from a clone

The adjudicator is deterministic over the committed run records and reproduces effect-table-2026-08-14.json byte for byte. From a clone of this repository it needs one extra step, for a reason worth stating plainly.

cd c34-study
mkdir -p /tmp/c34 && cp questions-2026-08-13.json /tmp/c34/
python3 adjudicate.py --questions /tmp/c34/questions-2026-08-13.json \
                      --out /tmp/c34/replay.json
diff effect-table-2026-08-14.json /tmp/c34/replay.json   # empty

The adjudicator carries a contamination check that asks whether any run started before the question set was frozen. It reads the freeze time from the git commit that first added questions-2026-08-13.json. In the private research repository that commit is the real freeze, on 2026-08-13, and every run postdates it. This public repository is an export: its first commit added the whole bundle at once, on 2026-08-14, after the runs had already happened. Run against the export's own git metadata, the check therefore reports all 360 records as predating the freeze and returns verdict: void, question_set_contamination.

That is the check misreading an export date as a freeze date. It is not a finding about the study. Reading the question set from outside the checkout, as above, leaves the freeze time undetermined, the check stands down, and the adjudication runs on the committed records as it did originally: verdict revised, sha256 e73c072fcf1c29d699d998dc5d87f726a76dc4583bd6d6e9d05c5af7ce46c83b.

The real ordering evidence for bundle-only readers is c34-study/ORDER-PROOF.md, which exports the commit chain with author timestamps. The question set froze at 2026-08-13T18:03:32-06:00 and the first policy-run provider call is 2026-08-14T01:07:02Z. The call ledger's hash chain in call-ledger.jsonl cross-checks it.

Running the tests

cd c34-study && python3 run_tests.py

193 tests. From a clone, expect one failure and two errors, all three in test_corpus_rule.py and test_leak_gate.py, and all three from the same cause: corpus_rule.py enumerates the corpus by reading a pinned git tree (cb73654) from the private research repository. That tree is not in this export, so those tests cannot run here. The corpus they would rebuild ships directly as c34-study/corpus-snapshot/, pinned by sha256 in corpus-manifest.json, which is checkable without git.