Broad-topic discovery experiment v1

Question and controls

Can a fresh agent find and apply relevant observations with only a problem statement, without a campaign identifier, publisher title or saved reference?

Four isolated OpenCode 1.18.27 readers use ~z-ai/glm-flash-latest, with 240 seconds, 40 steps, an 8,192-token output allowance and at most six search queries each. The problem asks for the lowest measured energy acoustic-modem profile for stationary freshwater nodes 200 m apart sending 64-byte packets, subject to 97/100 deliveries and p95 latency at most 180 ms. Readers receive no observations.

The fixed corpus contains one applicable synthetic table and eight distractors: incompatible medium, distance, packet size, motion, theoretical advice, batteries, scheduling simulations and optical transport. An unpredictable observation ID and random energy values distinguish retrieval from invented or memorized answers. The operator seeds publications to isolate reader behavior from publisher variation.

The four cases cross direct versus paraphrased Beacon wording with first versus last position of the target among returned matches. Each uses a separate native Rust HTTP experiment node; a finite test proxy only rearranges returned matches. It never inserts a target that failed the native whole-word query. This is an experimental ordering intervention, not a change to the runtime's ID ordering. A query that matches only the target cannot exercise the ordering contrast.

Evaluation

A successful reader retrieves the actual Thread, checks applicable conditions, recomputes the published digest from exact message bytes, and selects the oracle's profile and energy. Record search requests, returned positions, Thread reads, verification and final output. Correct final fields alone do not establish the retrieval path. Unavailable evidence should produce an explicit refusal to guess.

Digest agreement checks byte integrity; this hosted-publication experiment does not authenticate a publisher or establish scientific truth. Signed saved-reference verification is covered separately by RESEARCH_DISHONEST_MIRROR_V1.md.

Reproduction and limits

experiments/sprite-cohort/research_broad.py creates one finite hub and four fresh reader Sprites, builds the current native runtime and deletes resources after collection. It requires the existing Sprite authentication and dedicated model-key file. Output is single-use at outputs/research-broad-v1/; raw transcripts and private routes stay there. No credential belongs in public result files.

This is one same-model observation per cell, on a single known node with nine records and no truncation pressure. It does not estimate a failure rate or test global routing, large-corpus semantic search, spontaneous protocol adoption, independent publication or adversarial prompt injection. The task explicitly asks readers to investigate the supplied node.

Results — 7 September 2026

All four traces contain the actual observation response and a computed matching SHA-256 digest. Both paraphrased readers initially missed the target with acoustic modem queries, then found it by changing terms to freshwater or stationary freshwater observations. All answered within six queries and the deadline.

Result-order robustness remains untested. Both last-position readers narrowed the successful query to a single match. The first-position readers encountered multiple matches, but neither last-position reader actually selected a useful result behind distractors. Preserve this as an unexercised experimental condition, not a successful order-bias test. A separate forced multi-match selection probe is needed; it should be labeled separately from unguided query formulation.

Reader 1 unnecessarily explored several serialization variants before confirming the exact raw event text matched. It also attempted to change to /tmp/opencode, outside the requested task directory; that tool attempt returned no output before it retried successfully. Reader 2's prose said “2” queries in one sentence although its trace and answer list contain one search. Counts above use actual search calls.

Model usage was $0.041735350. All five temporary Sprites were deleted. Public fixtures, oracle, sanitized search/verification commands and evidence hashes are in results/research-broad-v1/. Raw transcripts remain private. No native protocol code or frozen control changed for this trial.