Skip to content

Latest commit

 

History

History
79 lines (65 loc) · 3.79 KB

File metadata and controls

79 lines (65 loc) · 3.79 KB

ARB orientation-capsule trajectory decision

Date: 2026-07-27

Decision: no_measured_win. Keep the capsule benchmark-only and do not change production retrieval, ranking, tool descriptions, or defaults.

Scope and provenance

The final experiment used the two already prepared local ARB repositories at their pinned revisions. It scheduled one prewalk and one prewalk_capsule run per task with gpt-5.6-terra at high reasoning for the frontier and gpt-5.6-luna for the executor. Both arms used the same clean LeanToken runtime, 4,000-token per-call retrieval bound, 30-call frontier bound, and 20-call executor bound. No larger ARB corpus was downloaded.

The redacted machine report binds raw report BLAKE3 0b0fcd47eaff4fe7095afe06e8a694d4b4b06ab8cddfdbfac384ed9c7a7c9c44, classifier source 68106f7df83e011217d8661f79ddbd41a468a9111e40f3e6bc5a4783b6659673, classifier binary ade1d435775438cb8731c831a342ce04c3c9eae3bfe0fc96e801fffbe5de88cb, and ten available trajectory artifacts. The raw report and private trajectories remain local because they contain absolute paths and repository content.

Two earlier attempts are retained locally and were not reused. V1 completed four invalid cells before an external LeanToken SKILL.md read was found to be misclassified as native task-repository retrieval. V2 was stopped after its first invalid cell exposed two more bounded preflight shapes: exact external skill discovery and a compound Git status/diff check. The final adapter allowlists only exact bounded forms, charges every call, and has negative tests for chained task-source reads and alternate shell connectors.

Result

Both Clap patches passed the frozen offline validator:

Clap metric Prewalk Capsule Candidate delta
Official success yes yes no regression
Retrieval calls 12 9 -3
Retrieval source tokens 8,525 4,327 -4,198 (-49.2%)
Complete capsule prompt tokens 0 103 +103
Net retrieval tokens after prompt 4,095 saved
Dead-end source tokens 3,869 0 -3,869
Reread tokens 4,633 351 -4,282
First owner evidence sequence 5 3 two calls earlier
Provider input tokens 1,037,118 814,503 -222,615 (-21.5%)
Duration 253,163 ms 230,468 ms -22,695 ms (-9.0%)

The Click pair did not produce comparable completed trajectories. The capsule frontier created a hash-verified handoff, but its executor exceeded the frozen 20-call limit; the trace and trajectory were therefore absent. The baseline created a complete handoff and followed the labeled owner, but its executor used forbidden native repository retrieval at merged tool call 27. Neither run reached authoritative validation.

The classifier fails closed on that missing candidate trajectory. It reports one unclassified run, one unverified candidate owner route, and null token, dead-end, and reread deltas for the paired decision instead of treating missing data as zero. There were zero observed success regressions only because both Click arms failed; that is not positive evidence.

Interpretation

The Clap pair is strong mechanism evidence that a 46-token owner capsule can focus a successful prewalk and save substantially more downstream retrieval than its 103-token complete injected prompt. It is still one task and is counterbalanced only by a Click pair that failed in different ways. The pre-registered all-pairs gate therefore does not pass.

A future experiment may reuse the bounded contract and fail-closed classifier, but it needs a newly frozen task set on which both control and candidate can complete under the same executor policy. Limits or prompts must be committed before those runs; this result must not be repaired post hoc by raising only the candidate budget.