Skip to content

Aider standalone diagnostic canaries

Prepared from closed local evidence on 2026-10-25. These four single-task runs are excluded from the campaign population. They diagnose harness behavior, not model quality or serving speed.

  • summary.json: exact source/adapter/oracle-image identities, all four outcomes, counts, durations, limits and missing evidence.
  • native-records.jsonl: native-shaped sanitized plan, selection, terminal artifact, result, process, retention, grader and request-event records. Every existing request row is retained in order.
  • projection-ledger.json: original private SHA/bytes, declared omissions, source-to-record slices and public derivative hashes. Missing adapter/grader records remain explicit.
  • projection-validation.json: bounded source/record/hash/privacy checks, not a native replay/validation claim.
  • friction-log.md, restoration.json, source-registry.json: causal limits and operation boundary.
  • artifact-manifest.json: all ten publication roles and the closed public inventory.

R1 failed with no retained exact boundary cause. R2 proves output overflow. R3 completed with false/true grader outcomes despite false C++ header lint affecting its conversation. R4 completed with false/true outcomes after the scoped correction and passed independent single-task review. Neither first grader failure has an assigned cause. Missing outcomes never count as passes.

The three corrections preserve typed failures, increase only the finite Aider process-capture cap, and select the correct grammar for admitted C++ solution headers. The latter is a declared local methodology correction, not official leaderboard equivalence. Counts and timings are descriptive observations only; no ratios or speedup/quality comparison are warranted.

Raw prompts, replies, reasoning, chat, command/environment strings, host/network identity and private paths are omitted. Bulky preparation inventories remain private. Schemas and remaining native fields retain their nesting; redaction means native internal hashes still refer to private originals and cannot validate these derivatives. The manifest only binds the public bytes. This bundle was assembled after the canaries, not frozen before requests. Full publication/matrix integration and campaign acceptance remain separate.