From 67699e8737101bba7936b0860d8782429dc0344a Mon Sep 17 00:00:00 2001 From: Marius Ciepluch <11855163+norandom@users.noreply.github.com> Date: Wed, 30 Sep 2026 08:35:41 +0200 Subject: [PATCH] bench(plane): increment gate met after the tie-break -- 16/20 agreed at $0.0205/candidate A third verify pass on the 10 disputed candidates, 2-of-3 agreement (Req 6.4), run on ax by extending the same Run: 16/20 declared sites agreed (reference 16), $0.0205 per candidate with recorded triage (reference $0.022). Caveats recorded: a design change on the same set; 17 vs 19 sites flagged in any pass; three non-declared pgadmin candidates promoted. Co-Authored-By: Claude Opus 5.5 Claude-Session: https://claude.ai/code/session_01EmM1j46qgKMh3zqD1QMwEo --- .kiro/specs/ai-service-plane/tasks.md | 2 +- benchmarks/independent/plane-increment-2.json | 65 +++++++++++++++++++ 2 files changed, 66 insertions(+), 1 deletion(-) create mode 100644 benchmarks/independent/plane-increment-2.json diff --git a/.kiro/specs/ai-service-plane/tasks.md b/.kiro/specs/ai-service-plane/tasks.md index 91bfb27..a1b2dd8 100644 --- a/.kiro/specs/ai-service-plane/tasks.md +++ b/.kiro/specs/ai-service-plane/tasks.md @@ -130,7 +130,7 @@ mypy are clean, and the commit is gated on pytest's own exit code. Script-level if not met, revise the design before task 9. - _Requirements: 6.1, 6.2, 6.3_ -- [ ] 8.1 Tie-break pass and re-measurement +- [x] 8.1 Tie-break pass and re-measurement - `verify` takes `OUSAST_INPUT_ONLY` (pass c on the disputed candidates only); `agree` takes `OUSAST_INPUT_PASS_C` and decides 2-of-3 where pass c asked, the a/b rule elsewhere (`metrics.tiebreak`). - `validation-46` gains `-vc` and `-final` per case, appended so the first 60 entries keep their diff --git a/benchmarks/independent/plane-increment-2.json b/benchmarks/independent/plane-increment-2.json new file mode 100644 index 0000000..b45acb7 --- /dev/null +++ b/benchmarks/independent/plane-increment-2.json @@ -0,0 +1,65 @@ +{ + "measurement": "ai-service-plane increment re-measured after the design revisit (Requirement 6.4): passes a and b as recorded in plane-increment-1.json, plus a tie-break pass c on the 10 candidates a and b disputed; agreement 2-of-3. Executed on google/ax by rerunning the same Run (the 60 done tasks were reused, 30 new tasks ran).", + "date": "2026-09-30", + "run": { + "name": "validation-46", + "tasks": 90, + "status": "done", + "runner_image": "localhost:5001/ousast-runner@sha256:fb9a743027518976659fd11505e58cc3fc84fb42e07627196e13b76ac01d0617" + }, + "previous": "benchmarks/independent/plane-increment-1.json", + "gate": { + "declared_sites_agreed": { + "plane": 16, + "reference": 16, + "of": 20, + "required": ">= 16", + "met": true + }, + "cost_per_candidate_with_recorded_triage_usd": { + "plane_over_46": 0.0205, + "plane_over_43_after_triage": 0.02195, + "reference_over_46": 0.022, + "required": "< 0.022", + "met": true, + "note": "the reference divides by the 46 candidates before triage; on that basis the plane is 7% cheaper, on the 43 after triage it is 0.2% under" + }, + "result": "MET" + }, + "detail": { + "tiebreak": { + "asked": 10, + "flipped_to_agreed": 5, + "declared_sites_recovered": [ + "froxlor listing", + "froxlor listingCount" + ], + "declared_site_still_disputed": "froxlor validateIpAddresses (b only)", + "non_declared_flipped_to_agreed": [ + "pgadmin enable_disable", + "pgadmin get_job_args_value", + "pgadmin run_job" + ] + }, + "agreed_candidates": { + "two_pass": 23, + "two_of_three": 28 + }, + "declared_sites_flagged_in_any_pass": { + "plane": 17, + "reference": 19, + "of": 20, + "missed_by_every_plane_pass": [ + "budibase read", + "fuxa getDaqValue" + ] + }, + "usd": { + "plane": 0.8567, + "recorded_triage": 0.0872, + "with_triage": 0.9439, + "reference": 1.0123 + } + }, + "reading": "The plane on ax matches the script pipeline's pre-registered recall (16/20 agreed) at 7% lower cost, so Requirement 6.2 is met after the 6.4 revisit. Two caveats travel with the result: (1) the tie-break is a design change chosen after seeing the first measurement on the same validation set, so this compares plane with script like for like and qualifies nothing on an independent population; (2) the plane's passes flag fewer declared sites than the reference's (17 vs 19 in any pass), and 2-of-3 also promoted three non-declared pgadmin candidates, whose truth is unassessed (an earlier pgadmin spot check found 3 of 5 findings truly privileged). Precision is not part of this gate and remains the open problem." +}