Files
CyberGril/docs
zonggongandmultica-agent 81d50f28af HEL-868: fixed 13-call real A/B retest executed in one bounded round per rework860 _retest_plan under the original '同意测试' authorization after manager HEL-867 review lifted the stop gate - criteria/fixtures/orchestration untouched at e06dcaf (driver only captures per-call raw input/sent-context/answers and writes per-scenario outputs, never overwriting ab_scenarios.json); preflight zero-embed zero-network audit861_reuse868 + audit867_reuse868 both exit 0 with prior evidence and ledgers unchanged; 13/13 calls sent HTTP 200 (ADD1-5 anchored real extraction + QA1-8), real ledger continued on a SQLite-consistent backup of the original 26-row ledger (prefix line digest 03f92557 identical, original read-only preserved) now 39/60, embed ledger 458->498/500 (+40 dense, pre-computed debit guard held), usage 163041 tokens recorded per call; honest per-scenario results: S1/S4/S5/S8/S9 pass, S2/S3 mem0 side and S7 fail fail-closed because base-group batch-anchored infer hits carry the unshared t1 companion row in the group anchor and companion_unshared rejects them at every exit (zero leak markers in all sent contexts - the gate holds; row-level provenance for batch extractions is our adaptation gap, not visible to offline row-anchored fake payloads), S3 retrieval layer strictly better re-confirmed (11:11 target hits vs baseline 0:0) but model-use verdict fails without the gated content, S6 orchestration/status-in-block/unresolved-contrast all pass while both real answers are descriptive '资料里说...' unsupported by the pre-declared bounded grammar and honestly fail closed; no rerun, no criteria relaxation, no debug requests, no merge, no deploy; dual-target requirement checks 4/4 pass with handoff-HEL-868.json
Co-authored-by: multica-agent <github@multica.ai>
2026-10-26 03:49:33 +08:00
..
HEL-868: fixed 13-call real A/B retest executed in one bounded round per rework860 _retest_plan under the original '同意测试' authorization after manager HEL-867 review lifted the stop gate - criteria/fixtures/orchestration untouched at e06dcaf (driver only captures per-call raw input/sent-context/answers and writes per-scenario outputs, never overwriting ab_scenarios.json); preflight zero-embed zero-network audit861_reuse868 + audit867_reuse868 both exit 0 with prior evidence and ledgers unchanged; 13/13 calls sent HTTP 200 (ADD1-5 anchored real extraction + QA1-8), real ledger continued on a SQLite-consistent backup of the original 26-row ledger (prefix line digest 03f92557 identical, original read-only preserved) now 39/60, embed ledger 458->498/500 (+40 dense, pre-computed debit guard held), usage 163041 tokens recorded per call; honest per-scenario results: S1/S4/S5/S8/S9 pass, S2/S3 mem0 side and S7 fail fail-closed because base-group batch-anchored infer hits carry the unshared t1 companion row in the group anchor and companion_unshared rejects them at every exit (zero leak markers in all sent contexts - the gate holds; row-level provenance for batch extractions is our adaptation gap, not visible to offline row-anchored fake payloads), S3 retrieval layer strictly better re-confirmed (11:11 target hits vs baseline 0:0) but model-use verdict fails without the gated content, S6 orchestration/status-in-block/unresolved-contrast all pass while both real answers are descriptive '资料里说...' unsupported by the pre-declared bounded grammar and honestly fail closed; no rerun, no criteria relaxation, no debug requests, no merge, no deploy; dual-target requirement checks 4/4 pass with handoff-HEL-868.json
2026-10-26 03:49:33 +08:00