Fix the six mis-executions and three unclarified invalid-slot cases that the A3 recall baseline exposed, in the rule layer (no retrieval, no model changes - the baseline showed top-1 at 0.968 with all misses being rule-layer semantics).
Why: semantic negations (不要/别/不用/先不要 + action term) matched as positive intents - "不要重启任务" triggered a restart; a multi-intent tie-break preferred a bare greeting over a slot-bearing action; multi-candidate values, charset violations and symbols-only slot values started without clarification.
Evidence: one implementation unit (negation-aware per-hit check in compiledSpec.evaluate; specificity tie-break in adjudication; slot extraction multi-value/charset invalidity -> clarify) + the nested-pin cascade; the eval corpus mis_exec is now 0 with 37/37 cases passing and no regression in expired_confirmation, generation_consistency or overall top1 (verify-misexec2.json, race); full make test green (verify-misexec-combo2.json). Independent review lands in the task card before merge.
Impact: four component pins advance (native be556f4, intent 7f3ff6a, squad 95b387d, backend bd5c0f3) with matching nested bumps; clarify/fallback rates rise as dangerous inputs now clarify instead of executing. The retrieval/model-tuning decision stays open with a cleaner baseline.
Fix the six mis-executions and three unclarified invalid-slot cases that the A3 recall baseline exposed, in the rule layer (no retrieval, no model changes - the baseline showed top-1 at 0.968 with all misses being rule-layer semantics).
Why: semantic negations (不要/别/不用/先不要 + action term) matched as positive intents - "不要重启任务" triggered a restart; a multi-intent tie-break preferred a bare greeting over a slot-bearing action; multi-candidate values, charset violations and symbols-only slot values started without clarification.
Evidence: one implementation unit (negation-aware per-hit check in compiledSpec.evaluate; specificity tie-break in adjudication; slot extraction multi-value/charset invalidity -> clarify) + the nested-pin cascade; the eval corpus mis_exec is now 0 with 37/37 cases passing and no regression in expired_confirmation, generation_consistency or overall top1 (verify-misexec2.json, race); full `make test` green (verify-misexec-combo2.json). Independent review lands in the task card before merge.
Impact: four component pins advance (native be556f4, intent 7f3ff6a, squad 95b387d, backend bd5c0f3) with matching nested bumps; clarify/fallback rates rise as dangerous inputs now clarify instead of executing. The retrieval/model-tuning decision stays open with a cleaner baseline.
Refs: evaluation/BASELINE.md; feature/catalog-recall-eval-a3
Why: establish the authorized task workspace and its initial baseline.
Evidence: preserved precheck receipt in .task/feature-recall-misexec-fixes.md; R0=bccb88f294038ce52a37fc5592a6daebd10e88b4.
Impact: task metadata only; the feature workspace starts at bccb88f294.
Refs: .task/feature-recall-misexec-fixes.md
Why: fix the design (negation-aware matching in compiledSpec.evaluate,
specificity tie-break in adjudication, slot multi-value/charset
invalidity) with the eval corpus as acceptance.
Evidence: A3 baseline analysis (evaluation/BASELINE.md).
Impact: task metadata only.
Refs: feature/catalog-recall-eval-a3 (baseline)
Why: land the A3-baseline-driven rule-layer fixes (negation-aware
matching, specificity tie-break, fail-closed slots) across the
composition.
Evidence: unit transport (misexec-git-run.json); eval mis_exec=0 with
37/37 pass (verify-misexec2.json); full combination re-run below.
Impact: four component gitlinks.
Refs: feature/recall-misexec-fixes
Why: the backend branch tip carries the corrected nested
intent/native/squad pins for the mis-execution fix composition.
Evidence: integration transport (misexec-git3-run.json); full
combination re-run below.
Impact: one component gitlink.
Refs: feature/recall-misexec-fixes
Why: close the task card inside its own PR so landing it needs no post-merge backfill or separate close PR.
Evidence: .task/feature-recall-misexec-fixes.md -> .task/archive/feature-recall-misexec-fixes.md; status 已完成（合入前闭合）; PR #717https://git-home.u1680832.nyat.app:30319/weichen/Agent-System-Dev/pulls/717.
Impact: task metadata only; the card is archived within PR #717 and lands with its merge commit.
Refs: task.sh close --self
Why: include the current integration baseline before merge.
Evidence: fetched origin/main=3dae91ec.
Impact: integrate upstream changes without rewriting task history.
Refs: task-sync
Why: include the current integration baseline before merge.
Evidence: fetched origin/main=a6485795.
Impact: integrate upstream changes without rewriting task history.
Refs: task-sync
weichen
merged commit 0650e81133 into main2026-10-23 05:39:42 +00:00
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
Fix the six mis-executions and three unclarified invalid-slot cases that the A3 recall baseline exposed, in the rule layer (no retrieval, no model changes - the baseline showed top-1 at 0.968 with all misses being rule-layer semantics).
Why: semantic negations (不要/别/不用/先不要 + action term) matched as positive intents - "不要重启任务" triggered a restart; a multi-intent tie-break preferred a bare greeting over a slot-bearing action; multi-candidate values, charset violations and symbols-only slot values started without clarification.
Evidence: one implementation unit (negation-aware per-hit check in compiledSpec.evaluate; specificity tie-break in adjudication; slot extraction multi-value/charset invalidity -> clarify) + the nested-pin cascade; the eval corpus mis_exec is now 0 with 37/37 cases passing and no regression in expired_confirmation, generation_consistency or overall top1 (verify-misexec2.json, race); full
make testgreen (verify-misexec-combo2.json). Independent review lands in the task card before merge.Impact: four component pins advance (native be556f4, intent 7f3ff6a, squad 95b387d, backend bd5c0f3) with matching nested bumps; clarify/fallback rates rise as dangerous inputs now clarify instead of executing. The retrieval/model-tuning decision stays open with a cleaner baseline.
Refs: evaluation/BASELINE.md; feature/catalog-recall-eval-a3