Skip to content

Commit 6d71fd8

Browse files
committed
fix(agent-workspace): disambiguate stale repeated evidence spans
1 parent d964476 commit 6d71fd8

5 files changed

Lines changed: 89 additions & 10 deletions

File tree

docs/diataxis/en/explanation/development-progress-dashboard.md

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -61,6 +61,7 @@ Fresh implementation evidence in this slice:
6161
- The runtime probe `conflicting_release_date_evidence_probe_en`, backed by `Knowledge_Base/ragdateconflict/release date conflict probe.md`, now verifies that same-section `2026-07-01` and `2026-08-15` release-date statements produce the same `conflict` RAG role and `borderline/conflict` degradation instead of publishing one stable schedule.
6262
- The runtime probe `conflicting_multi_document_evidence_probe_en`, backed by `Knowledge_Base/ragmulticonflict`, now verifies that contradictory `+/-0.10 mm` and `+/-0.50 mm` calibration-tolerance facts across two scoped documents produce a cross-document `conflict` fragment, stay visible in the one public answer, and degrade under `context_assembly`.
6363
- The runtime probe `full_document_scan_remote_conflict_probe_en`, backed by `Knowledge_Base/ragfullscan`, now verifies the corrected source-boundary rule: the local paragraph window is not the maximum range. Two selected documents are read in full, remote appendix facts outside the matched opening spans are scanned for structured conflict evidence, and only the budgeted `RagContextPack` is capped before model-visible answer composition.
64+
- Focused provenance tests now verify that repeated snippet text does not force the first textual occurrence when offsets are stale. If an offset candidate does not contain the evidence snippet but the line-range candidate does, `evidenceContextAssembler.ts` uses the line provenance to select the intended source block and parent section.
6465
- Focused hard-negative tests now verify that an unavailable graph-neighbor source window is not treated as complete graph evidence merely because a graph-neighbor direct span or title exists. `evidenceContextAssembler.ts` records `source_resolver_returned_no_content:graph_neighbor_support`, and `ragSufficiencyJudge.ts` keeps the answer `borderline/partial_coverage` with `graph_neighbor_evidence_missing` until grounded neighbor source context is available.
6566
- Focused graph-window ambiguity tests now verify that compare intent filters high-confidence `sequence` / `prerequisite` successors when `contrast` / `analogy` neighbors exist, while still failing open to structural successors when the graph has no compare-aligned neighbor. The runtime verifier now also asserts the graph diagnostics for the `graphintent` compare probes: at least two intent-aligned successor candidates, at least one intent-misaligned successor candidate, and no misaligned-successor fallback.
6667

docs/diataxis/zh/explanation/development-progress-dashboard.md

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -126,6 +126,7 @@
126126
- `conflicting_release_date_evidence_probe_en` 现在由 `Knowledge_Base/ragdateconflict/release date conflict probe.md` 支撑,验证同一 scoped section 内的 `2026-07-01``2026-08-15` release-date 证据也会进入 `conflict` RAG role,并保持 `borderline/conflict` 降级,而不是发布单一稳定 schedule。
127127
- `conflicting_multi_document_evidence_probe_en` 现在由 `Knowledge_Base/ragmulticonflict` 支撑,验证两个 scoped 文档中的 `+/-0.10 mm``+/-0.50 mm` calibration-tolerance 矛盾会进入跨文档 `conflict` RAG role,保留在单条公开回答中,并保持 `borderline/conflict` 降级。
128128
- `full_document_scan_remote_conflict_probe_en` 现在由 `Knowledge_Base/ragfullscan` 支撑,验证修正后的 source-boundary 规则:局部段落窗口不是最大范围。两个被选文档会先完整读入,命中开头段落之外的远端 appendix 事实也会参与结构化冲突扫描;真正被硬限制的是进入公开回答编排前的有界 `RagContextPack`
129+
- focused provenance 测试现在验证 repeated snippet 不会在 offset 陈旧时强制落到第一处文本命中。如果 offset 候选不包含 evidence snippet,而 line-range 候选包含它,`evidenceContextAssembler.ts` 会使用 line provenance 选择正确 source block 与 parent section。
129130

130131
后续推进方向:
131132

0 commit comments

Comments
 (0)