role-os 2.5.0 → 2.6.0

package/CHANGELOG.md

@@ -1,5 +1,17 @@
 # Changelog
 
+## 2.6.0
+
+### Changed
+
+#### `verify-citations --local-panel` now judges against prism's FULL abstract (not just the span)
+
+- The local panel previously re-checked each `supported` citation against only prism's `source_title` + the single `supporting_span` the groundedness lens surfaced. A faithful claim that the *full* abstract entails — but no single span does — was escalated as a panel disagreement (the wave-6 end-to-end Kambhampati false-escalation). `buildEvidence` now prefers prism's full `source_abstract` (surfaced by prism **v1.0+**), falling back to the span on older prism builds — so faithful claims land cleanly while genuine false-confirms are still caught.
+- `gateCitations` threads `source_abstract` through from prism's `citation_results`. Requires prism ≥ 1.0 to take effect; older prism builds omit the field and the span fallback preserves prior behavior. Pairs with `tensor-engine-knowledge` wave-9 (the 3rd verifier family + the full-abstract e2e).
+
+### Tests
+- 3 new tests (`buildEvidence` abstract-preference + span fallback; an end-to-end assertion that the panel's evidence carries the full abstract). **1199 total, all green.**
+
 ## 2.5.0
 
 ### Added

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "role-os",
-  "version": "2.5.0",
+  "version": "2.6.0",
   "description": "Role OS — a multi-Claude operating system where 61 specialized roles execute work through contracts, conflict detection, escalation, and structured evidence. 10 team packs, 9 missions including dogfood swarm (multi-pass convergence), deep audit with manifest-scaled dynamic dispatch, and brainstorm with traceable disagreement.",
   "homepage": "https://mcp-tool-shop-org.github.io/role-os/",
   "bugs": {

package/src/citation-panel.mjs

@@ -29,16 +29,18 @@ export const DEFAULT_OFFLOAD_SCRIPT =
 
 /**
  * Build the evidence string the panel judges the claim against: prism's retrieved source title +
- * the single supporting span prism surfaced. Thin by design — if even prism's OWN best span does
- * not entail the claim under a strict panel, that is exactly the false-confirm worth catching.
- * (Surfacing prism's full retrieved abstract would strengthen this — tracked as a prism follow-up.)
+ * its FULL retrieved abstract (prism v1.0+ surfaces `source_abstract`), falling back to the single
+ * supporting span on older prism builds. Judging against the whole abstract — not one sentence —
+ * stops the panel from escalating a faithful claim that the abstract entails but prism's single
+ * span does not (the wave-6 e2e Kambhampati false-escalation). A strict panel that STILL cannot
+ * entail the claim from the full abstract is exactly the false-confirm worth catching.
  * @returns {string} evidence, or "" when prism surfaced nothing to judge against.
  */
-export function buildEvidence({ source_title, span } = {}) {
+export function buildEvidence({ source_title, source_abstract, span } = {}) {
   const title = (source_title || "").trim();
-  const s = (span || "").trim();
-  if (!title && !s) return "";
-  return [title ? `Title: ${title}` : "", s].filter(Boolean).join("\n\n");
+  const body = (source_abstract || "").trim() || (span || "").trim();
+  if (!title && !body) return "";
+  return [title ? `Title: ${title}` : "", body].filter(Boolean).join("\n\n");
 }
 
 /** Default exec — execFileSync, capturing stdout even on a non-zero exit, no shell (args verbatim). */

package/src/verify-citations.mjs

@@ -196,6 +196,7 @@ export function gateCitations(prismResponse) {
     detail: cr.detail,
     span: cr.supporting_span ?? null,
     source_title: cr.source_title ?? null,
+    source_abstract: cr.source_abstract ?? null,
   }));
   // role-os enforces the deterministic floor ITSELF (it does not delegate it to prism's top-level
   // aggregation): any fabricated-existence citation BLOCKS and dominates a top-level "accept", so a