Bound VALIDATES edge precision before reward use

Context

PR #12638 delivered #9906 by adding durable TEST FILE -> VALIDATES -> CLASS edges in the Native Edge Graph. The cross-family review approved the PR and treated the precision concern as non-blocking because #9906 needed the relation contract first, but the concern becomes load-bearing before any reward / gap-downgrade consumer treats VALIDATES as stronger than token-heuristic evidence.

Live latest-open sweep: checked the latest 20 open issues on 2026-06-06 before filing; no equivalent found. Exact live duplicate search for VALIDATES edge token match precision returned no issues. Local issue/discussion sweep found related historical anchors (#9906, #9889, #10035, #9904) but no ticket for this post-#12638 precision bound. Knowledge Base semantic ticket sweep was attempted and unavailable (collection not found), so the filing relies on live GitHub + local synced substrate evidence.

The Problem

Current GapInferenceEngine.inferTestGapsFromSession() uses the same broad node-token regex predicate for two different claims:

• no token match -> emit [TEST_GAP]
• any token match -> suppress [TEST_GAP] and create a VALIDATES edge

That predicate is false-negative-tolerant for gap detection, but the inverse is false-positive-sensitive for graph evidence. Example risk from the review: a structural node named ButtonFeature tokenizes to button + feature; a path such as test/.../ButtonStore.spec.mjs can match button and receive a VALIDATES edge even though it does not validate ButtonFeature.

This matters because VALIDATES is now a structural relation. Once downstream RLAIF / reward / gap-downgrade consumers read it, a broad token match can become overconfident evidence.

The Architectural Reality

Current origin/dev after #12638:

• ai/services/graph/GapInferenceEngine.mjs tokenizes structural node names and matches any token against test/ file paths before calling linkTestEvidenceToStructuralNode().
• linkTestEvidenceToStructuralNode() writes VALIDATES edges with metadata: evidenceKind, evidencePath, inferredBy, validatedNodeName, and validatedNodeType.
• test/playwright/unit/ai/daemons/orchestrator/services/DreamService.spec.mjs covers the positive exact-match case (ButtonFeature -> ButtonFeature.spec.mjs) but not the sibling-token false-positive class.
• learn/agentos/DreamPipeline.md and learn/benefits/ArchitectureOverview.md document the current behavior as token match -> VALIDATES edge.

The Fix

Tighten the VALIDATES creation predicate or explicitly classify its confidence so downstream consumers cannot mistake broad token evidence for direct validation.

Preferred implementation shape:

1. Extract the test-file matching predicate into a small helper on GapInferenceEngine so the same logic is named, documented, and unit-tested.
2. Require stronger evidence for VALIDATES than a single shared token when the structural node has multiple significant tokens. For example: exact normalized basename match, all significant tokens present in the basename/path segment, or an equivalent source-grounded predicate.
3. Preserve [TEST_GAP] behavior for genuinely uncovered nodes.
4. Add a negative fixture where two structural nodes share one token and only one test file exists; verify only the actually-covered node receives VALIDATES.
5. Update Dream Pipeline / Architecture Overview wording so consumers understand the evidence contract.

Contract Ledger Matrix

Decision Record impact

Aligned with the Native Edge Graph / Dream Pipeline architecture. No ADR amendment required.

Acceptance Criteria

• GapInferenceEngine has a named, documented predicate for test-file evidence matching.
• A sibling-token false-positive fixture proves ButtonFeature does not receive VALIDATES from a ButtonStore-style test path.
• The positive exact-match VALIDATES fixture from #12638 still passes.
• [TEST_GAP] still emits for uncovered structural nodes.
• learn/agentos/DreamPipeline.md and learn/benefits/ArchitectureOverview.md describe the tightened evidence contract or confidence semantics.
• Focused unit verification runs against the existing DreamService / GapInferenceEngine test surface.

Out of Scope

• Replacing all test-discovery heuristics with a full parser or coverage engine.
• Changing GraphService.linkNodes semantics.
• Adding reward / gap-downgrade consumers for VALIDATES.
• Reopening #9906; #9906 delivered the first relation contract and remains complete.

Avoided Traps

• Treating all token matches as equal evidence. Good enough for a gap heuristic is not automatically good enough for a structural validation edge.
• Deleting the heuristic entirely. The Dream Pipeline still needs lightweight test discovery; the goal is a precision bound, not a full coverage rewrite.
• A tiny docs-only clarification. The graph predicate and unit fixture must move together; docs alone would leave the overconfident edge path intact.

Related

• #12638 — delivered #9906 and surfaced this follow-up in review.
• #9906 — original TEST -> VALIDATES -> CLASS relation contract.
• #9904 — parent RLAIF / Dream topology context.
• #10035 — prior GapInferenceEngine move away from fragile regex concept matching.

Origin Session ID

dbb1a88c-987f-4519-9645-8f13e9d71000

Handoff Retrieval Hints

• query_raw_memories(query="VALIDATES edge token match precision GapInferenceEngine false positive")
• query_summaries(query="PR 12638 VALIDATES precision follow-up")
• Source anchors: GapInferenceEngine.inferTestGapsFromSession, linkTestEvidenceToStructuralNode, DreamService.spec.mjs inferTestGapsFromSession links matching test files via VALIDATES edges (#9906).