conformance-judge

PENDING — the agentic core. Per touched node, judge whether a PR's diff honors that node's spec intent, fed the spec + its tree + the node's diff slice + its yatsu scenarios. Structured verdict, not free text. Deterministic spec-touch matrix is its input. No code yet.

The agentic heart of CI-Gate-Spex-forge — and the LLM judge spec-lint explicitly deferred ("whether the code still matches what the spec says is the LLM judge's job, async, not in the commit path").

Determinism routes; the agent judges. Routing is mechanical — the code: graph already maps each changed file to its node. The judge appears only where meaning must be weighed: does this slice of the diff still honor what its node's spec says?

Tier 0.5 — the spec-touch matrix (deterministic input). Per touched node, classify (spec touched?, code touched?):

• code changed / spec untouched — drift risk; the judge decides harmless detail vs. silent divergence (the most dangerous cell, invisible to lint/typecheck).
• spec + code changed — normal; the judge checks the two are self-consistent.
• spec only — intent moved ahead of code; a note, with a follow-up question.
• neither, file ungoverned — a coverage smell, not a judgment.

Tier 1 — one judge per touched node, bounded and structured. Not one agent over the whole PR. Each judge is fed: the node's spec body plus its parent/siblings/children (intent is only legible against the tree), the node's slice of the diff (already grouped by review-proof), and the node's yatsu scenarios (the expected zero-loss behavior). It returns a closed-schema verdict — { verdict: conforms | diverges | spec-missing, severity: block | warn | note, rationale, evidence } — never free prose. Judges run in parallel across nodes and are skipped entirely for a PR that touches no governed node.

Read-mostly. The judge produces a verdict; it never pushes a commit, never auto-merges, and never writes a node's status. Escalating to adversarial multi-vote review is a depth knob, not the default.

Out of scope: publishing the verdict (forge-write-seam), sequencing it with the Tier-0 gates and CI (forge-gate), and the Tier-2 suggestion agent (deferred, default off).