From 47f1ca73e7b72e7a13adc0a9022408c4124b48aa Mon Sep 17 00:00:00 2001
From: profit <profit@lakehouse>
Date: Wed, 22 Apr 2026 22:24:25 -0500
Subject: [PATCH] =?UTF-8?q?auditor:=20Level=201=20correction=20=E2=80=94?=
 =?UTF-8?q?=20keep=20think=3Dtrue,=20only=20temp=3D0=20is=20needed?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

The previous Level 1 commit set think=false which broke the cloud
inference check on real PR audits. gpt-oss:120b is a reasoning model;
at think=false on large prompts (40KB diff + 14 claims) it returned
empty content — verified by inspecting verdict 8-8e4ebbe4b38a which
showed "cloud returned unparseable output — skipped" with 13421
tokens used and head:<empty>.

Small-prompt tests passed because the model could respond without
needing to think. Real audits with the full diff + claims context
require the reasoning channel to produce any output at all.

The determinism we need comes from temp=0 (greedy sampling). The
reasoning trace at think=true varies in prose but greedy sampling
converges to the same FINAL classification from identical starting
state, so signatures remain stable.

max_tokens restored to 3000 for the think trace + response.
---
 auditor/checks/inference.ts | 25 ++++++++++++++-----------
 1 file changed, 14 insertions(+), 11 deletions(-)

diff --git a/auditor/checks/inference.ts b/auditor/checks/inference.ts
index 5cddbc7..c6c3dbf 100644
--- a/auditor/checks/inference.ts
+++ b/auditor/checks/inference.ts
@@ -112,19 +112,22 @@ export async function runInferenceCheck(claims: Claim[], diff: string): Promise<
           { role: "system", content: systemMsg },
           { role: "user", content: userMsg },
         ],
-        // Deterministic classification mode — temp=0 is greedy-sample,
-        // so identical input → identical output on the same model
-        // version. think=false disables the reasoning trace that was
-        // letting variable prose leak into the classification output
-        // and inflate the audit_lessons signature set (observed as
-        // sig_count creep across the 9-run empirical test).
+        // Deterministic classification — temp=0 is greedy-sample, so
+        // identical input yields identical output on the same model
+        // version. This kills the signature creep we observed in the
+        // 9-run empirical test (sig_count 16→27 from cloud phrasing
+        // variance at temp=0.2).
         //
-        // max_tokens tightened to 1500 — the structured JSON response
-        // fits comfortably in 1500 tokens for typical PRs (~7 claims);
-        // the old 3000 just gave the model room to wander.
-        max_tokens: 1500,
+        // IMPORTANT: keep think=true. gpt-oss:120b is a reasoning
+        // model; setting think=false caused it to return empty content
+        // on large prompts (observed during Level 1 validation: 13421
+        // tokens used, empty content returned). The reasoning trace is
+        // variable prose, but at temp=0 the FINAL classification is
+        // still deterministic because greedy sampling converges to
+        // the same conclusion from the same starting state.
+        max_tokens: 3000,
         temperature: 0,
-        think: false,
+        think: true,
       }),
       signal: AbortSignal.timeout(CALL_TIMEOUT_MS),
     });