From f29d3a45be6e08421be3cf4dd9de1f42dde27e5c Mon Sep 17 00:00:00 2001
From: wassname <1103714+wassname@users.noreply.github.com>
Date: Fri, 17 Apr 2026 05:57:42 +0800
Subject: [PATCH] fix: FM2 asks for sneaky/perverse failure, not undetectable
 one

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
---
 src/index.ts | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/src/index.ts b/src/index.ts
index 0cfbc30..7a72844 100644
--- a/src/index.ts
+++ b/src/index.ts
@@ -338,7 +338,7 @@ After this, task enters pending sign-off state — only completable via /lgtm <i
 
 - **evidence**: Auditable proof — command output, table, file path, link
 - **failure_mode_1**: Most likely way this could be wrong despite evidence
-- **failure_mode_2**: Most subtle/perverse failure -- silent fail, null hypothesis, or one that your evidence would NOT distinguish from success
+- **failure_mode_2**: Most perverse or sneaky failure -- one that looks like success superficially, corrupts silently, or only breaks under specific conditions (scale, time, edge case). E.g. feature active but wrong mechanism, works in tests but degrades in prod, correct output for wrong reason.
 - **falsification_test**: Concrete check that distinguishes your hypothesis from failure modes. Can be a command, a procedure, a log snippet to look for, or an experiment observation. Include: what to run or observe, what you expect if claim is true, what you expect if a failure mode is real, and why this check can't accidentally pass under the failure. Think especially about: null hypothesis (feature isn't active at all), silent failures (error swallowed, fallback triggered silently), and env mismatch (passes in test, broken in prod).
 - **evidence_files** (optional): File paths human should inspect -- must exist
 - **remaining_uncertainty** (optional): What's NOT tested, known limitations, deferred edge cases`,
@@ -346,7 +346,7 @@ After this, task enters pending sign-off state — only completable via /lgtm <i
       taskId: Type.String({ description: "Task ID to submit for sign-off" }),
       evidence: Type.String({ description: "Auditable proof with full reproducibility: exact command run and its output, commit hash, config/seeds used, output file paths. Must be re-runnable by the human. 'I wrote X' is not evidence -- 'I ran X and got Y' is. Include counts, snippets, test output." }),
       failure_mode_1: Type.String({ description: "Most likely way this could be wrong despite evidence" }),
-      failure_mode_2: Type.String({ description: "Most subtle or hard-to-distinguish failure: silent fail, null hypothesis (feature not active at all), or a bug your evidence would NOT distinguish from correct behavior" }),
+      failure_mode_2: Type.String({ description: "Most perverse or sneaky failure: looks like success superficially, corrupts silently, or only breaks at scale/time/edge case. E.g. correct output for wrong reason, feature active but wrong mechanism, passes tests but degrades in prod." }),
       falsification_test: Type.String({ description: "Concrete check (command, procedure, log inspection, or experiment observation) that distinguishes your hypothesis from failure modes. State: what to run/observe, expected result if claim is true, expected result if a failure mode is real, why this can't accidentally pass under the failure. Cover: null hypothesis, silent fail, env mismatch." }),
       evidence_files: Type.Optional(Type.Array(Type.String(), { description: "File paths to inspect (must exist)" })),
       remaining_uncertainty: Type.Optional(Type.String({ description: "What's NOT tested, known limitations, edge cases deferred. Be honest about scope boundaries." })),