CORE RULE
Community reports are leads, not truth.
GECKWiki, Nexus posts, plugin changelogs, source-atlas, NVSE symbols, and model interpretations can all point us toward an answer. None of them substitutes for measured PC evidence and, for high-confidence behavioral claims, a runtime experiment.
PIPELINE
Seven gates
Capture the claim
Record the exact reported behavior, source, platform, date, and any claimed workaround. Do not silently rewrite the report into a root-cause theory.
Check current status
Look for maintained fixes in JIP LN, lStewieAl's Tweaks, JohnnyGuitar, YUP, TTW, kNVSE, ShowOff and other active projects. Classify duplicates and partial fixes.
Make it deterministic
Build the smallest vanilla test that produces the bug. Record build identity, load order, save state, inputs, expected result and observed result.
Find the engine path
Use mnehmos.re, Ghidra, xrefs, call graphs, NVSE PDB context, source-atlas candidates and FNVEdit record truth to locate the responsible path.
State the narrowest cause
Promote only what the evidence supports. Preserve alternative explanations and retractions. Semantic names remain provisional until independently supported.
Change as little as possible
Prefer a minimal NVSE hook, branch correction, missing lifecycle call, or data fix. Keep the patch separable from the explanation.
Try to falsify the fix
Repeat the original reproduction, negative controls and neighboring cases. High-confidence claims require cited runtime artifacts under the RE confidence policy.
Status vocabulary
Seed: listed by a source; not independently reproduced. Investigating: current evidence/reproduction work is active. Partially mitigated: an existing fix addresses some but not all reported cases. In the RE desk, static chain recovered means the path is documented but not runtime verified, while open means the requested consumer is not yet established. Future states will include reproduced, root-caused, patched, runtime-verified, fixed-upstream, debunked and intended-behavior.
Confidence discipline
The engine KB is stricter than the public bug catalog. Model judgment or community naming alone cannot justify high confidence. Static concurrence can support medium confidence; runtime evidence is required for high-confidence behavioral claims. If later evidence contradicts an interpretation, the semantic claim is retracted while the underlying measurements remain in the audit trail.