Reproducible recipe

Interpretability probes

Diagnose a measured model failure without treating a visualization as causal proof.. Aggregate scores cannot explain which layers, tokens, or representations drive a failure.

Status
reference only
Learning path
interpretability
Regression slices
4
Execution
Fresh experiment required

Target

Diagnose a measured model failure without treating a visualization as causal proof.

Failure this recipe addresses

Aggregate scores cannot explain which layers, tokens, or representations drive a failure.

Data contract

A frozen set of successful and failed traces from one checkpoint.

Method or policy

Use logit lens, probes, activation patching, causal tracing, or SAE analysis with an explicit hypothesis.

Evaluation contract

Prediction change under a controlled intervention, not visual salience alone.

  • control prompts
  • random intervention
  • layer/token stability
  • replayability

Budget and stop rule

Closed as a learning/reference lane; future use must be attached to a specific model failure.

Stop if the probe has no discriminating prediction or intervention control.

Decision rule

Use findings only to change a recipe when a controlled intervention supports them.

Learning exercise

Inspect one logit-lens or activation-patching fixture and write the causal claim it does and does not support.

Separate correlation, localization, and causal intervention.

How to read this dossier

This canonical page separates the retained repository decision from the material that informed it. A source may describe an external claim, historical measurement, local observation, or planned procedure; those evidence classes are not interchangeable. Follow the provenance links before reusing a number or method.

Status and disposition describe what PostTrainLLM retained when this record was indexed. They are not a live product promise, a newly run benchmark, or permission to restart historical work. Unknown or unmeasured fields remain unknown rather than being treated as zero.

To reuse the record, first name the exact claim you need and trace it to the linked source. Then check whether the original environment, model revision, data split, evaluator, hardware, and budget match the proposed use. If they do not, treat the record as a hypothesis or design reference and run the smallest fresh comparison that can falsify it. Preserve negative outcomes and regressions beside any improvement; a local win on one slice does not silently become a general capability claim.

Source provenance

The normalized record comes from docs/recipes/registry.json. The links below are the tracked evidence and explanatory sources preserved with the record.