How Does a Historical Case Become an Executable Professional Method for the Agent?

Why Incident Post-Mortem Materials Require Re-Structuring

Incident reports typically emphasize conclusions and containment actions, whereas future investigations require knowing what was reviewed first, why specific methods were chosen, what data was missing, and what evidence altered judgment. Candidate methods can be extracted from reports, specs, tool manuals, and expert interviews without requiring an exhaustive historical database first.

Five Essential Tasks: From Incident to Method

• Distinguish facts, hypotheses, executed actions, and verification outcomes.
• Declare applicability conditions across materials, processes, tools, and metrology.
• Specify required inputs, comparative baselines, and missing data policies.
• Preserve inspection sequences, dependencies, method selections, and decision boundaries.
• Peer review by experts, verified using normal lots, excursions, negative controls, and novel conditions.

For example, an edge residue investigation involving maintenance changes produces a reusable method: 'verify pre/post detections, reference baselines, and maintenance logs,' rather than a brittle rule: 'edge residues always require replacing a seal ring.'

Separating Method Specs from Live Instance Data

Method: Preliminary Edge Anomaly Screening
Scope: Inspection coverage and reference baselines verified
Inputs: Images, inspection recipes, existing analyses, tool logs
Checks: Verify inspection conditions -> Compare zone divergence -> Check process drift
Bounds: Correlation does not directly confirm component failure
Missing Data: Log evidence gaps, formulate verification requests
Governance: Owner, version hash, positive/negative tests, revision rationale

Method specs can be authored directly by experts or drafted by AI models for expert revision. Models do not automatically prove method correctness. Current Rails govern engineering judgment scope and available methods, not a rigid script forced across all cases.

Method Validation: Preventing Outdated Conclusions on New Problems

At minimum verify: do positive cases yield grounded recommendations, do normal baselines remain quiet, do visually similar negative controls preserve alternative explanations, do missing datasets halt or request evidence, and do revisions avoid regression errors? Historical post-mortems may carry uncertainty and cannot serve unconditionally as ground truth.

Validation thresholds and sample sizes depend on the application, risk, and data. Without sufficient normal samples, a one-in-a-thousand false-alarm rate cannot be claimed; passing a small historical replay does not confer production qualification.

Version-Controlled Methods That Embrace Human Correction

Record the reason for changes, affected scope, and applicable conditions, and continue evaluating with new data. The value is reducing repeated exploration and preserving team methods, not guaranteeing that a new engineer equals a senior expert or automatically delivering fixed yield or cycle-time gains.

Related Reading & References

Below are verifiable industry references; specific processes, permissions, and acceptance criteria are governed by fab site specifications.

• NIST: Randomized Block Designs