Pattern Mining

Pattern Mining

Decide whether repeated code demonstrates a reusable rule or only a plausible

hypothesis. Similar names and syntax are not enough; the abstraction must

survive examples it was not designed around.

Constraints

• To prevent lineage copies from faking recurrence, use independently

implemented exemplars with repository anchors.

• Because the candidate must generalize, form it without seeing the holdout and

back-apply every holdout-driven refinement.

• To keep weak evidence from becoming architecture, route hypotheses to

no-action; only a fully proven promotion may reach operationalize.

Workflow

• State the candidate pattern and collect independently implemented

exemplars with repository anchors. Use research when coverage is unclear.

• From the exemplars, separate required invariants, legitimate variation

points, and incidental similarity.

• Require at least three distinct exemplars before promotion is possible.

Form the candidate abstraction without using the holdout.

• Test it against every exemplar, then a separate holdout. Back-apply the

refined abstraction to the original exemplars so the holdout fix cannot

silently break them.

• Emit outcome: promote only when the exemplar floor, holdout, and

back-application all pass. Route that evidence to operationalize, which

decides whether the eventual shape is a skill, gate, library, template, or

no action.

• Otherwise emit outcome: hypothesis with route: no-action. Keep the

evidence bounded and name what additional observation would retest it.

Diff/align across exemplars

Invariants are extracted mechanically, not remembered. Lay the exemplars side

by side, align them structurally (same role, same position in the flow — not

same variable names), and diff: what survives every alignment is a candidate

invariant; what varies by site is a variation point; what varies with no

functional consequence is incidental. Work pairwise before generalizing — an

"invariant" derived by skimming all exemplars at once is usually the first

exemplar's shape with the others squinted into agreement. Stop condition: every

line of the candidate abstraction is traceable to a surviving alignment across

all exemplars, or it is deleted. The named failure mode is eyeball

convergence — declaring similarity from memory of the files rather than from an

explicit alignment, which smuggles one lineage's incidentals into the rule.

Explicit search recipes

Exemplar discovery is part of the evidence and must be replayable. Record the

exact queries used — rg patterns, glob scopes, structural searches — in the

output packet, alongside which hits were kept and why the rest were excluded.

A recipe that returns hits you did not examine is unfinished coverage: either

examine them or narrow the recipe and record the narrowed form. The named

failure mode is convenience sampling — mining only the files already in

context, which biases the exemplar set toward one author or one era of the

codebase and fakes independence at step 1. If no recipe can be written that

finds the exemplars, the recurrence claim is unverifiable and the outcome

stays hypothesis.

Packaging-shape catalog

A promoted pattern lands in exactly one shape, and naming the intended shape

before promotion sharpens the holdout test. The catalog, in ascending

commitment: no-action (evidence retained, nothing built), **checklist

line (rule in an existing doc or skill), template** (copyable exemplar),

library/helper (shared executable code), gate (deterministic check

that blocks). Match commitment to evidence strength: three exemplars and one

holdout justify a checklist line or template; a gate needs demonstrated cost

of violation, not just recurrence. This skill only records the recommended

shape in the packet — operationalize owns the packaging decision. The named

failure mode is shape inflation: routing a barely-promoted pattern straight to

a gate or library because building feels like progress, which turns weak

evidence into architecture the same way skipping the holdout would.

Output Specification

• Artifact directory: .agents/scratch/pattern-mining/<run-id>/
• Filename convention: pattern-mining.json
• Format: pattern-mining.v1 JSON containing the outcome, distinct

exemplars, invariants, variations, incidental details, holdout result,

back-application result, and route.

• Validation command: skills/pattern-mining/scripts/validate-output.sh <pattern-mining.json>
• Downstream handoff: pass a validated promote artifact to

operationalize; retain a validated hypothesis artifact as bounded

evidence with route: no-action.

Promotion requires at least three distinct exemplars, one separate passing

holdout, successful back-application, and at least one invariant. Any weaker

packet remains a hypothesis and cannot route to reusable packaging.

The validator is the machine boundary:

skills/pattern-mining/scripts/validate-output.sh <pattern.json>

This skill owns evidence for the pattern. It never creates the reusable

artifact itself and never promotes a failed or untested holdout.

Executable behavior:

references/pattern-mining.feature.

Quality

• Exemplars are independent and repository-anchored; copied implementations do

not inflate the evidence floor.

• Invariants, legitimate variations, and incidental similarities stay distinct

through holdout testing and back-application.

• The named validator passes before a promotion reaches operationalize or a

hypothesis is retained as no-action evidence.

Do not

• Count copies from one implementation lineage as independent exemplars.
• Hide variation by calling it incidental.
• Route a hypothesis directly to a skill, rule, gate, or library.