Clone check

Two builds are not the same task because they share a device. They are the same task when the thing a solver has to see is the same thing.

Part 0. What this is for

The final_verdict reviewer excludes a package that fingerprints against anything delivered, and the client has said these builds read as templated. The question is whether a reviewer holding both builds learns nothing new from the second, which is about the driver, not the surface. The check reports at the severity a reviewer would attach and never repairs; a clone is answered by a redraw, which is the author's.

Where it runs. Every build runs the free half, guard.py heart and guard.py surface, at stage 3. The adjudicated check (the clone-auditor agent, medium effort) runs only when the author asks, through /clone-check, because it spends tokens. The corpus baseline is corpus/fingerprints.json (tracked); the working cache under .cache/ is seeded from it and never committed. The author refreshes the baseline on a checkout holding every pack: fingerprint.py extract --out .claude/skills/clone-check/corpus/fingerprints.json. The mechanical screen is fingerprint.py extract then fingerprint.py screen (default the last 15 builds), whose hard tells (byte-identical files, same-seed tables surviving a rename, identical column sets, shared multi-word invented names, masked prompt wording) route straight to adjudication.

Part 1. The discrimination

• Reusing a component is legitimate; reusing the driver is not. A proven device is meant to be spent again. The driver, the one insight the solver has to reach, is what a reviewer remembers and a solver can be primed against. Familiar moves in a new order with a new decisive move is a new task; new moves arriving at the same insight is the old task reskinned. A similarity score ranks these backwards, so the check is never a score.
• The stumping strategy is not a component. Different drivers can share a strategy (a withheld rule under an official decoy; a certified record blessing the wrong method until a forward condition breaks it). The client names the first as about 40 percent of what it receives and rejects replicated approaches, so a strategy carrying a large share of a batch is a finding even when every pair is clean.
• The prompt shape is a component (guide-to-prompt/references/shapes/): it supplies the 25 criteria and will repeat. Score a shared shape as Track A; escalate only when it hides the same decisive rung. It is very visible, so vary it across a batch and check it in pre-flight.
• Read the position, not the inventory. A device decisive in one build and at rung 1 in the other is demoted, which is reuse. Decisive in both is spent twice, whatever the surface distance.

Part 2. Two tracks, never merged into one score

Track A, what a reviewer sees: prompt wording, file names, schemas, row counts, values, invented names, dated windows, deliverable species, write-up phrasing. A Track A hit against a delivered task is exclusion level by itself. The repair is a rename, not optional. Run ../guide-to-prompt/references/voice-check.py bare on a batch first: the script measures the surface (question forms, repeated phrases, length, file count, format mix) and this skill adjudicates. A monoculture spread thinly is invisible pair by pair and obvious in the counts (eighteen prompts, no two clones, one skeleton).

Track B, the mechanism: gap, pattern, decisive rung, calibration form, decision family. A hit means the draw failed; the repair is a redraw that usually invalidates the data, so catch it early. Check the mechanism axes independently, never behind a surface threshold: a fully reskinned surface over a repeated driver scores near zero on surface and is the worst case.

Part 3. The tiers, by cost to answer

Tier	What it is
Exact	The same artifacts.
Same seed	Renamed surface over the same generated values (row counts, column sets, numbers survive). Reads as hiding.
Dataset	The same spine dataset or schema; banned between consecutive builds (stumping Part 6).
Template	The same driver: gap, pattern and decisive rung line up whatever the nouns. The tier the client's note was about, and the only one surviving a full rename.
Shape echo	Different mechanism, same outline (ask sentence, deliverable species, decision shape, calibration organ). Soft flag.
Idiom	Author habits across the corpus (a file role, an ask construction, a deliverable pairing). Corpus wide, reported apart from pairs.
Strategy monoculture	Read each card's stump_strategy, decisive_rule_filed, official_decoy, validated_decoy (design notes for older builds). A finding, builds named, when the withheld rule with an official decoy exceeds two in any twelve, the certified-record decoy exceeds three, or any one strategy holds a third of the batch. Corpus wide.
Clean, shares components by design	A verdict, stated out loud when devices are shared and drivers differ, so the author keeps reusing what works.

Part 4. Evidence that is not there

Read the fingerprint cards first (../fingerprint/cards/): one vocabulary, a domain-stripped driver sentence, and mechanism_source: submission_only where the mechanism was never recorded. Back-fitted labels are accurate about the mechanism and approximate about the name, so two builds sharing one share a guess. Where the mechanism cannot be read, report the pair as checked on Track A only. A driver reconstructed from a prompt is a fabrication and sends an author to redraw a sound build.

Part 5. How the check goes wrong

• Flags legitimate reuse. Proximity is not a finding; no Template finding without the shared decisive move named in one sentence.
• Misses the reskin. The mechanism axes need their own route to adjudication.
• Scores instead of deciding. A number is triage; the verdict comes from reading both builds.
• Treats the corpus as flat. A collision with a delivered build excludes; with a failed or retired build it is weaker; with the slot's own earlier architecture it is not a finding.

Part 6. What a finding carries

The tier, the track, the two builds and both delivery statuses, the shared thing stated as a mechanism in one sentence (or it is not a finding), and the one axis to redraw. Say when a finding is Track A only: the repair is a rename, not a rebuild.