Design spec: similarity/containment detection algorithm for plagiarism/copycat detection (#1969)

[JSONbored]

Problem: #1969 (plagiarism/copycat detection, effort: XL) has shipped only its first slice — PR #4140 added the gate.copycat.mode config scaffold, explicitly deferring "the actual detection engine (which carries real false-accusation risk and needs careful design)" to a documented follow-up. That design has not been scoped as its own issue yet, and #1969 itself has no design artifact to implement against.

Area: ORB / anti-slop

Proposal: Produce a concrete design for the similarity/containment scoring algorithm #1969 needs: what "containment" is measured against (repo history window, cross-PR corpus, or both), the actual scoring method (n-gram/AST-fingerprint/embedding-similarity — with a stated false-positive-rate target), the direction-by-timestamp attribution logic (earlier submission = victim) for edge cases (near-simultaneous submissions, forked lineage), and how the tiered response (warn → label → block → strikes) maps to concrete score thresholds. This is a maintainer-only design pass precisely because a wrong call here produces a false plagiarism accusation against a real contributor — the same reasoning gate.copycat.mode's own scaffold PR used to defer this exact piece.

Deliverables:

• A written design spec: algorithm choice + justification, false-positive-rate target and how it's measured pre-launch, attribution-direction edge case handling, and threshold-to-tier mapping.
• Explicit sign-off before any detection-engine implementation issue is filed/decomposed from Plagiarism / copycat detection #1969.

Acceptance criteria:

• The design spec is reviewed and approved by the maintainer before implementation work begins.
• The false-positive-rate target and its measurement method are stated concretely (not "low false positives" — an actual number/method).

Resources:

• Plagiarism / copycat detection #1969 (parent issue, Requirements/Deliverables sections).
• PR feat(review): add gate.copycat.mode config scaffold for copycat detection #4140 (gate.copycat.mode scaffold — the deferred-scope decision this design spec resolves).
• src/signals/duplicate-winner.ts (existing overlapping-scope duplicate clustering, a different signal this is a sibling to, not a replacement for).

Boundaries:

• Design only — no detection-engine implementation in this issue. Implementation should be filed as a fresh, decomposed issue (or issues) once this design is approved.

---

Part of #1969.

[linear-code]

JSO-3991

[JSONbored]

Resolved by PR #5999 (`feat(review): wire the copycat containment engine into the gate`, merged 2026-07-15), which shipped the full design this issue asked for — mapping each acceptance criterion to what actually landed in `packages/loopover-engine/src/signals/copycat.ts`:

• Containment measured against: same-repo prior art only (earlier open siblings + recently-merged PRs), bounded to 25 candidates (`src/queue/copycat-detection.ts`'s `MAX_COPYCAT_CANDIDATES`).
• Scoring method: deterministic 3-line code-shingle containment (`codeShingleList`/`containmentScore`) — asymmetric, candidate-shingles-found-in-prior-art — not embedding/AST-based, a deliberately simpler and more auditable choice.
• False-positive-rate target, stated concretely: `DEFAULT_COPYCAT_MIN_SCORE = 85` — only high containment (≥85%) trips the check, mirroring the existing miner-side self-plagiarism threshold. Mode defaults to `off` fleet-wide (byte-identical-when-off), so there's no live FP exposure until an operator opts in.
• Direction-by-timestamp attribution for edge cases: `copycatDirection` — the earlier submission is always the victim; any missing/unparseable timestamp or exact tie resolves to `ambiguous` and is excluded from consideration entirely, so it's structurally impossible to accuse the wrong side or a coincidental near-simultaneous filer.
• Tiered response → threshold mapping: `MODE_SEVERITY` (`off`→no finding, `warn`→info, `label`→warning, `block`→critical).

Parent #1969 is also already closed. No further design work needed here — this issue's acceptance criteria are fully satisfied by shipped code, not just a written spec.