fix(prism): fair review for generic LM tropes

[echobt]

Summary

• Prod rejects for 6d0824f9… / b19998da… were cheap LLM similarity-v2 false positives (RMSNorm/RoPE/SwiGLU evidence at score 0.7), not AST copy-gate or agentic. AST vs peers/baseline is ~6100–7400 bps (under AST_SUSPICIOUS_BPS=8500).
• similarity-v3 + parser coercion: generic modern-LM components are never plagiarism evidence; pre-pod hard-rejects Copied only (Suspicious advisory).
• Copy / LLM similarity / agentic corpus = champions only (Score>0 tops + ex-tops) + baseline.
• Public site gallery/leaderboard hide non-champions; operator challenge API still lists all.
• Prod ops: reopened prism gating to open for both miner hotkeys (no scores invented).

Test plan

• cargo test -p prism-review -p prism-challenge -p site-data -p site-api -p challenge-agentic -p prism-store -p prism-pipeline
• CI green
• Deploy prism-challenge + gateway (site filter); confirm precheck still works
• Miners can submit again under same hotkeys

Summary by CodeRabbit

• New Features

  • Public Prism galleries and leaderboards now show only current and historical champion submissions with positive scores.
  • Similarity reviews use champion architectures and baseline references while excluding standard language-model components as plagiarism evidence.

• Bug Fixes

  • Suspicious similarity findings are now advisory instead of automatically producing a zero score.
  • Clear copying and agentic cheating or suspicious behavior continue to trigger hard-zero scoring.
  • Generic architecture terminology is no longer treated as sufficient evidence of copying.

• Documentation

  • Updated Prism scoring, review, corpus, and public API guidance to reflect these changes.

[coderabbitai]

Caution

Review failed

The pull request is closed.

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro Plus

Run ID: bc67e59b-6ce6-40dd-aa83-8a63c7fd2cfc

📥 Commits

Reviewing files that changed from the base of the PR and between b0ef8b0 and 8b37134.

📒 Files selected for processing (21)

• crates/challenge-agentic/src/prompts.rs
• crates/db/src/prism_store.rs
• crates/prism-challenge/src/api.rs
• crates/prism-challenge/src/orchestrator.rs
• crates/prism-challenge/src/score.rs
• crates/prism-challenge/tests/copy_gate.rs
• crates/prism-pipeline/src/precheck.rs
• crates/prism-review/prompts/similarity_v3.md
• crates/prism-review/src/generic_arch.rs
• crates/prism-review/src/lib.rs
• crates/prism-review/src/llm.rs
• crates/prism-review/src/prompts.rs
• crates/prism-store/src/dbprism.rs
• crates/prism-store/src/store.rs
• crates/site-api/src/handlers.rs
• crates/site-data/src/map.rs
• docs/COMPLETENESS.md
• docs/PRISM.md
• docs/PRISM_RECIPE.md
• docs/SITE_API.md
• docs/external-miner/prism.md

---

📝 Walkthrough

Walkthrough

PRISM now limits copy and review corpora to positive-score champion submissions, updates similarity review to v3, ignores generic language-model components as plagiarism evidence, makes suspicious similarity advisory, and exposes champions only in public listings.

Changes

PRISM review and champion filtering

Layer / File(s)	Summary
Champion corpus retrieval and wiring crates/prism-store/..., crates/db/..., crates/prism-challenge/..., crates/prism-pipeline/...	The store adds champion listing support. Copy, agentic, and similarity flows use champion submissions instead of general or terminated submissions.
Similarity prompt and generic-architecture handling crates/prism-review/prompts/similarity_v3.md, crates/prism-review/src/..., docs/COMPLETENESS.md, docs/PRISM_RECIPE.md	Similarity v3 defines architecture-only judgments. Generic language-model evidence is coerced to Original under the defined confidence thresholds.
Scoring and anti-cheat verdict handling crates/challenge-agentic/..., crates/prism-challenge/..., docs/PRISM.md, docs/external-miner/prism.md	Only Copied similarity results hard-gate scoring. Agentic Cheat and Suspicious results remain hard-zero outcomes.
Public champion filtering crates/site-data/..., crates/site-api/..., docs/SITE_API.md	Public Prism submissions and BPB leaderboard entries now include positive-score champion submissions only.

Estimated code review effort: 4 (Complex) | ~45 minutes

Sequence Diagram(s)

sequenceDiagram
participant Candidate as Candidate submission
participant PrismStore
participant SimilarityReview as Similarity review
participant Scoring as Scoring gate
Candidate->>PrismStore: Load positive-score champion corpus
PrismStore-->>SimilarityReview: Return champion submissions
SimilarityReview-->>Scoring: Return Copied, Suspicious, or Original
Scoring-->>Candidate: Apply hard gate or advisory result

Loading

Possibly related PRs

• BaseIntelligence/base#82: Modifies PRISM similarity corpus construction in related challenge and store code.
• BaseIntelligence/base#85: Covers related PRISM pre-pod similarity and agentic anti-cheat handling.
• BaseIntelligence/base#89: Adds related precheck and corpus helpers used by this champion-listing change.

✨ Finishing Touches 💡 1

🛠️ Fix failing CI checks 💡

• Create stacked PR
• Commit on current branch

📝 Generate docstrings

• Create stacked PR
• Commit on current branch

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch fix/prism-agentic-tops-only

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}