feat(sn60): screener-safe DeepSeek depth miner (jimcody1995-20260719-01)

[jimcody1995]

Summary

• Replaces PR feat(sn60): evidence-first DeepSeek depth miner (jimcody1995-20260718-01) #161 (kata:invalid) — round-start screener failed with room connection drop / non-JSON report after a 5-call budget that could overrun the TEE.
• Hardened for the execution screener: max 3 inference calls, 700s run cap, adaptive HTTP timeout, always returns {\"vulnerabilities\": [...]}.
• Pipeline: static rank → map triage → two deep audits (value + auth). DeepSeek-V3.2-TEE via Chutes-sealed key.
• Challenges current king bohdansolovie / PR feat(sn60): DeepSeek depth-audit miner (bohdansolovie-20260717-01) #156.

Test plan

• Intake screening → kata:pending
• Round-start execution screener completes with valid JSON
• Round scoring vs king / other challengers

Made with Cursor

[jimcody1995]

Pushed an upgrade aimed at the current king (#160 / Dexterity104, 16 TP):

• Full multi-language discovery: Solidity / Vyper / Rust / Move / Cairo
• Triage → contiguous deep audit → risk-windowed wide second pass
• High-precision structural supplements (no fingerprint banks)
• Still screener-safe: 3 calls / 760s, always returns JSON
• Chutes-sealed DeepSeek-V3.2-TEE credential rebound to this bundle

Ready for re-screening / next round.

[carlos4s]

Thanks for the PR. I’m pausing this one before it enters a challenge.

Short version: the miner looks too benchmark-shaped. The bot overlords are suspicious, and this time they brought receipts.

What tripped the wire:

• This PR has kata:pending, but the current head commit has not passed intake screening. This usually means new commits were pushed after screening, or an intake webhook was missed. Re-run /kata review before the PR can enter a competition challenge.

Why this matters: Kata agents need to analyze unseen projects. If the code recognizes benchmark fingerprints and spits out canned findings, that is replay, not mining.

How to fix it: remove project/finding-specific branches, static report banks, copied titles, and benchmark-only pattern matches. Keep reusable detectors and let the agent reason from the project source.

Maintainer commands: /kata approve accepts it, /kata review re-runs screening, and /kata close closes it.
You can also push an update and the checks will run again.

[carlos4s]

/kata review

[carlos4s]

Review re-run by carlos4s.

Screening passed. I marked this PR as kata:pending.

[carlos4s]

This challenger beat the current king and is safe to merge.

Submission: submissions/sn60__bitsec/miner/jimcody1995-20260719-01
Challenge run: pr-162

Score comparison:

• Primary: detection candidate 28.6% vs king 14.3%; delta +14.3; PASS projects 2/7 vs 1/7.
• Result: The candidate beat the current king in the SN60 sandbox duel.
• Perfect PASS projects: candidate 2 vs king 1
• Invalid/error evaluations: candidate 0 vs king 0
• True positives: candidate 5/16 vs king 3/16
• Precision / F1: candidate 5.8% / 9.8% vs king 10.7% / 13.6%

Merged. Your agent is the new king, and the temporary submission directory has been cleaned up from main. Nice work.