feat(sn60): add Dexterity104-20260723-01

[Dexterity104]

Adds one miner bundle for the sn60__bitsec miner lane.

[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:

• agent.py: Screening review required: submission agent is highly similar to the current lane king implementation (similarity 0.94).

LLM sanity check:

• Reason: I do not see benchmark IDs, canned answer text, or leaked validator secrets in the shown source. However, the deterministic similarity hit to the current king agent is strong enough that this needs manual review for copycat behavior rather than a clean pass.
• Verdict: suspicious. Confidence: 58% medium - plausible concern; manual review needed.
• Audit artifact saved for maintainer review.
• Proof: File header and overall structure present a polished multi-pass auditor, but deterministic screening found very high similarity to the current king agent (0.9398), which is the primary concern under the rules. (agent.py:1)
• Proof: Large prompt blocks (SYSTEM/CHECKLIST/ENUM_TAIL/LOCALIZE/JSON_RULES/SCHEMA and the triage/deep/focused/recheck prompts) form a specific multi-pass audit scaffold consistent with a copycat risk rather than a minimal independent implementation. (agent.py:90)
• Proof: The inference request path uses only the validator-provided inference proxy via INFERENCE_API / INFERENCE_API_KEY, this is not a secret leak and weighs against rejection. (agent.py:350)

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]

A new commit was pushed while this competition challenge was running, so the challenge result belongs to an older version of this PR. I left the PR open and marked it kata:stale; re-run /kata review for the latest commit before the next challenge.

[carlos4s]

This PR beat the king, but it isn't currently safe to merge, so the king wasn't updated.

Reason: Kata command produced no JSON output: king promote --challenge-run /srv/kata-bot/work/promotions/Autovara__kata-pr197-960843e9c100/challenge_summary.json --submission-path /srv/kata-bot/work/kata-bot-preverify-9ua_m692/kata-bot-kata-preverify-6xdlvp1s/submissions/sn60__bitsec/miner/Dexterity104-20260723-01 --public-root /srv/kata-bot/work/kata-bot-preverify-9ua_m692/kata-bot-kata-preverify-6xdlvp1s --json
stderr: Using CPython 3.12.3 interpreter at: /usr/bin/python3
Creating virtual environment at: .venv
Building kata @ file:///srv/kata-bot/work/kata-bot-preverify-9ua_m692/kata-bot-kata-preverify-6xdlvp1s
Built kata @ file:///srv/kata-bot/work/kata-bot-preverify-9ua_m692/kata-bot-kata-preverify-6xdlvp1s
Installed 1 package in 1ms
Traceback (most recent call last):
File "", line 198, in _run_module_as_main
File "", line 88, in _run_code
File "/srv/kata-bot/work/kata-bot-preverify-9ua_m692/kata-bot-kata-preverify-6xdlvp1s/kata/main.py", line 4, in
raise SystemExit(main())
^^^^^^
File "/srv/kata-bot/work/kata-bot-preverify-9ua_m692/kata-bot-kata-preverify-6xdlvp1s/kata/cli.py", line 585, in main
return args.handler(args)
^^^^^^^^^^^^^^^^^^
File "/srv/kata-bot/work/kata-bot-preverify-9ua_m692/kata-bot-kata-preverify-6xdlvp1s/kata/cli.py", line 370, in handle_king_promote
result = promote_submission_result(
^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/srv/kata-bot/work/kata-bot-preverify-9ua_m692/kata-bot-kata-preverify-6xdlvp1s/kata/submissions/workflow.py", line 454, in promote_submission_result
raise ValueError(
ValueError: Submission is not safe to promote. Challenge result is stale because the king artifact has changed.

Once that's sorted out, a maintainer can run /kata promote to finish this already-decided win without re-scoring. If the cause was a merge conflict, rebase onto the latest default branch and push -- that re-runs intake and re-enters the PR to challenge the current king.

[carlos4s]

Thanks for the challenge! Your agent went up against the current king and came up just short this time, so I've closed this PR. Here's exactly how it played out:

Here's how you stacked up against the king's average over its last 2 challenges, in the order the crown is decided — reading top to bottom, you take the crown by clearly beating the king on a line — by more than that line's margin — without falling behind on a higher one:

#	Signal	You	King	Win margin
1	Project pass score projects solved on a 2-of-3 majority	0% (0/7)	7%	>14.3% ← decided here
2	Projects passed projects with at least one passing run	0	0.5	>1
3	True positives real bugs found out of bugs planted	0/32	2	>2
4	Fewer invalid runs runs that errored out — fewer is better	4	4.5	>2
5	Precision how many of your findings were real	0.0%	33.3%	>5%
6	F1 score overall balance of precision and detection	0.0%	12.9%	>5%

How the crown is decided: read the lines top to bottom. On each line — if you're behind the king, the king keeps the crown and you can't make it up on a lower line; if you clearly beat the king (by more than that line's margin) you take the crown; if you're close (within the margin), that line is a tie and it moves to the next one. So to take the throne you have to clearly beat the king on some line (pass score first) without falling behind on a higher one. This stops a single lucky run from flipping the crown.

(The king's numbers are averaged over its whole reign — 2 challenges — since that's what it's judged on, so beating it on a single run isn't enough.)

So it came down to project pass score (tier 1) — that's where you fell behind the king, and since you can't make up a deficit on a lower line, the king kept the crown. To take it you'd need to stay even with the king (within the margin) down to a line where you clearly beat it.

Tweak your agent and open a fresh PR whenever you're ready — you can absolutely take another run at the throne. 👑