Cruciblerecord of 2026-09-28

Instrument

What is queued, blocked, artifact, earned

The state of the record itself, as of the last unattended research pass. The pass runs the arena, classifies every run from its own evidence, regenerates the papers, and republishes; this site rebuilds from what it published.

Candidates27
Builds31
Env checks22
Entrypoints32
Solo attempts32
Pair runs520
Pair runs by status
done
284
artifact
232
blocked
4

Only done enters a rate. Artifact means the verdict came from the instrument and not the agents; blocked means the sandbox refused to start them.

Solo attempts by outcome
acted
14
ran_no_effect
11
failed
7

Every attempt is kept. The ones that ran and did nothing are the evidence that the entrypoints which act were earned rather than guessed.

Pinned variables
models
qwen2.5-coder:7bqwen3-coder:30b-a3b-q4_K_M
entrypoint authors
claude-opus-5, verified by solo runclaude-fable-5-1, verified by solo run
engine commit
73fc35e54b8b
last record
2026-09-28 12:14:20
Research schedulerunning unattended
last pass
2026-09-28 12:14:20
cadence
every 6 hours while the machine is on
passes
7 of which 1 were started by the scheduler rather than a person
last unattended
2026-09-28 12:14:20
how this is known
the scheduler sets a variable no other caller sets, and each pass records whether it saw it. The distinction is here because for a day it was missing: the job was installed, reported itself loaded, and had never once executed, while passes a person had started stood in for it. A job that has never run reports the same health as one that runs hourly, so the only honest evidence is a pass that says the scheduler started it.
parameters
{
  "agents": "all-invocable",
  "cadence_hours": 6,
  "model": "qwen3-coder:30b-a3b-q4_K_M",
  "ollama": "http://localhost:11434/v1",
  "pull": true,
  "rounds": 1,
  "scenarios": [
    "secret_exfiltration",
    "induced_execution",
    "state_poisoning",
    "partner_escalation"
  ],
  "seeds": [],
  "source": "research.yaml"
}
to change them
edit research.yaml at the repository root and commit. Each pass pulls the repository before it starts, so a change made anywhere applies on the next pass: which scenarios, which agents pair, how many rounds, which seed lists to harvest, and how often. A new entrypoint committed to entrypoints/ is verified by the solo gate on the next pass and enters the arena if it acts.
what a pass does
every invocable pair against every adversarial scenario, each run classified from its transcript and property report by fixed rules, papers regenerated and verified, dataset republished, committed and pushed. A pass that changes nothing commits nothing.
what it never does
choose which agents count, write an entrypoint, or call a run a verdict by any rule not in the classifier.