Reading a score

The fields of a lab’s score, one by one, and what they mean.

Where to find the score

lab score, lab baseline and lab run print the score as JSON. It is also saved in a score.json file stored with the hidden answers, outside the lab directory, next to a Lean module that holds every accepted result.

An example

As an example, here is the real score of the agent’s run on seed 7, rules hidden (reformatted onto fewer lines).

Output

{
 "lab_id": "20261001-031725-8f1cad",
 "mode": "blackbox",
 "used": {"proposals": 90, "cell_updates": 888},
 "budget": {"proposals": 200, "cell_updates": 200000},
 "verdicts": {"certified": 90},
 "conservation": {"found": 20, "hidden": 20},
 "structures": {"found": 70, "hidden_within_bounds": 70, "beyond_bounds": 0,
                "bounds": {"width": 7, "period": 8}},
 "agent": {"name": "claude", "model": "sonnet", "exit": 0, "seconds": 221,
           "cost_usd": 0.428, "turns": 13, "audit": {"clean": true, "markers": []}}
}

The fields

Field What it means

used, budget proposals and cell updates used, and their caps

verdicts number of proposals per verdict: certified, refuted, trivial, invalid, repeat

conservation.found law dimensions found, counted as added rank, trivial laws excluded

conservation.hidden law dimensions in the hidden answer

structures.found structure species found that are in the hidden answer

structures.hidden_within_bounds species in the hidden answer, exhaustive within the bounds

structures.beyond_bounds verified species found outside those bounds

structures.bounds the answer’s bounds: maximum width and period

agent name, model, exit code, duration, cost, number of turns, audit

Field	What it means
`used`, `budget`	proposals and cell updates used, and their caps
`verdicts`	number of proposals per verdict: `certified`, `refuted`, `trivial`, `invalid`, `repeat`
`conservation.found`	law dimensions found, counted as added rank, trivial laws excluded
`conservation.hidden`	law dimensions in the hidden answer
`structures.found`	structure species found that are in the hidden answer
`structures.hidden_within_bounds`	species in the hidden answer, exhaustive within the bounds
`structures.beyond_bounds`	verified species found outside those bounds
`structures.bounds`	the answer’s bounds: maximum width and period
`agent`	name, model, exit code, duration, cost, number of turns, audit

How to read it

A structure species is counted once per background, minimal period and velocity.
Width bounds depend on the world: 7 for three-state worlds, 6 for four-state worlds, with a period of at most 8.
beyond_bounds counts real, verified findings, but the hidden answer cannot say how many exist: this field has no denominator.
audit.clean is true when no marker of access outside the lab was found in the transcript. It is not a sandbox.
A score is compared with the scripted reference at the same budget, not in absolute terms.