A JSON determining what the agent may do: names, typed parameters, budgets, gates, the goal.
A general agent under a referee. In any world you connect, the agent discovers what its tools do, works its goal into subgoals and internalizes what it learns. The referee makes it predict before it acts, grades every prediction in code and keeps everything on the record. First benchmark: frontier performance on ARC-AGI-3, along with its audit trail.
An agent with these capabilities is different from the LLM it runs on. The same constitution, conditions, budgets, and audit run identically across worlds. ARC-AGI-3 was the first benchmark. Factorio and OOLONG publish next under the identical standard.
| System | Backbone | Regime | RHAE | Games | Levels | Actions | Audit trail |
|---|---|---|---|---|---|---|---|
| arc-skill | Opus 5 | uncapped | 100.00 | 25/25 | 183/183 | 7,645 | — |
| ASSAY (ours) | Opus 5 | hard caps, gated | 96.54 | 24/25 | 177/183 | 8,157 | full |
| Prime Agent (median card) | Opus 5 | uncapped | 95.24 | 24/25 | 178/183 | 11,245 | — |
| PRO-LONG | Fable 5 | uncapped | 94.71 | 19/25 | — | 11,156 | — |
Same 25 game instances, verified by id. PRO-LONG ran on a stronger backbone and is directional context only. Prime Agent's best-of-3 reaches 183/183; ASSAY's numbers are single-protocol, no best-of-N.
The consolidated verification replay of all 25 recorded action sequences is a public ARC scorecard. Per-game action counts and per-level outcomes are on the card.
A broker daemon owns the environment session and its credentials. The agent process, it doesn't hold them.
The key to success was the record we've created with Assay that proves agent does exactly what it says and nothing is done without being recorded. We've created a harness that the journal is append-only under a rolling hash chain with external anchors. A third party recomputes integrity from the artifacts alone. You can audit it transparently and there is not depdency to a specifiec world.
The checker is one Python file. Verifying the dc22 journal against its published head takes three commands.
On six occasions the agent proved a level impossible. Every proof was consistent with every recorded transition. Every proof was wrong and wrong precisely where no transition had ever gone.
“The switch is unreachable; the remaining blocks are isolated.”
All 74 inert probes had been taken in a single world-state. One coordinate had been inert-then-live earlier — the gate was geometric arming, and the blocks were connected.
“Level 7 is provably unwinnable.”
The load-bearing rotation rule had only ever been exercised in the occupied failure mode. One graded probe of the off-board case showed rotation clips rather than refuses — un-trapping the arm and breaking level 8's length cap too.
“Unsolvable — by exhaustive search over 8.2 million states.”
The search was fully correct given its transition rules. One of them — freed blocks re-hook bottom-only — fit all 103 recorded transitions and had never been graded. Re-hooking is positional; the action written off as a no-op was SELECT.
“Throughput ceiling; hostiles effectively unkillable.”
All three hostile rules were false. The kill action works on any adjacent hostile — the prior session had seven free kills in front of it — and movement is one orthogonal cell per action.
“No switches remain; the map is certified pixel-complete.”
Nine switches sat above the map, visible only from inside the “dead-end” shaft. The deepest instance: the unexercised premise was the observation frame itself — “rendered” standing in for “exists”. Won at 95.27; running over baseline on late levels costs 0.19 set points.
“The switch is unreachable.”
The switch sat above the drifted map — the map had drifted, not the world. Two of the six proofs broke on this one game.
Five of the six converted to wins inside their existing action budgets — cheap precisely because the audit step is free: it runs offline over the journal. The sixth remains open (roadmap).
Published work on this problem since early 2026. The progression Sensi → ARG → ASSAY, in public, with the negative results left in.
The philosophical companion to the exercised-claim finding. Written in February about the epistemology of uncertainty; in August the harness broke six impossibility proofs by auditing what had never been exercised.
Read on Substack →Timestamped priority on the problem formulation.
Read on Substack →Lineage, and the negative-results discipline on display.
Read on Substack →Test-time-compute foundation.
Read on Substack →Test-time-compute foundation.
Read on Substack →A long-horizon industrial world attached the same way ARC was: one adapter, one action registry, the same kernel and constitution unchanged.
The one game ASSAY did not clear: 4/10 levels at 364 actions, n=3, with a located diagnosis. It costs 3.27 of the 3.46 set points between ASSAY and 100.
Independent researcher. Builds agent instruments where the referee, not the agent, decides what counts as knowledge — and where the record is the product. Other works
Reasoning, test-time learning, verified autonomy. I build agents that can prove what they did.
Request access →