Research program / August 2026 → February 2027

From leaderboard
to mechanism.

Our north star is not the strongest chess harness. It is a reproducible chain from observation to mechanism to intervention to held-out replication.

01
“A scalar score can tell us that a system changed. It cannot tell us which part learned to see.”

Chess is the measurement environment. The objects of study are broader: grounded state, proposal quality, comparative selection, search, value estimation, uncertainty, and long-horizon coherence.

Elo remains the external-validity test. Paired decision experiments are the main instrument for causal understanding.

Mechanism map

Five failures hiding inside one move.

02

Latent failure stage

Can the agent reconstruct the board it is acting on?

Instrument

Matched FEN, board, piece-list, and deliberately inconsistent interfaces.

What would prove us wrong

If representation changes legality but not move quality, grounding was not the main bottleneck.

Featured causal instrument

A105 holds the chess state fixed and randomizes access to redundant board and legal-action representations.

Open the experiment explorer

Portfolio

Six linked workstreams.

03
01

External validity

Can the complete agent survive the consequences of its own decisions?

INSTRUMENTFull games
STATEFrozen Luna; DeepSeek remains provisional
02

Proposal

Does a strong move enter the model’s consideration set?

INSTRUMENTNested recall
STATERetrospective bridge; paired replication next
03

Selection

Can the model choose the best action already available?

INSTRUMENTPaired regret
STATEPreregistered
04

State

Does the interface preserve a grounded, legal board model?

INSTRUMENTRepresentation cross
STATEControls frozen; causal ablation designed
05

Search + value

Can verified consequences extend reliable decision depth?

INSTRUMENTTransition scaffold
STATEProspective
06

Efficiency

What strength is purchased per token, second, tool call, and dollar?

INSTRUMENTPareto frontier
STATELuna frozen; cross-model provisional

Critical path

Six months, gated by evidence.

04
00
Now

Freeze the current generation

Terminal cohorts, content-addressed evidence, no mutable dependency.

01
Weeks 1–3

Resolve the retrospective signal

Separate proposal regret from selector regret without causal overclaim.

02
Weeks 2–6

Build the paired substrate

600 locked positions, nested candidates, independent labels, grouped splits.

03
Months 2–3

Identify the mechanism

Candidate-width replication, overload controls, notation and order ablations.

04
Months 3–5

Search, value, architecture

Promote modules one at a time on held-out decisions.

05
Month 6

Replicate and release

Benchmark v1, model cards, correction log, paper, clean-room reproduction.

Standards of evidence

The result must be harder to publish than to imagine.

  1. 01Identical positions for causal comparisons.
  2. 02Randomized order, stored exactly.
  3. 03Scores and ranks hidden unless intervened on.
  4. 04Provider failures reported, never converted to losses.
  5. 05Every estimate travels with N and uncertainty.
  6. 06Every mechanism owns an explicit falsifier.
Read the evidence doctrine