Skip to content

Flagship experiments

Two ChemWorld tasks currently have formal multi-world Participant campaigns: Electrochemical Conversion and Reaction to Crystallization. Three additional tasks now have a five-world development-only comparison, bringing complete comparative execution to five tasks. This page records the design and results; see the authoritative current-status page for release readiness.

1. Questions

The experiments separate two questions:

  1. S0 v1.0, no dossier: how does the Participant compare with classical optimizers under anonymous materials and a fixed 20-experiment budget?
  2. S0 v1.2, three arms: does correct material information help, does a targeted error mislead, and can experimental feedback support recovery?

The nominal dossier contains correct but limited attributes for anonymous materials. It does not reveal the answer; performance still has to be learned through experiments.

2. Frozen common contract

Item Frozen value
Tasks Electrochemical Conversion; Reaction to Crystallization
Independent worlds seeds 0–9
Autonomous exploration 20 complete experiments per task × world × arm
Participant Codex subscription, gpt-5.6-sol, medium reasoning
Primary endpoint paired blind-validated final-recommendation score
Replay exact for every formal unit
Inference unit independent world, not provider call or algorithm seed
Arm Visible information Causal role
opaque anonymous IDs, no dossier no-information reference
nominal correct attributes attached to each ID correct information
misindexed one target field has a fixed two-row swap targeted wrong prior

Within each task and world, the arms share the semantic world, keyed observation noise, budget, model, scaffold, and blind endpoint. The paired contrast therefore targets the dossier condition rather than world difficulty.

3. S0 v1.0: Participant and classical baselines

Task Codex 95% world interval Best information-matched Best privileged calibration
Electrochemical Conversion 0.7150 [0.6283, 0.7861] Structured RF-EI 0.6159 Descriptor RF-EI 0.6441
Reaction to Crystallization 0.5355 [0.5045, 0.5644] LHS 0.5708 n/a

Electrochemistry has a descriptive +0.0991 difference from the best information-matched baseline, but is not stable against the best privileged calibration baseline. Crystallization is below LHS. No superiority threshold or multiplicity plan was preregistered, so these are formal descriptive results—not broad SOTA or provider-effect claims.

4. S0 v1.2: three-arm results

Value of correct information

Task Opaque Nominal Paired difference Familywise 97.5% interval Decision
Electrochemical Conversion 0.7150 0.7874 +0.0724 [+0.0074, +0.1546] positive information value
Reaction to Crystallization 0.5355 0.5615 +0.0260 [−0.0130, +0.0630] inconclusive

Wrong prior and recovery

Task Misindexed Misindexed − Nominal Manipulation Action correction Recovery to Opaque Overall
Electrochemical Conversion 0.6853 −0.1020; [−0.2101, −0.0078] pass pass fail fail
Reaction to Crystallization 0.5845 +0.0229; [+0.0046, +0.0419] pass fail pass fail

The joint recovery rule requires a material early behavioral effect, later correction away from the misleading action, and practical performance recovery to the opaque arm. Passing only one component is not evidence that the model identified and corrected the wrong dossier.

The higher misindexed crystallization score is only a benefit of this fixed swap in the sampled worlds. Because action correction failed, it is not evidence of error discovery.

5. Accounting and audit

Item S0 v1.0 S0 v1.2 three-arm
Formal units 20 60
Participant provider calls 420 1,260
Participant physical experiments 760 2,280
Physical experiments including baselines 28,060 baselines not counted again
Automatic retries 5
Method failures 0 0
Exact replay all passed all passed

The three-arm campaign reuses the v1.0 opaque results; those observations are not independent duplicates.

6. Evidence

7. Boundary with mechanism adaptation

The three-arm study asks how prior information changes search in a static world. Mechanism adaptation asks whether an Agent detects, attributes, and recovers from a law change during a campaign. They share the evidence-correction narrative but are not interchangeable experiments.

RC28 Gate A certified environment identifiability and online attainability only on its historical frozen source. Its current-source binding is stale, and Participant Gates B–E remain pending.

8. Five-task post-qualification extension

On source 74cfcdaa0d9780de2d21424ef8c329079554f8b5, five tasks used the same task-neutral Codex prompt, world seeds 0–4, 20 exploration experiments, and 3+3 blind validations. Five classical methods used the same public task information. This is audited development evidence, not a preregistered superiority study.

Task Codex Best classical method Difference Wins/ties/losses against each world's best
Electrochemical Conversion 0.7454 ± 0.0522 RF-EI 0.6622 +0.0832 3 / 0 / 2
Reaction to Crystallization 0.5206 ± 0.0681 RF-EI 0.6071 −0.0866 1 / 0 / 4
Reaction to Distillation 0.4795 ± 0.0264 GP-EI 0.4192 +0.0603 4 / 0 / 1
Partition Discovery 0.5426 ± 0.0870 GP-EI 0.5511 −0.0085 1 / 0 / 4
Flow Reaction Optimization 0.1627 ± 0.0131 GP-EI 0.2145 −0.0518 0 / 0 / 5

All 150 method-by-world cells and 3,900 physical experiments completed exact replay. The new 13-dimensional reaction–distillation task reached its task threshold in every world. No method reached Partition Discovery's frozen 0.58 cross-world threshold. Absolute scores are not comparable across tasks.