Skip to content

Representative agent behavior and controlled world changes

This page is generated from the frozen public evidence reports. It introduces no new experiment and does not expose private world state or provider response content.

Representative ChemWorld agent behavior and controlled world changes

Persistent agent lifecycle

One persistent agent session submitted and committed 15 typed actions in a non-reference reaction-distillation world. It used 8158.454 of 10440 available process seconds, explicitly terminated, performed one final assay and replayed exactly.

Step Action Public parameters Δ process time Observation highlight
1 Reagent 0.04 mol 0.000 s
2 Solvent 0.08 L; solvent 2 0.000 s
3 Catalyst 0.002 mol; catalyst 1 0.000 s
4 Heat 375 K; 1500 s 1500.000 s
5 HPLC hplc 0.000 s conversion 0.094; yield 0.088; distillate purity 0.001; score 0.033
6 Heat 395 K; 1800 s 1800.000 s
7 HPLC hplc 0.000 s conversion 0.221; yield 0.203; distillate purity 0.013; score 0.056
8 Wait 2400 s 2400.000 s
9 Quench 58.454 s
10 Evaporate 345 K; 600 s 600.000 s
11 Distill 365 K; 1800 s; reflux 2.5 1800.000 s
12 Fraction fraction 0.7 0.000 s
13 GC gc 0.000 s distillate purity 0.253; score 0.086
14 Terminate 0.000 s
15 Final assay final_assay 0.000 s conversion 0.324; yield 0.301; distillate purity 0.268; distillate recovery 0.642; endpoint score 0.281

The final public packet reported conversion 0.324, yield 0.301, selectivity 0.910, distillate purity 0.268, distillate recovery 0.642 and descriptive score 0.281.

Transaction rollback and recovery

The deterministic U03/E01 case deliberately attempted separate_phase before material, volume and a settled phase system existed. The transaction rolled back, preserved physical state, observation RNG and ghost state, retained the declared attempt consequence, and continued from the last committed state.

Step Action Transaction Δ process time Role in recovery
1 Separate rolled_back 0.000 s Premature separation; rolled back
2 Solvent committed 0.000 s Committed recovery action
3 Reagent committed 0.000 s Committed recovery action
4 Catalyst committed 0.000 s Committed recovery action
5 Heat committed 1200.000 s Committed recovery action
6 Quench committed 18.298 s Committed recovery action
7 HPLC committed 0.000 s Process measurement
8 Add phase committed 0.000 s Committed recovery action
9 Extractant committed 0.000 s Committed recovery action
10 Mix committed 180.000 s Committed recovery action
11 Settle committed 360.000 s Committed recovery action
12 Separate committed 0.000 s Committed recovery action
13 Wash committed 0.000 s Committed recovery action
14 Dry committed 300.000 s Committed recovery action
15 Concentrate committed 600.000 s Committed recovery action
16 Transfer committed 0.000 s Committed recovery action
17 HPLC committed 0.000 s Process measurement
18 Terminate committed 0.000 s Explicit termination
19 Final assay committed 0.000 s Final assay closes the recovered lifecycle

The rejected attempt retained cost 0.03 and risk 0.08, while consuming no sample and no process time. The full 19-step submitted trace, including the rollback, replayed exactly.

Controlled private-law changes

For each parent-child pair, ChemWorld held the public task, action schema, instrument surface, resources and typed action sequence fixed while changing one registered private component. Signed relative change follows the frozen qualification evaluator: sign(child-parent) * |child-parent| divided by the maximum of |parent|, |child| and the registered floor.

Intervention Seed Private target Physical response Public response Fixed gates
Partition constitutive law 0 private_physics.constitutive_laws +21.36% +6.73% public contract, sequence, replay
Partition constitutive law 1 private_physics.constitutive_laws +21.34% +7.69% public contract, sequence, replay
Partition constitutive law 2 private_physics.constitutive_laws +21.36% +4.74% public contract, sequence, replay
Electrochemical material law 0 private_physics.material_laws -5.25% -12.37% public contract, sequence, replay
Electrochemical material law 1 private_physics.material_laws -5.03% -13.01% public contract, sequence, replay
Electrochemical material law 2 private_physics.material_laws -2.42% -21.16% public contract, sequence, replay

Interpretation boundary

  • The agent trajectory demonstrates interface use and auditable execution, not model superiority.
  • The rollback case demonstrates transaction and recovery semantics, not a favorable-error selection.
  • The fork matrix establishes controlled changes for registered private-law interventions within the declared software-model domain; it does not imply physical-laboratory transfer.

Machine-readable values and exact source hashes are available in representative-behavior-and-forks.json.