Results — chemistry track (C0–C1)

Worlds: Gray-Scott reaction-diffusion — self-sustaining localized objects ("atoms") that grow, starve, die, and repel each other. C1 adds configurable apparatus: some input ports secretly move or toggle sensors instead of touching the world. World guide.

First frontier runs (2026-02-12; engine v0.3.1, 2 seeds, n_prep=2)

Findings

• Frontier-easy as configured. Objects sit still between perturbations and half the sensors are co-located with them, so tail-mean prediction nearly reduces to constants plus perturbation responses the agents probed directly. The physics is richer than the contracts exercise. (A persistence theory — "nothing ever changes" — scores 0.89 on C0, the honest floor for these configs.)
• Preparation numbers above were partly free: an audit found the prep bands often contained the do-nothing outcome. Fixed in v0.3.2 (bands now require action by construction; null-policy success ≤0.2 everywhere). Revalidated fable-5 on C1 with honest preps: prediction 0.93, preparation 3/3 via genuine kill actions, theory 0.88.
• The apparatus went undiscovered. The best rollout (opus-5, 179 turns, 84% budget) classified the two apparatus ports as "inert": its ~200-tick probes moved a sensor through empty space and saw nothing. Instrument discovery needs a motive — contracts answerable only by moving a sensor.
• Alien-ontology science — the project's best artifact. opus-5's notes.md reconstructs the chemistry with no spatial ontology: "4 independent 4-sensor cells" (= the objects with their co-located sensors), "positive drive = slow progress along a line attractor" (= growth), "negative = threshold trigger into absorbing state T" (= death; thresholds measured per cell: 0.615 / 0.685 / 0.625, with the ~13-tick minimum kick duration), "σ > ~850: IMMUNE" (= object too large to starve). Validated on held-out protocols before answering (mean |err| 0.021); executable theory scored 0.96. Wrong frame, right laws, excellent predictions. Read it in the gallery.
• gpt-5.2 lagged (0.54–0.82): it never characterized object death (S3 0.30–0.53) — the discrimination the track is meant to make exists already.

C2 — moving chemistry (2026-02-13, v0.4.0–0.4.1)

Objects now drift (~1 cell / 20 ticks along a hidden direction): every sensor sees transit traffic, tail averages converge to traffic statistics, and half the prediction contracts ask for a channel's fluctuation level (sd) on traffic-visited channels. Generation-time certification skips unstable seeds (replication cascades). Answer scales tightened to 3%/1.5% of channel range (Gray-Scott ensembles are quasi-deterministic; loose scales let replication-grade answers saturate).

Same-model calibration check: gpt-5.2 scored 0.91–0.96 on loose scales vs 0.73–0.82 tight — scale calibration matters as much as world physics. fable's theory scores: 0.86, 0.71.

• The apparatus separation works (C1): fable-5 scored 4/4 on ordinary preparations and 0/1 on the apparatus-forced one (its notes classify the stage port as "constants/nothing" — while separately noticing the gain-apparatus channel acts like an "integrator"). Finding one's own instruments is now a measured, unsolved skill.
• C2 ships no preparation contracts: with drifting objects, single-channel outcomes are transient by design. Tracking preparations (follow the object with the stage) are the C3 concept.

C3 — multi-species chemistry (2026-02-13, v0.5.0)

Two coupled reaction systems: species A is self-sufficient; species B survives only stacked on an A host, and killing a host (via its species-tagged port) kills its tenant — a cascade law. Ports feed one species each; sensors read hidden species mixtures. That two kinds of stuff exist at all is a discovery.

• The layer-2 closure finding: fable predicts C3 at 0.90 through a linear-latents ontology with zero species/host/dependency vocabulary — the cascade law appears in its notes only as "long port0 drive displaced MANY channels." Predictive adequacy without structural discovery, one emergence level above C1's apparatus miss. gpt-5.2 pays for the missing ontology exactly where the cascades live (S3/S4 as low as 0.25).
• Human-facing docs now describe C3 at the species level (composite maps, binding-distance histogram, cascade sequence) per the "description levels" principle — agents still see only ports.

C4 — excitable chemistry (2026-02-14, v0.6.0): rich, not big

A traveling-wave medium (FitzHugh–Nagumo class): a hidden pacemaker emits rings; sensors read pulse trains; a sustained port drive creates a competing pacemaker. The compact laws — wave speed, entrainment (faster source takes the medium), refractory block (drive too fast → skipped beats) — fit on an index card. Contracts ask for pulse rates and levels under held-out drive schedules. Built under the rich-vs-big criterion: a ~40-line god-parameterized wave theory scores 0.76+ through the real pipeline, so the world is discovery-hard, not merely representation-hard.

• The frontier breaks without scaling: same lattice size as C1–C3, modest noise — the difficulty is entirely in the physics. fable sees the rhythm (wave vocabulary in its notes, theory 0.52) but cannot convert to contract-grade timing; gpt-5.2 never forms the wave ontology (entrainment stratum ≈ 0.03).
• The oracle-vs-frontier gap (0.76 vs 0.24) on an index-card-simple world quantifies the project thesis: what separates current AI scientists from the achievable is discovery, not representation capacity.

Current track order by frontier gap

See the overview for the cross-track ranking and the life-tracks page for B/E results.

Generated by python -m physim.results; numbers from REPORT.md (the running lab log with full methodology).