Platform & handoff

Experiments — validation and assay planning.

The Experiments surface where validation runs and assay plans are scheduled, tracked, and tied back to claims.

Experimental · computation and biology

DiscoveryLab keeps speculative science executable, inspectable, and bounded.

The experimental track holds ideas outside the primary product path until they survive concrete artifacts: run configs, generated mechanisms, verifier receipts, tests, and explicit claim boundaries. RuliadBench tests observer-relative computation; the epigenome track tests whether chromatin-fidelity peptide hypotheses can be configured without pretending they are validated biology.

Experiment design

Predict what an observer can measure, not the whole microscopic future.

The tensor is Y[rule, init, time, observer, observable]. The model learns embeddings over computations and predicts coarse readouts. This keeps the target aligned with computational irreducibility: exact microscopic evolution may require running the computation, while observer-level summaries can still be compressible.

Universe

All 256 elementary cellular automata

State

12 initial conditions · 64 cells · 16 horizons

Observers

density · entropy · block histograms · compression · sensitivity

Model

feature-backed CP residual over a per-observer mean baseline

Artifact

experiments/ruliadbench + TensorLang static smoke module

Wolfram Science connection. The ruliad frames physics and computation as an enormous space of possible rules and histories. RuliadBench is a small, executable slice of that idea: sample a known computational universe, choose observers, and measure which observer-relative facts can be learned without pretending to shortcut all microscopic computation.

Strict 256-rule result · MSE lower is better

Split

ID CP

Feature CP

Linear baseline

Held-out rules

Can the model predict unseen ECA rules?

0.055552

0.004619

Feature-backed rule descriptions transfer; ID embeddings do not.

0.023061

Held-out initial conditions

Can it generalize to unseen starting states?

0.083956

0.004347

Initial-state summaries carry real predictive structure.

0.019302

Held-out long horizons

Can it forecast beyond trained time steps?

0.085229

0.023497

Coarse behavior extrapolates better than ID lookup, but this remains the hard split.

0.024925

Held-out observer

Can it infer a new measurement operator?

0.204300

0.074883

Observer metadata helps, but real observer transfer needs richer operator semantics.

0.114442

The first strong result is not interpolation.

A learned ID embedding gets very low random-row error, but that is mostly memorized geometry. The feature-backed residual model is the meaningful result because it has rule bits, initial-state summaries, time features, and observer metadata for truly excluded axes.

Observer-relative compression is measurable.

Exact microstate prediction at held-out horizons still has roughly 22-32% bit error in this run, while coarse observers remain substantially more predictable. That is the scientific point: the model is not replacing computation, it is learning what an observer can compress.

The bridge to biology is structural, not metaphorical.

Replace rule, initial condition, time, and observer with cell context, perturbation, dose/time, and readout. The same tensorized intervention question becomes whether unseen biological state transitions have learnable, observer-relative structure.

Current run: build/ruliadbench/eca_256_feature_strict/model/strict/SUMMARY.md. The section reports experiment evidence, not product claims or wet-lab evidence.

Epigenome rescue track

Chromatin-fidelity hypotheses now have a concrete run configuration and a verifier failure surface.

The latest configuration turns the Information Theory of Aging discussion into a bounded DiscoveryLab experiment. It focuses on peptides for chromatin-interface reader/writer/eraser targets, DNA-repair-coupling, and epigenome-output assays, with forbidden states treated as first-class verification targets.

Run ID

EPI-CHROMATIN-DISCOVERY-V1

Swift plan

EpigenomeDiscoveryRunFactory + EpigenomeRunPlan

Targets

BRD4_BD1 · KDM1A_LSD1 · EZH2_SET

Lanes

sequence binder · seed-local optimization · macrocycle · controls

TensorLang

biology_epi_v1.tl + PeptiterMechanismReasonerEpiV1

Lean

PathwayLean.Examples.BiologyEpiV1

Current conclusion

Epigenomic aging is handled as a falsifiable target class.

SINCLAIR.md frames the information-theory-of-aging thread as a useful hypothesis, not a product claim. The configured objective is epigenomic fidelity: restore cell identity and chromatin order while penalizing dedifferentiation, DNA damage, oncogenic proliferation, and broad chromatin disruption.

The run keeps controls in the denominator.

The Swift plan includes parent, negative, and near-miss calibration controls alongside discovery candidates, so the batch cannot collapse into top-hit-only reporting. The local in-silico submission completes across the assay panel and records the denominator policy.

The verifier is allowed to reject the hypothesis.

The SIRT6 / heterochromatin TensorLang slice deliberately exposes risky repair-chromatin couplings. The manifest is not fully green; failed efficacy and safety obligations are the point because they identify where the encoded mechanism overreaches.

Wolfram Science connection. The same observer-relative discipline from RuliadBench applies here: the system does not claim to solve the full biological evolution. It chooses a finite mechanism slice, names the observer readouts, and lets TensorLang plus PathwayLean show exactly which claims pass, fail, or require wet-lab evidence.

Current artifacts: experiments/epigenome-discovery, Packages/PathwayLean/tensorlang/biology_epi_v1.tl, and Packages/PeptiterDiscovery/Sources/PeptiterDiscovery/EpigenomeDiscoveryRun.swift.

Related on this track

AI Scientist

Research loop and next-action reasoning.