Not a forecaster. It learns a distribution over every history that could connect an initial state to a terminal constraint. The science is in the geometry of that distribution — not in any single path.
Generate samples the path measure — from one state, the space of possible histories erupts at once, and the streams that converge are mechanism families.
A path that breaks a conservation law is rejected. A goal with no sampled path is not impossible — only when competence, coverage, and an independent invariant or numerical oracle all agree.
Each survivor becomes a typed, scoped, falsifiable claim in an append-only, auditable ledger — with evidence, competence, and coverage attached. Belief is updated by rules, never by a learned policy.
Discovery nodes (blue) feed the green claim-certification path — the Claim Compiler types samples, the Competence/Coverage model scores them, and nothing reaches the central Claim Ledger until the Certification Gate admits it. The skepticism engine (teal) attaches calibrated belief; the write-up stage (violet) renders only what the ledger supports. Amber dashed = the planner's next experiment fed back to discovery. Click any node.
Click any node above for its agents, its role, and how it learns. Start with the amber hub — the Claim Ledger — which everything reads from and writes to.
It needs three coupled models, not one monolith. Two are learned by gradient descent; the third — belief — is updated only by auditable rules. That is why it is easy to miscount as two.
Each module mapped to the inner-ring process it serves and the world model it belongs to. M1 is the physical world model; M3 realizes the action–outcome model; the ledger (W_epi) is not a trained module — it is the object these write into.
| Module | Symbol | Process / role | World model |
|---|---|---|---|
| M1 Path generator | ℙθ | Generate | W_phys |
| M2 Forward model | f_φ | Assimilate / Generate critic | feeds W_phys |
| M3 Latent forward model | h_ψ | Explore (EIG rollouts) | W_act |
| M4 Assimilation model | p_η | Assimilate | — |
| M5 Mechanism encoder (LCS) | e_ξ | Represent | — |
| M6 Claim compiler | c_χ | Represent → ledger | outer ring |
| M7 Competence model | κ_α | Quantify / certify | self-knowledge of W_phys |
| M8 Coverage model | γ_β | Assimilate / Quantify | — |
| M9 Literature retriever + NLI | r_ζ | review (outer) | outer ring |
| M10 Context policy | π_ctx | all agents (read) | outer ring |
| M11 Planner policy | π_ν | Explore | uses W_act |
| M12 Refiner policy | π_edit | Refine | edits W_phys |
Review is a dynamic panel whose domains change with each claim — grounded in its own retrieved evidence, not personas over one context. A mandatory falsifier tries to break it; a judge adjudicates against the ledger and log, weighting reviewers by their calibration track record. Disagreement is an output, not a forced verdict.
Each agent is stateless except through four shared memory layers — so an agent's design is its read/write contract. Any agent may read any layer; writes are restricted so no learned policy can author belief or rewrite history.
| Agent | Reads | Writes (how) |
|---|---|---|
| Planner | ledger, episodic, log | pre-registration → log |
| Generator | probe, W_phys | ensembles → log |
| Assimilator | ensembles, observations | weights / CRPS → log |
| Representation | ensembles | LCS fields → log |
| Claim compiler | ensembles, LCS, κ, γ | candidate claims (gated) |
| Falsifier | ledger, log | counterexample probes |
| Retriever | ledger, literature graph | evidence packets (not the log) |
| Reviewer / Judge | ledger, log, packets | opinion update (rule) |
| Refiner | log, episodic, metrics | model edit (gated by V) |
The leap from generator output to belief is where the make-or-break failure lives. Three nodes turn it into a gated pipeline; no claim is admitted without a typed object, scope, falsifier, log reference, and competence/coverage annotation.
Memory is evidence; context is policy. A learned agent decides what to retrieve into context — never what the log says or whether evidence existed.
Writing is not a free-standing rhetorical stage — it is a derived view bound to the ledger, so the output can never exceed the evidence. The three artifacts are kept strictly separate.
Manuscript synthesis (N8) is five components, each bound to the ledger:
| Component | Job |
|---|---|
| Pre-registration / outline agent | writes the prediction & falsification criterion to the log before the experiment |
| Section synthesizer | renders a ledger subgraph; numbers locked to log entries; hedged vs. assertive language from (b,d,u); explicit "what is unknown" |
| Figure service | plots real ledger/log data with a verification loop checking each figure against its log entry |
| Citation binder | verifies a source supports (or contradicts) the claim — warrant, not mere existence |
| Refinement controller (N9) | accepts an edit only if it improves calibrated fit against the log; reverts otherwise |
The curriculum runs from physical competence to epistemic action — forward & bridge models, representation, competence/coverage, claim compiler, literature + committee, then planner/memory/refiner policies, and only last the closed loop. Evaluation is a graded ladder.
Headline metric: the certified-impossible false-positive rate. The decisive first experiment is whether the system can separate reachable · rare · impossible · out-of-competence.