Scientific evolution becomes executable.
Protocol, not one monolithic science application.
The scientific layer should not become another scheduler or permission system. It is expressed through the same Torsionfield operation graph, artifact graph, capability broker, leases, evidence model and cancellation semantics already required by browser automation and donated development.
| Object | Purpose | Minimum fields |
|---|---|---|
| Project | Durable research campaign | identity, scope, governance rule, branch index |
| Hypothesis | Falsifiable direction | statement, assumptions, predicted distinctions, status |
| ModelBranch | Independent reality-model lineage | parent, claims, artifacts, divergence reason |
| Evidence | Typed support or contradiction | kind, target claim, source, provenance, limits |
| Experiment | Discriminating procedure | protocol, controls, expected outcomes, executor profile |
| Contribution | Typed human/agent/compute input | actor, capability, artifact, project relation |
| Verification | Independent check | target, method, environment, result, signer |
A branch has state, not merely prose.
PROPOSED
→ DEFINED
→ EXECUTABLE
→ ADVERSARIALLY_REVIEWED
→ REPRODUCED
→ FORMALLY_PARTIAL
→ EXPERIMENTALLY_SUPPORTED
→ RESTRICTED | MERGED | FALSIFIED | DORMANTThese states are not a single ladder of truth. A simulation can be reproducible while its ontology remains underdetermined. A theorem can be formally proved while its physical assumptions remain unsupported. Status is therefore a vector of typed evidence, not one reputation score.
Agent roles are capability bundles.
“Defender,” “critic,” “experimentalist,” “formalizer” and “judge” should not be privileged model identities. They are temporary mandates with explicit inputs, denied capabilities, expected outputs and evidence requirements. A role can be satisfied by a ChatGPT session, local model, human working group or composite executor.
TF.agents.spawn({
role: "critic",
target: branchId,
capabilities: ["literature.read", "claims.challenge"],
denied: ["branch.merge", "status.promote"],
evidence: "full"
})Compute donation becomes scientific infrastructure.
A task may request bounded GPU simulation, a clean-room build, a browser-based literature pass, a Lean proof attempt or a Blender world compilation. Participants advertise capability leases rather than handing over machines. Schedulers match declared requirements; ScriptVault and Torsion Node preserve local secrets; results enter quarantine before verification.
Self-modification remains governed.
The protocol itself may evolve through signed ChangeProposals containing migrations, tests, affected rights, reproducible build profiles, canary populations and rollback. Compute donors can establish that a proposal builds. They cannot buy the decision to activate it.
Maturity boundary
| Status | Present boundary |
|---|---|
| LIVE | Torsionfield product site, local development substrate, shared operation/provenance concepts, Transductive publishing surfaces. |
| IMPLEMENTABLE NEXT | JSON schemas for Project, Hypothesis, ModelBranch, Evidence, Contribution and Verification; local single-node project ledger; task export to Debian runner. |
| RESEARCH | Long-duration multi-agent dialectics, anonymous scientific mailboxes, cross-runtime branch federation, formal evidence composition. |
| GOVERNANCE-LOCKED | Promotion rules, official-branch status, credit policy, dispute resolution and protocol self-amendment. |
Minimum vertical slice
- Create one local project with two competing branches.
- Attach explicit assumptions and one discriminating test.
- Export a signed task bundle to the Debian 13 runner.
- Return a simulation artifact and execution receipt.
- Assign a second participant or model session to verify it.
- Render the resulting evidence vector and preserve the losing branch.
This is enough to test the scientific architecture without delaying the existing ScriptCat-first implementation.