Roadmap

This is an auditable engineering plan. It is not a promise of evidence about the universe. Every stage must produce inspectable artifacts, pass declared acceptance checks, and stop when its evidence requirements are not met.

Stage 0: Record and boundary

Status: Current

Purpose: Preserve the completed evidence and separate it from proposed engineering and optional cosmology.

Deliverables

  • Theory and claims ledger
  • Simulator specification
  • Public methods, deviations, and release manifests
  • OSF retrospective-registration package
  • Human and AI contribution policies

Acceptance

Every test has an explicit status, reproducible artifacts where available, and visible provenance limitations. Missing or conflicting evidence status blocks new claims.

Stage 1: Minimal simulator foundation

Purpose: Prove that the combined architecture can be implemented and audited before attempting a large world.

Stage 1A: Core world contracts

Define the authoritative trunk, event schema, world clock, identities, relationships, obligations, unresolved events, causal dependencies, bounded views, dump records, merge rules, and cost accounting.

Acceptance

  • Schemas are versioned and machine-readable.
  • Every authoritative mutation has provenance.
  • Protected entries cannot disappear without a typed loss record.
  • Conflicting writes are rejected or resolved by a frozen deterministic rule.

Stage 1B: Deterministic world MVP

Implement one trunk, two deterministic regions, a small population, a versioned protected ledger, bounded local views, dump residuals, deterministic merge and still operations, checkpoints, and cost counters.

Acceptance

  • Identical seeds and inputs produce identical state and observation digests.
  • Exact checkpoint replay succeeds over the declared horizon.
  • Tests cover stale writes, duplicate events, conflicts, identity loss, and protected-ledger violations.
  • The run publishes its schema, seed, command, fixtures, outputs, and manifest.

This stage establishes buildability only.

Stage 1C: Research council workspace

Create the internal workspace that coordinates human and AI contributors. The first version must operate on real project records rather than placeholder activity.

Workspace responsibilities

  • Display proposed questions, protocols, assignments, and agent disclosures.
  • Track isolated implementations, critic findings, validation results, and unresolved disagreements.
  • Connect claims to supporting and contradicting artifacts.
  • Show provenance manifests, checksums, deviations, and experiment costs.
  • Require human approval before freezing a protocol, opening reserved seeds, accepting a claim, merging authoritative simulator changes, classifying an independent replication, or publishing a record.

Acceptance

  • Agents can submit structured proposals and evidence without directly modifying accepted claims.
  • Human decisions are attributable and reversible before publication.
  • Failed and rejected contributions remain visible.
  • No agent can approve its own scientific conclusion.

Stage 1D: World simulator dashboard

Build a functional control and observation surface for the deterministic MVP.

Dashboard responsibilities

  • Display authoritative trunk state, world time, regions, and inhabitants.
  • Inspect identities, relationships, obligations, unresolved events, and causal dependencies in the protected ledger.
  • Show proposed, accepted, rejected, and conflicting events.
  • Inspect dumps, reconstruction assumptions, checkpoints, and replay results.
  • Report memory, CPU, storage, merge, and reconstruction costs.

Acceptance

  • Every displayed value comes from the running simulator or its immutable run artifacts.
  • Dashboard actions cannot bypass event validation or human approval gates.
  • A user can trace a visible world change back to its source event and provenance record.

Stage 2: Bounded payload benchmark

Compare trunk-only, split-without-dump, dump-only, and split/dump/merge architectures on frozen finite workloads. Predeclare payload fidelity, memory, CPU, storage, dump, merge, and reconstruction metrics.

Acceptance

  • All inputs, outputs, failures, costs, and exclusions are published.
  • The benchmark produces a complete comparison table.
  • A clean independent rerun is attempted.
  • No winner is generalized beyond the tested workloads.

Stage 3: AI research council automation

Allow specialist agents to propose protocols, build isolated implementations, generate tests, search for counterexamples, audit provenance, and prepare claim proposals.

Acceptance

  • Agent roles, tools, instructions, inputs, outputs, and model disclosures are recorded.
  • Independent implementations do not share implementation details during their isolated phase.
  • AI-assisted work is not labeled independent replication solely because it used another model, agent, or prompt.
  • Reserved execution and publication remain human-authorized.
  • Compute and tool budgets are enforced.

Stage 4: Narrative closure benchmark

Specify forward and return constraints, causal continuity, and the inhabitant-visible narrative criterion. Test finite graphs and histories with known positive and negative controls.

Acceptance

  • The benchmark distinguishes recurrence from mutual temporal constraint.
  • Seams, failures, ledger costs, and conflicting histories are reported.
  • A passing result supports only internal simulated narrative closure under the frozen model.

The closure constraint is currently untested.

Stage 5: Independent replication and registration

Have a separately accountable reviewer or team run the frozen benchmark in an independently controlled environment. Deposit prospective protocols and results through OSF Registration, while the repository and website remain the active collaboration locations.

A mismatch requires a deviation record, not a quiet correction.

Stage 6: Conditional research after reciprocal closure

Status: Proposed only. Do not begin unless a reserved reciprocal-closure study produces a valid result and a separately accountable team independently replicates it.

These studies test possible benefits rather than assuming them. A valid cycle would establish only the frozen closure predicates within the declared system boundary; it would not by itself establish better reasoning, lower cost, indefinite stability, or practical superiority over established methods. For LLM-informed worlds, long-horizon contradiction, memory drift, diversity loss, and synthetic-data degradation are separate outcomes. Reciprocal closure may be tested as a detector or constraint for some of these failures, but closure alone does not prevent model collapse, hallucination, or factual error.

Stage 6A: Guided-generation comparison

Test whether ending constraints can guide history generation rather than only reject completed histories. Compare forward generation, forward generation with final validation, planning or constraint solving, checkpoint replay, reciprocal closure without guidance, and reciprocal closure with goal-conditioned guidance.

Acceptance

  • Tasks, long-range constraints, resource budgets, and failure rules are frozen before execution.
  • Constraint satisfaction, rejected candidates, contradictions, interventions, protected-fact retention, unsupported additions, diversity, hidden-state dependence, runtime, and compute are reported.
  • Any advantage is limited to the evaluated tasks and matched baselines.
  • “Flawless consistency” and physical retrocausality are not inferred.

Stage 6B: Comparative efficiency benchmark

Measure whether reciprocal closure reduces search, replay, storage, or verification cost on any declared task class. Compare it against event sourcing, deterministic replay, checkpoint restoration, planning and search, constraint solvers, model-predictive control, and generate-and-rank systems where those are applicable.

Acceptance

  • Baselines receive matched inputs, outcome criteria, and resource budgets.
  • Successful, unfavorable, and invalid comparisons are all retained.
  • No general efficiency claim is made from a task-specific advantage.

Stage 6C: Repeated-cycle stability study

Test whether independently validated cycles can repeat without accumulating state corruption, unresolved obligations, hidden dependencies, or unbounded resource costs. In LLM-informed arms, distinguish context or memory drift from training-data model collapse; do not use one as evidence of the other.

Acceptance

  • The number of cycles, stopping rules, fault injections, and resource ceilings are declared in advance.
  • Closure success, ledger growth, numerical error, unresolved obligations, identity preservation, contradiction rate, protected-fact loss, diversity change, recovery behavior, and external-state dependence are measured separately.
  • Compare against retrieval, event-sourced memory, checkpoint, schema-validation, and forward-only controls under matched model and resource conditions.
  • Bounded stability is not described as infinite operation or perpetual motion.

Stage 6D: Bounded world-simulation pilot

Build a controlled world-simulation study containing a founding artifact or obligation, LLM-informed autonomous actors, a long-range ending condition, conventional validation baselines, and a reciprocal-closure treatment.

Acceptance

  • Contract satisfaction, narrative coherence, human intervention, runtime cost, agent autonomy, and human-rated experience are separate outcomes.
  • A closed history is not assumed to be natural, correct, enjoyable, or superior without the corresponding measurements.
  • Internally consistent generated claims remain distinct from measured inputs and may be false.
  • The pilot remains a bounded application study and does not expand the evidence into claims about consciousness, cosmology, or physical time.

Evidence guardrails

Tests 01 through 04 remain retrospective in the current OSF record. Test 03 failed its support criterion. Test 04 Revision 2 is limited to exact replay in its fixed model and tested horizon. Revision 1 is aborted.

No roadmap stage licenses claims about physical recurrence, consciousness, subjective continuity, cosmology, entropy removal, or universe-scale resource savings.