Consolidating Cognitive Substrate (CCS)

Programme-level umbrella for a research programme on multi-timescale continual agents. Theory, terminology, claim ledger, experiment dependency graph, and evidence status.

Status: no claim in this programme has empirical support.

Reconciled 2026年09月04日. The programme has its first admissible confirmatory result: plasticity-routing/EXP-001 passed every preregistered gate on five held-out seeds. It supports CCS-C4 partially — allocation content matters under matched budgets — and does not establish CCS-C4's separability statement, because the experiment contains no independently manipulable COMMIT_INTERNAL policy to hold fixed.

CCS-C4 stays unresolved, and was not narrowed to fit the result. No claim has empirical support. 0 of 4 synthesis-readiness tiers are cleared.

A component result is not evidence that the decomposition composes. This repository exists to keep that distinction visible.

A thin coordination layer. It maintains:

• the architecture hypothesis — explicitly a conjecture
• component definitions and terminology
• the claim ledger — every claim, its evidence status, its falsifier
• the resource envelope — the common accounting the integration experiment needs before it can be stated
• the evidence map — claim ↔ experiment, in both directions
• the experiment dependency DAG
• the paper roadmap, novelty boundaries, integration criteria, failure criteria
• synthesis readiness — what must be true before a synthesis paper is defensible

• Not an experiment implementation repository. No models, no training code, no runs. Experiments live in sibling repositories and stay there.
• Not a place where untested architecture is described as results. The decomposition below is a conjecture. tools/validate_ledger.py scans committed prose for result language as a tripwire.
• Not a controller of sibling protocols. Reconciliation is read-only by construction. See RECONCILIATION-POLICY.md.

ACCUMULATE ──▶ ALLOCATE ──▶ COMMIT_INTERNAL
 retain decide decide whether and when
 candidate where a change to durable
 experience capacity state is made
 is assigned
 ————— a separate commitment domain —————
 COMMIT_EXTERNAL
 decide whether and when cognition is
 released outward: WAIT / THINK / ASK
 / RESPOND / ACT

...with state maintained at multiple timescales: ephemeral, latent, fast parametric, slow parametric, over a frozen foundation.

This is an organising conjecture, not a finding. It is written down so it can be attacked. Four specific ways it may be wrong — including the live possibility that ALLOCATE is not separable from COMMIT_INTERNAL — are in ARCHITECTURE-HYPOTHESIS.md.

COMMIT was split on 2026年09月03日. One word had been covering two decisions — making state durable, and releasing an utterance — which presupposed that a single mechanism governs both. That is now an explicit claim with four falsifiers (CCS-C12), not a naming convention. Whether COMMIT_EXTERNAL belongs in the pipeline above is unresolved.

Each existing track declares its own results non-evidential and declares that they do not travel to other tracks. The umbrella honours those declarations: see the live reconciliation.

Eleven evidence entries are on file; two are admissible, both partial-scope — and the second points against its claim. Three of thirteen claims have no runnable test: CCS-C5, CCS-C8 and CCS-C12 — the last two because their experiments have no host repository, deliberately.

The ledger distinguishes full from partial evidence scope. A partial entry records what it establishes and what it leaves open, so an admissible component result can be logged without narrowing the claim to fit it. The validator refuses a full-scope supporting admissible entry on an unpromoted claim, and requires a held_back_reason whenever admissible evidence does not promote. Full table: docs/CLAIM-LEDGER.md.

Two claims are worth reading before the rest. CCS-C6 carries two against-direction development results and should be expected to end at falsified; its live successor is CCS-C11, added rather than substituted so the original is not quietly rewritten to follow the evidence.

CCS-C11 is now itself under pressure. A preregistered frozen-grid run found that pooling beats refusing-to-admit only above a threshold in absolute skill count, and that its advantage grows as the ceiling loosens — the opposite of the claim's motivating intuition, which pointed at high pressure. The claim has not been rewritten; the tension is recorded on it, and the successor mechanism was added as CCS-C13 (consolidation needs pressure and candidate diversity) rather than substituted.

The programme's endpoint might be titled something like Consolidating Cognitive Substrate: Toward Multi-Timescale Coherent Continual Agents.

Its Results section will not be written until the evidence exists, and this repository will not host speculative drafts of it. The gates are in SYNTHESIS-READINESS.md; none are cleared. What is honestly publishable at each stage short of synthesis is tabulated there too.

No CI. No GitHub Actions. Run locally:

make check

Python 3 standard library only, by design.

Machine-readable, under ledger/:

• claims.json — the claim ledger
• repos.json — sibling registry and observed state
• dag.json — experiment dependency graph
• resource_envelope.json — cross-track resource accounting

Markdown tables in docs/ are generated from these. Edit the JSON, then make render.

See CONTRIBUTING.md and WRITING-RULES.md. The short version: a claim's status may only rise when a committed artefact in a sibling repository supports it, and never because something was implemented, looks promising, or is needed for a deadline.

Apache-2.0, matching the sibling repositories.