Adaptive Commitment

Working research project: when should an agent stop deliberating and commit outward — WAIT, THINK, ASK, RESPOND, ACT — under streaming evidence?

Status: charter only. No literature audit, no benchmark, no preregistration, no code, no result. This repository holds no claim, and the CCS umbrella deliberately records none against it.

The domain: COMMIT_EXTERNAL

The CCS programme separates two commitment decisions that a single word used to cover (terminology):

	decision	track
COMMIT_INTERNAL	whether information becomes durable in the substrate	in-c0/state-promotion
COMMIT_EXTERNAL	whether cognition is released outward as a response or action	this repository

They are not assumed to share a mechanism. Whether one learned policy can govern both is CCS-C12, a conjecture with four falsifiers and no implementation. This track is not evidence for it, and must not be built as though the answer were known.

This track is not a refinement of state-promotion's consolidation threshold. The umbrella recorded it that way until 2026-09-03 and the record was wrong; the correction is in that reconciliation.

The question, not yet a hypothesis

Under streaming evidence, when should an agent choose among WAIT, THINK, ASK, RESPOND and ACT — trading accuracy against latency, compute, intervention cost, uncertainty, and consequence/reversibility?

That is a research question, not a falsifiable claim. Turning it into one is the first job here, and it must happen before any benchmark is designed — otherwise the benchmark decides the hypothesis.

The hard part is the objective, not the policy

Every other CCS track measures resources denominated in the machine: parameters, writes, FLOPs, stored bytes, evidence reads. This track's two native costs are not:

• latency — borne by whoever is waiting, not by the system;
• intervention cost — the price of asking an unnecessary question, or of an action that cannot be undone.

The umbrella's resource envelope records both as INCOMMENSURABLE: they exist, and they have no non-arbitrary exchange rate with a parameter write. That is open problem OP-1, it currently blocks the programme's integration experiment, and it is this track's problem to define.

A trade-off objective that silently picks an exchange rate between a wrong answer and a slow one has not solved this; it has hidden it. Whatever this track chooses, it must be stated as a choice, with the result reported across a range of rates rather than at one flattering point.

Novelty boundary — provisional, pending the audit

The field is crowded. Not claimed as novel, pending confirmation by the literature audit: anytime/interruptible algorithms; speed–accuracy tradeoffs and drift-diffusion accounts of decision timing; the option to abstain, defer or escalate to a human; active learning and the value of asking; early-exit and adaptive-computation inference; bandits with a stopping decision; deferral under uncertainty in agent systems.

If the audit finds the question is already answered, that is the result, and this track should report it and close rather than build a benchmark to rediscover it.

Execution order — gated

1. Literature audit. What is already owned, and what survives.
2. Falsifiable claim. One sentence, with its falsifier, derived from the audit — then proposed to the CCS umbrella as a claim.
3. Cost model. Address OP-1 explicitly, or declare it unaddressed.
4. Benchmark, calibrated by a criterion naming no candidate policy.
5. Preregistration, committed before any runner exists.
6. Development phase, disjoint seeds.
7. Confirmatory, one shot.

Nothing below step 2 may be built first. A benchmark written before the claim will encode the claim, and the audit will then be scored against a question the benchmark already assumed.

Dependencies

Not blocked on state-promotion. COMMIT_EXTERNAL was scoped as a separate domain, so this track does not wait on the state pipeline reading out. It is currently the only CCS track that could produce an admissible result without it.

No dependency on plasticity-routing, modular-consolidation or lifetime-integrity in either direction. Their results are not evidence here and this track's are not evidence there.

Inherited admissibility rules

This track inherits the programme's admissibility rules by default. A departure must be declared here and recorded at the next reconciliation. The ones most likely to bite in this domain:

• No held-out signal in a decision path. A policy that decides when to respond may not consult the label it will be scored against.
• Decision-time compute counted separately. The cost of deciding whether to speak is part of the method, not overhead to be hidden. state-promotion's pilot measured its gate at 72.4% of total algorithmic compute; a deliberation policy can plausibly be worse.
• Never weaken the primary comparator. A fixed threshold rule, tuned on the same budget as the learned policy, is the comparator — not a courtesy baseline.
• Preregistered criteria must be satisfiable by some possible policy, checked before freezing.
• Short-horizon results do not license long-horizon claims — now confirmatory in lifetime-integrity, mean Spearman ρ 0.335 across horizons.
• "Better than doing nothing" is not "it works." A policy that responds slightly less badly than answering immediately has not demonstrated deliberation is useful.

Local gates

No GitHub Actions, by programme convention. Gates run locally:

make check

Licence

Apache-2.0.