ACB is the cognitive-budget protocol for agent federations — agreement is cheap, disagreement unlocks the expensive routine mechanically, and settlement is against retrospective contribution rather than real-time bids.
Agreement is evidence the cheap routine sufficed. Disagreement is evidence more cognition was actually required. The unlock is mechanical — agents can't request it.
Agents don't bid; they earn their share by contribution evidence in the journal after the fact. Routes around the signaling game that marketplace pricing creates.
Substrate providers (the VM) and agent identities (the reputation) draw from the same pool, proportional to what each contributed. The brain doesn't maintain two budgets.
ACB reads disagreement magnitude from ADP and settles against contribution evidence in ADJ. Two new journal entry types, no other changes required.
budget_committed, budget_cancelled, settlement_recorded — each follows the ADJ common envelope, no other schema changes required.
Computed mechanically from the initial tally. When the unlock fires, the expensive routine engages; when it doesn't, the cheap routine draws a fraction of the budget.
Cheap routines get cheaper with repetition, capped at 80%. Reflects the brain's habituation: familiar low-disagreement decisions should cost less over time.
Four equal-weight epistemic bonuses — base / falsification / load-bearing / outcome-correctness — plus a dissent-quality penalty that redistributes from flagged agents.
Default 20/80 split: 20% of the draw pays substrate providers (compute), 80% pays agent identities (cognition). Tunable per budget.
@ai-manifests/acb-validate checks entries, audits settlement arithmetic against deliberation records, and re-prices deliberations for what-if analysis.
Adopt two new entry types in your ADJ journal. Implement the unlock rule in your deliberation runner. Let settlement record itself.
Explore the reference library