acb-manifest

A TypeScript reference implementation of the Agent Cognitive Budget (ACB) protocol — the metabolic-budget layer for deliberative multi-agent systems. ACB provides append-only journal entries, pricing models, habit-memory discounts, and settlement distribution that mirror the brain's resource allocation for routine vs. contested decisions.

This library is one of several reference implementations (C#, Python) of the same spec. The spec itself is at adp-manifest.dev and is the source of truth; this library implements what the spec says.

Zero runtime dependencies. Pure TypeScript, ESM.

Install

npm install @ai-manifests/acb-manifest

Quick example

import {
  InMemoryBudgetStore,
  computeDisagreementMagnitude,
  selectRoutine,
  computeExpensiveDraw,
  buildSettlementRecord,
  Routine,
  TerminationState,
  type BudgetCommitted,
  type Tally,
} from '@ai-manifests/acb-manifest';

const initialTally: Tally = { approveWeight: 0.71, rejectWeight: 0.64, abstainWeight: 0.18 };
const magnitude = computeDisagreementMagnitude(initialTally);  // ≈ 0.948 (contested)

const routine = selectRoutine(pricingProfile, initialTally, /*roundCount*/ 1, TerminationState.Converged);
// routine === Routine.Expensive (unlock threshold exceeded)

const draw = computeExpensiveDraw(pricingProfile, /*participants*/ 3, /*rounds*/ 1, /*habitDiscount*/ 0.80);
// draw = 200 × 3 × 1.5^1 × (1 − 0.80) = 180 EU

API

All exports are re-exported from the package root.

Entry types

AcbEntry, BudgetCommitted, BudgetCancelled, SettlementRecorded

Value types

Denomination, PricingProfile, SettlementProfile, SettlementMode, BudgetConstraints, SubstrateDistribution, EpistemicDistribution, ContributionBreakdown, Tally, HistoricalDeliberation, ParticipantContribution, SubstrateReport, SettlementInputs, BudgetState

Enums

Routine, TerminationState, SettlementMode

Pricing

• computeDisagreementMagnitude(tally) — 1 − |approve − reject| / (approve + reject), in [0, 1]
• selectRoutine(pricing, tally, roundCount, termination) — returns Routine.Cheap when the decision is an agreed-on routine; Routine.Expensive when contested
• computeCheapDraw(pricing, participantCount, habitDiscount?) — cheap-routine draw
• computeExpensiveDraw(pricing, participantCount, roundCount, habitDiscount?) — expensive-routine draw with round multiplier
• computeDraw(pricing, tally, participantCount, roundCount, termination, habitDiscount?) — convenience wrapper that picks a routine and computes the draw
• computeHabitDiscount(history) — habit-memory discount function, capped at MAX_HABIT_DISCOUNT (0.80)
• MAX_HABIT_DISCOUNT — exported constant

Settlement

• distributeSubstrate(pool, reports) — substrate pool distribution proportional to reported cycles
• distributeEpistemic(pool, contributions) — default-v0 epistemic scoring with the four equal-weight bonuses (base, falsification, load-bearing, outcome correctness) plus dissent-quality penalty
• buildSettlementRecord(inputs) — builds a complete SettlementRecorded entry from contributions and substrate reports

Store

• InMemoryBudgetStore — thread-safe in-memory budget store suitable for tests and prototypes
• BudgetStore — interface for custom backends

Testing

npm test

Spec

This library implements the Agent Cognitive Budget protocol specification. Read the spec at adp-manifest.dev. If the spec and this library disagree, the spec is correct and this is a bug.

License

Apache-2.0 — see LICENSE for the full license text and NOTICE for attribution.