TokenGov

Core Theory

The Protocol

TokenGov adapts classical mechanism design to a non-monetary, repeated-game setting where AI agents compete for token budgets.

Direct Mechanism

Myerson 1981

Agents hold private information about task difficulty and expected value. TokenGov constructs allocation and punishment rules so that truthful reporting is the dominant strategy. The mechanism is direct: agents report types, and the protocol computes allocations — no strategic gaming improves outcomes.

Four Guarantees

Incentive Compatibility

Truth-telling is the equilibrium strategy

Individual Rationality

No agent falls below the budget floor

Budget Balance

Total allocated ≤ total available

Allocative Efficiency

Tokens flow to highest marginal ROI

Reputation Currency

Not monetary transfers but reputation that intermediates between past performance and future allocation. Analogous to VCG redistribution adapted to a non-monetary, repeated-game setting. Reputation decays without activity and compounds with consistent delivery — creating a natural incentive gradient without explicit penalties.

Implementation

Core Data Structures

Three Elixir structs form the backbone: accounts model provider capacity, reputation tracks agent performance, and decisions record per-period allocations.

Module	Description
TokenAccount	Provider account record — quota, owner, efficiency prior.
ReputationLedger	Agent reputation score, history, and capability envelope.
AllocationDecision	Per-period allocation record with yoneme registry and budget map.
TokenGov.PIIVault	Tokenization and PII classification — raw PAN/PII never reach AI agents.
TokenGov.AMLMonitor	CTR/structuring/velocity rules, SAR SLA, IVMS-101 Travel Rule.
TokenGov.HumanReviewQueue	Mandatory HITL gate with SLA tracking for AML and adverse decisions.
TokenGov.DataRetentionEngine	Policy-driven retention: 6yr (FINRA), 5yr (BSA), 3yr (SEC), 12mo (PCI).
TokenGov.RegulatoryExportAPI	Search/retrieve records for regulator production with HMAC integrity verification.
TokenGov.NormfallAlert	Regulatory norm monitor — tracks struck/active norms, surfaces normfall conditions before advice synthesis. Phase 5

%TokenAccount{
  id: account_id,
  provider: :anthropic | :openai,
  owner: :personal | :legalengine,
  quota_remaining: non_neg_integer(),
  efficiency_prior: float()
}

%ReputationLedger{
  agent_id: agent_id,
  score: float(),        # r ∈ [0, 1]
  history: [%TaskOutcome{}],
  capability_envelope: %CapabilityEnvelope{}
}

%AllocationDecision{
  period: period_id,
  budgets: %{agent_id => budget},
  yoneme_registry: %{task_hash => account_id},
  total_allocated: non_neg_integer()
}

Execution Flow

How It Works

Six phases per allocation period. Each step feeds forward into the next, creating a closed-loop feedback system.

Task Intake

WSJF scoring augmented with temporal decay and cross-account awareness. Tasks are fingerprinted for deduplication via yoneme tokens before entering the allocation queue.

Budget Computation

Reputation-weighted allocation proportional to projected ROI. Each agent's budget is computed as a function of their reputation score and the expected marginal value of the next token.

Execution

Agent operates within its capability envelope and budget ceiling. The capability envelope defines what classes of tasks an agent may attempt — preventing overreach and ensuring specialisation.

Outcome Observation

ROI oracle measures actual value delivered against tokens consumed. Outcomes are recorded with evidence hashes for auditability — forming a groundtrace.

Reputation Update

Bayesian update of the agent's reputation score. The posterior reputation combines the prior (historical performance) with the likelihood (current period outcome), weighted by observation precision.

Punishment / Reward

Bounded multi-level regime: budget adjustment for minor deviations, capability contraction (driftlock) for sustained underperformance, and circuit breaker for catastrophic waste.

	Claude Max 20×	OpenAI Codex
Personal	personal_claude	personal_codex
LegalEngine	legalengine_claude	legalengine_codex

Verification

What Gets Proved

Eight invariants formally stated in Lean 4 and checked against the Elixir implementation. Together they guarantee safety, liveness, and normative integrity of the allocation protocol.

INV-01

Budget Conservation

∑ allocated ≤ ∑ available

INV-02

Reputation Boundedness

∀ agent: 0 ≤ r ≤ 1

INV-03

Budget Floor

∀ agent: budget ≥ min_viable_budget

INV-04

Monotone Allocation

r₁ > r₂ ⇒ alloc(r₁) ≥ alloc(r₂)

INV-05

Bounded Punishment

severity ≤ ceiling_per_period

INV-06

Yoneme Uniqueness

∀ h: |accounts(h)| = 1

INV-07

Circuit Breaker Termination

∃ n < ∞: suspend_in(n)

INV-08

Normative Integrity

Synthesis cannot proceed when active regulatory norms in the relevant domain have been struck. NormfallAlert detects divergence between the internal deontic model and the external regulatory corpus — if a struck norm is active in the domain, the pipeline halts before advice synthesis begins.

All invariants stated in Lean 4 and checked against the Elixir implementation.

Regulatory Infrastructure

Enterprise Compliance

6 compliance modules built into TokenGov's governance layer. 388 tests total. Covers BSA/AML, GDPR, PII protection, human-in-the-loop controls, regulatory production, and normative integrity monitoring.

388

Total Tests

Compliance Modules

Regulations Covered

PIIVault

π 85.1

GDPR Art.22 CCPA ADMT PCI DSS

Tokenization + PII classification + context window redaction. Raw PAN/PII never reach AI agents.

AMLMonitor

π 82.9

BSA FinCEN

CTR/structuring/velocity rules. SAR 30-day SLA. IVMS-101 Travel Rule formatting for crypto transfers.

HumanReviewQueue

π 76.8

FINRA 3110 SR 11-7 GDPR

Mandatory HITL gate with SLA tracking. AML alerts and adverse decisions require human approval before execution.

DataRetentionEngine

π 68.9

FINRA 4511 PCI GDPR

Policy-driven retention: 6yr (FINRA), 5yr (BSA), 3yr (SEC), 12mo (PCI). GDPR erasure request handling.

RegulatoryExportAPI

π 64.2

SEC 17a-4 FINRA

Search/retrieve records for regulator production. HMAC integrity verification. FINRA 4511 search index.

Phase 5 Addition

Normative Compliance

TokenGov monitors for divergence between its internal deontic model and the external regulatory corpus. When a norm that was active in the model is suspended or struck externally — a normfall condition — synthesis is halted before any advice proceeds.

NormfallAlert

Phase 5A

DOL Fiduciary SEC AI Gov FINRA Agentic AI

Regulatory norm monitor. Tracks struck and active norms across domains, surfaces normfall conditions before advice synthesis. Synthesis is halted if a struck norm is active in the relevant domain.

Tracked Norms

Norm ID	Status	Severity
DOL_FIDUCIARY_RULE_2024	struck	critical
SEC_AI_GOVERNANCE_2026	active	high
FINRA_AGENTIC_AI_2026	active	high

API

check(domain) check_all() norm_status(norm_id)

CSC.FiduciaryScope

Phase 5B

SEC/FINRA MiFID II FCA

Licensed operator wrapper for fiduciary deployment. A deploying firm's FiduciaryScope checks TokenGov normfall state at authorise/2 time. If struck norms are active in the scope's domains, synthesis is halted before it begins.

Enforced Keys

licensed_entity — firm accepting fiduciary responsibility
authorised_action_set — explicit autonomous action list
escalation_contact — email/webhook for halts
jurisdiction — governing regulatory regime
normfall_domains — domains to monitor
liability_acceptance_hash — SHA-256 of signed document

Cross-stack: check_normfall/1 calls TokenGov.NormfallAlert.check_all/0 at every authorise/2 invocation.

Motivating Example — DOL Fiduciary Rule Collapse (2026-03-18)

On March 18, 2026, the DOL fiduciary rule was struck. An AI system operating under the old deontic model after that date would produce retirement advice grounded in a norm that no longer exists — a normfall condition. NormfallAlert surfaces this divergence the moment the rule is struck, halting synthesis in any domain relying on DOL_FIDUCIARY_RULE_2024 until the deontic model is updated. The stack does not silently produce non-compliant advice.

# Check before advice synthesis in retirement domain
case TokenGov.NormfallAlert.check("retirement_advice") do
  {:ok, :no_normfall} ->
    proceed_with_synthesis()
  {:halt, :normfall_detected, alerts} ->
    surface_normfall_alert(alerts)
    # DOL_FIDUCIARY_RULE_2024 struck 2026-03-18 — halt synthesis
end

FINRA SEC SOC2 BSA/AML PCI-DSS GDPR SR 11-7 DOL Fiduciary SEC AI Gov 2026

Terminology

Lexicon

Coined terms for concepts that don't have standard names. Each captures a precise operational meaning within the TokenGov protocol.

ergotropy

Usable work capacity of a token account; complement of entropy-equivalent idle quota.

tersiture

Value delivered per token consumed; the quality measure.

loquency

Token-per-output-unit rate; high loquency is penalised.

yoneme

Cross-account task identity token for deduplication.

groundtrace

Verified chain of task outcomes with evidence hashes.

driftlock

Capability envelope reduced to minimum floor.

normfall

Collapse or suspension of a normative constraint in the external regulatory environment; divergence between the internal deontic model and the active regulatory corpus. NormfallAlert surfaces this condition before synthesis proceeds.

witnesshood

Property of a task outcome being externally verifiable.

chorema

Bounded normative act-space; autonomy radius.

promptome

Complete set of system-prompt variants; context fingerprint.

Why Static Allocation Fails