DIGEST
Traditional software systems execute actions through deterministic code paths. LLM-driven systems propose actions through probabilistic inference. EAAP defines a strict deterministic control plane between AI proposal and real-world execution — ensuring no action reaches production without verified admissibility.
Architecture Overview
Exogram implements four strictly separated control layers.
1. Motivation
When AI systems transition from generating text to executing real-world actions, the risk surface expands:
“Inference is probabilistic. Execution must be deterministic.”
EAAP prevents unauthorized AI actions, prevents execution under unresolved conflict, enforces role/policy/confidence thresholds, guarantees cryptographic state integrity, and produces a tamper-detectable audit record.
EAAP does not improve model intelligence. It governs model admissibility.
2. Threat Model
EAAP assumes the following conditions. Under all such conditions, EAAP fails closed.
2.1 Model Unreliability
LLMs hallucinate facts, contradict stored facts, misclassify sensitive data, or produce semantically false statements.
2.2 Prompt Injection
Attackers override constraints, induce self-contradiction, mask harmful instructions, or manipulate tool usage.
2.3 State Drift (TOCTOU)
Between evaluation and execution: roles change, constraints mutate, facts update, conflicts emerge, confidence decays.
2.4 Infrastructure Degradation
LLM outages, embedding unavailability, vector index degradation, circuit breaker activation, database latency.
2.5 Insider Risk
Manual database modification, audit log tampering, or replay of previously approved actions.
3. Architectural Layers
Purpose: Enforce ledger integrity.
PII scrubbing via deterministic pattern detection, encryption at rest, semantic indexing, conflict detection, confidence scoring, fact locking, and audit event logging.
Purpose: Assemble bounded, deterministic context.
Namespace isolation, deterministic relevance scoring, temporal decay weighting, conflict surfacing, context health classification, snapshot generation, and HMAC snapshot signing.
Purpose: Deterministic admissibility evaluation.
Authority validation, fact consistency enforcement, constraint evaluation, confidence threshold enforcement, and escalation classification.
Purpose: Guarantee execution integrity.
Claim extraction from payload, pre-flight conflict detection, full state hashing, evaluation record creation, commit validation, and immutable action ledger.
4. Protocol Invariants
Mandatory and non-configurable. Cannot be weakened without a major version change.
5. Evaluation Protocol
5.1 Required Inputs
{
"action_type": string,
"actor_role": string,
"impact_scope": string,
"payload": object
}5.2 Execution Sequence
state_hash = SHA-256(
sorted(relevant_objects) ||
policy_version ||
namespace_id ||
floor(timestamp, window)
)Evaluation does not execute the action. It generates an immutable trace with a signed context snapshot and recorded state hash.
6. Commit Protocol
{ "evaluation_id": string }The commit cannot modify the evaluation decision. The commit fails under state drift. The final state hash is recorded.
7. State Hash Requirements
The state hash must include: namespace state of ledger entries, constraints, actor roles, action types, and conflicts.
Algorithm: SHA-256 over canonical JSON serialization
Requirements: Deterministic key ordering, float normalization, no timestamp inclusion, no nondeterministic fields
8. Failure Modes
EAAP fails closed under all degradation conditions:
9. Versioning & Security Guarantees
Major versions require a major version increment and a 12-month deprecation window. Protocol invariants may not be weakened without a major version change.
EAAP does not guarantee LLM factual correctness, absolute uptime, or external API reliability. EAAP guarantees admissibility integrity, not model perfection.
Conclusion
The Exogram Action Admissibility Protocol establishes a deterministic control plane between AI inference and real-world execution. Inference remains probabilistic. Execution becomes admissible.