RFC-0009 · Reference catalog
Versioned draft snapshot
RFCs, reference catalogs, and evidence reflect Spec 3b2e3e7a114fb826b1445d39a48756b8fff629cd. Counts and pending/status notes belong to that snapshot. Read the dated project update for later progress.
Source repositories are private during staging. Links marked “repository access required” need access to the linked repository.
Generated reference catalog. This page renders structured Lean data for RFC-0009: defined terms, diagnostics, dependencies, and invariants. It does not establish proof coverage or runtime conformance. The authored narrative lives in
RFC-0009.This page is generated from the
lean/Gnomon/Spec/module for RFC-0009. Edit that source and regenerate this catalog.
Title: AI Participation Contract
Status: draft
Owners: AI Platform, Security, Semantic Core
Proof citations in source: 29 / 29. Standalone proof verification is pending.
Definitions
| Term | Meaning |
|---|---|
| Privilege escalation | When an AI agent invokes another agent or presents itself as a different actor with broader permissions than its own declared scope. |
| Prompt injection | Class of attacks where adversarial content embedded in external data is interpreted as instruction by an AI agent, causing unauthorized actions. |
| anomaly_policy | Declared policy on Claim |
| classify | Typed AI protocol assigning an gnomon instance or field value to one of N declared categories; output epistemic status: inferred. |
| extract | Typed AI protocol extracting structured field values from unstructured input; output status observed (raw) or inferred (interpreted). |
| fill | Typed AI protocol suggesting a value for a missing/incomplete field; requires user confirmation to commit as asserted. |
| flag | Output-only typed protocol raising an ambiguity, conflict, or anomaly for human review; AI MUST NOT resolve flags autonomously. |
| map | Typed AI protocol mapping external or natural-language terminology to a controlled vocabulary. |
| propose | Typed AI protocol creating a draft gnomon instance for human review; output status: estimated. |
| recommend | Typed AI protocol proposing an action with rationale; output must be draft and require human review; status: estimated. |
| request_confirmation | Typed AI protocol explicitly requesting human confirmation of an AI-produced value or action; blocks downstream operations. |
| summarize | Typed AI protocol producing a text summary of one or more gnomon instances; output status: inferred. |
Diagnostics
| Code | Severity | Description |
|---|---|---|
M09800 | Error | Agent declared without permissions block |
M09801 | Error | Agent declared without prohibitions block |
M09802 | Error | Agent permissions references protocol not in canonical set |
M09803 | Error | AI-produced output missing required evidence field |
M09804 | Error | Claim wrapper missing confidence where agent requirements mandate it |
M09805 | Error | Claim source does not reference a known model ID |
M09810 | Error | AI actor attempts to trigger transition to terminal/high-authority state |
M09811 | Error | AI actor attempts to write to field not covered by declared permissions |
M09812 | Warning | AI-authored declaration missing authored_by: ai tag |
M09820 | Error | AI-authored declaration approved without recorded human review |
M09821 | Error | AI-initiated deletion without human confirmation |
M09822 | Error | AI-produced .gnm file content not passed through gnomon fmt before writing |
M09823 | Error | AI-authored declaration missing draft epistemic modifier before human review |
M09830 | Error | Prompt template variable lacks declared type constraint — injection risk |
M09831 | Error | Claim |
M09832 | Error | Agent submitted command with actor identity other than its own declared identity |
M09833 | Error | Agent submitted command for ActionType not in its declared capabilities list |
M09834 | Error | Claim |
Invariants
- AI actors must not perform any action not listed in their permissions block. (enforced by: runtime)
- AI actors must not trigger transitions to terminal or high-authority states without explicit human involvement. (enforced by: runtime)
- AI-assisted authoring mutations to module declarations must be identifiable as AI-originated in the audit log. (enforced by: runtime)
- AI-authored declarations must carry the draft epistemic modifier until a human author has reviewed and explicitly approved them. (enforced by: compile-time)
- AI-produced .gnm content must be passed through gnomon fmt before writing or returning to the caller. (enforced by: compile-time)
- AI-produced Claim
values on gnomons with epistemic_status estimated or inferred must not be used in authorization decisions without explicit human or rule-based adjudication. (enforced by: runtime) - AI-produced fields must use Claim
wrappers carrying at minimum confidence and source. (enforced by: compile-time + runtime) - Agent declarations must be reviewed and explicitly approved before activation in non-development environments. (enforced by: install-time)
- Agents must not invoke another agent on behalf of a human actor without an explicit Command-confirmed delegation, and may not autonomously chain-invoke across trust boundaries. (enforced by: runtime)
- Agents writing Claim
values must strip or escape DSL syntax, shell metacharacters, and instruction-like patterns, and the kernel must reject claims exceeding the declared length bound. (enforced by: runtime) - An agent must submit commands with its own declared actor identity; any other identity must produce a kernel rejection. (enforced by: runtime)
- An agent’s invokable ActionTypes are exactly those in its permissions rule-block; commands for any other ActionType must be rejected before precondition evaluation. (enforced by: runtime)
- Declared anomaly policies must be enforced before the kernel processes Claim
writes; violations must be rejected and must not proceed. (enforced by: runtime) - Every AI actor must be declared as an agent with explicit permissions, prohibitions, and requirements blocks. (enforced by: compile-time)
- Model version changes must trigger re-evaluation of affected derivations, and stale derivations must be detectable. (enforced by: runtime)
- Prompt templates must structurally separate the declare-time instruction portion from the runtime data portion. (enforced by: compile-time)
- Prompt-template variable substitutions must pass through the declared PromptVariable type contract; untyped interpolation is forbidden. (enforced by: runtime)
- When agent A invokes agent B, the effective permission set is the intersection of A’s and B’s declared capabilities; neither may grant the other broader scope. (enforced by: runtime)
Normative Statements
Statements are structured source data. Proof citations, where present, have not passed a standalone Gnomon proof gate.
AI Authoring Contract
- MUST_NOT — AI MUST NOT delete existing declarations autonomously; deletions require human confirmation.
- MUST — AI-authored declarations MUST be reviewed by a human author before being marked as approved in the install plan. (proof citation:
rfc0009_b45_grounded; verification pending) - MUST — AI-authored declarations MUST be tagged with authored_by: ai and the model ID in the compiled bundle. (proof citation:
rfc0009_b40_grounded; verification pending) - MUST — AI-authored declarations MUST pass the same compiler validation as human-authored declarations. (proof citation:
rfc0009_b47_grounded; verification pending) - MUST — AI-authored declarations MUST use the draft epistemic modifier until human review is complete (M09823). (proof citation:
rfc0009_b64_grounded; verification pending) - MUST — Every AI tool producing .gnm content MUST pass it through gnomon fmt before writing or returning it (M09822).
- MUST_NOT — MCP interface for AI authoring MUST NOT bypass schema validation; it MUST return structured diagnostics. (proof citation:
rfc0009_b23_grounded; verification pending)
Compatibility and Migration
- MUST — Confidence threshold changes MUST be reflected in the install plan diff.
Evidence Obligations
- MUST — AI actors producing output MUST satisfy the evidence obligation declared in their requirements block.
- MUST_NOT — Enforcement MUST NOT rely solely on agent-side honesty.
- MUST — Outputs failing required evidence MUST be rejected with diagnostic M09803. (proof citation:
rfc0009_b50_grounded; verification pending) - MUST — The runtime MUST enforce evidence obligations at the point of gnomon write. (proof citation:
rfc0009_b11_grounded; verification pending)
Goals
- MUST_NOT — AI actors MUST NOT finalize decisions or trigger high-authority actions without explicit human approval. (proof citation:
rfc0009_b31_grounded; verification pending) - MUST — AI actors MUST be definable in terms non-technical reviewers can assess. (proof citation:
rfc0009_b55_grounded; verification pending) - MUST — AI-assisted authoring MUST remain auditable and reversible. (proof citation:
rfc0009_b0_grounded; verification pending) - MUST — AI-produced outputs MUST be structurally differentiated from human-verified data at the type level.
- MUST — Evidence obligations MUST be enforceable by the compiler and runtime. (proof citation:
rfc0009_b29_grounded; verification pending)
Invariants
- MUST_NOT — AI actors MUST NOT perform any action not listed in their permissions block. (proof citation:
rfc0009_b4_grounded; verification pending) - MUST_NOT — AI actors MUST NOT trigger transitions to terminal/high-authority states without human involvement.
- MUST — AI-assisted authoring mutations MUST be identifiable as AI-originated in the audit log. (proof citation:
rfc0009_b15_grounded; verification pending) - MUST — AI-produced fields MUST use Claim
wrappers carrying at minimum confidence and source. (proof citation: rfc0009_b2_grounded; verification pending) - MUST — Agent declarations MUST be reviewed and approved before activation in non-development environments. (proof citation:
rfc0009_b25_grounded; verification pending) - MUST — Every AI actor MUST be declared as an agent with explicit permissions, prohibitions, and requirements. (proof citation:
rfc0009_b35_grounded; verification pending)
Security and Authorization
- MUST — AI actors MUST be subject to the same RLS and authorization policies as human actors. (proof citation:
rfc0009_b57_grounded; verification pending) - MUST — Agent activation in non-dev environments MUST require a deployment approval step in the install plan. (proof citation:
rfc0009_b59_grounded; verification pending) - MUST — Agent capability scope MUST be enforced at the runtime boundary, not only at authoring time. (proof citation:
rfc0009_b17_grounded; verification pending) - MUST_NOT — Agents MUST NOT concatenate raw user input or untrusted content directly into prompt templates; all substitutions MUST pass through PromptVariable type contract.
- MUST_NOT — Agents MUST NOT invoke another agent on behalf of a human without explicit human Command confirmation, nor chain-invoke across trust boundaries autonomously. (proof citation:
rfc0009_b27_grounded; verification pending) - MUST — Agents MUST submit commands with their own declared actor identity; spoofed identity MUST be rejected (M09832). (proof citation:
rfc0009_b6_grounded; verification pending) - SHOULD — Agents producing Claim
SHOULD declare an anomaly_policy on output fields (max_value_length, allowed_confidence_range, disallowed_patterns). - MUST — Agents writing Claim
MUST sanitize value/rationale of DSL syntax, shell metacharacters, or instruction-like patterns; kernel MUST reject oversized values. - MUST — Commands for ActionTypes outside an agent’s permissions block MUST be rejected before precondition evaluation (M09833). (proof citation:
rfc0009_b13_grounded; verification pending) - MUST_NOT — Estimated/inferred Claim
values MUST NOT be used in authorization decisions without human or rule-based adjudication. (proof citation: rfc0009_b62_grounded; verification pending) - MUST — Prompt templates MUST structurally separate instruction (declare-time) from data (runtime) portions.
- MUST — The runtime MUST enforce declared anomaly policies before the kernel processes Claim
writes; violations produce M09834. (proof citation: rfc0009_b19_grounded; verification pending) - MUST — When agent A invokes agent B, effective permissions are the intersection of A’s and B’s declared capabilities. (proof citation:
rfc0009_b69_grounded; verification pending)
Training and Supervision
- MUST — AI models in derivation pipelines MUST declare model ID/version, training provenance, capability scope, and confidence calibration baseline. (proof citation:
rfc0009_b66_grounded; verification pending) - MUST — Model version changes MUST trigger re-evaluation of affected derivations. (proof citation:
rfc0009_b33_grounded; verification pending) - MUST — Stale derivations from older model versions MUST be detectable. (proof citation:
rfc0009_b9_grounded; verification pending)
Typed AI Protocols
- MUST_NOT — AI MUST NOT resolve flags autonomously; flag is output-only.
- MUST — An agent declaration MUST enumerate only protocols from the canonical typed-protocol set.
- MUST — propose output MUST be in a Draft or equivalent non-final state.
- MUST — recommend output MUST be a draft state and MUST require human review before finalization. (proof citation:
rfc0009_b42_grounded; verification pending)
Linked References
(no cross-RFC dependencies)