Add Fraud Decision Intelligence agent admission profile

[mdheller]

Purpose

Add an agent admission profile for Fraud Decision Intelligence agents.

Upstream doctrine anchor:

• SocioProphet/ProCybernetica:docs/capabilities/fraud-decision-intelligence-control-plane.md

Agent capabilities to classify

Fraud agents may be admitted for bounded activities such as:

• ingesting public-synthetic fraud events;
• producing evidence-bundle summaries;
• comparing model and legacy-rule outcomes;
• generating residual-fraud lift reports;
• proposing scenario candidates;
• preparing false-positive and false-negative reviews;
• proposing analyst-routing recommendations.

Prohibited or gated activities

Fraud agents must not, without explicit higher authorization:

• deny, suspend, or restrict a real person/account/customer;
• promote a model score to fraud proof;
• canonize a fraud label without provenance;
• write analyst notes into memory without review;
• convert a scenario candidate into an enforcement rule;
• ingest private customer evidence into public fixtures;
• execute live fraud-control actions.

Required receipts

Define admission and action receipt fields for:

• agent id;
• capability profile;
• event refs;
• evidence bundle refs;
• decision receipt refs;
• policy authorization refs;
• human-impacting action flag;
• memory-writeback review status;
• claim boundary.

Acceptance criteria

• Positive fixture: report-only fraud-analysis agent admitted for public-synthetic evidence review.
• Negative fixture: agent attempts human-impacting denial without policy receipt and is rejected.
• Negative fixture: agent attempts memory writeback before review and is rejected.
• Negative fixture: agent promotes model score to proof and is rejected.
• Documentation states Agent Registry owns admission metadata, not runtime fraud case management.

[mdheller]

Implementation progress:

• Added schemas/fraud-agent-admission-profile.v0.1.schema.json.
• Added valid report-only fixture examples/fraud-agent-admission-profile.report-only.example.json.
• Added invalid human-impact/no-policy fixture examples/fraud-agent-admission-profile.human-impact-no-policy.invalid.json.
• Added invalid memory-writeback/unreviewed fixture examples/fraud-agent-admission-profile.memory-writeback-unreviewed.invalid.json.
• Added stdlib validator tools/validate_fraud_agent_admission_profile.py.
• Added validate-fraud-agent-admission-profile Makefile lane and wired it into make validate.

Commit trail observed from connector:

• Schema: 510017ffa9a1e071e275b0e98b1f971e16210273
• Valid fixture: d852ae8479a3d026bd99ff34f4b886bbb9d5e238
• Invalid human-impact/no-policy fixture: 079dce8aff2e45aef81fc1bc5959e5c1fd011632
• Invalid memory-writeback/unreviewed fixture: 7eb70fdea278b73edae3c9c3685c1fcaef5ec0ad
• Validator: 8eebacaa21b4905028ac4b593c08364f2945e50e
• Makefile lane: 805703f9fff8c95c6061aacc9f33852c838ee3e0

Validation intent:

• Report-only fraud agent profile passes.
• Policy-gated/human-impacting profile without policyAuthorizationRef fails.
• Memory writeback profile without reviewReceiptRef fails.

Remaining before closure:

• Observe make validate-fraud-agent-admission-profile green.
• Observe full make validate green.

Boundary preserved: Agent Registry owns admission metadata only; runtime execution remains AgentPlane and policy authorization remains downstream policy/control-plane territory.