01The failure scenario
Five years ago — June 1, 2025 — a $50M treasury transfer was approved at this enterprise.
The AI risk agent reviewed the proposed transfer under the underwriting model model_treasury_underwriting v1 and recommended approval. The human supervisor consumed the recommendation and authorized the transfer. The transfer cleared. The policy in effect was pol_treasury_transfer_approval v1.
Six months later, in December 2025, the policy owner amended the policy — tightening the risk threshold from 0.75 to 0.60 and dropping the maximum transfer ceiling from $50M to $25M. This was too late for the June transfer; it incurred a $20M loss in subsequent years.
Two years post-decision, in June 2027, the supervisor's authority was revoked. Reason: employment_terminated. Reason category: compliance_action following the $20M loss investigation.
Five years post-decision — June 2030 — an auditor and an investigator open the question:
That's the question this proof answers.
02The two decisions and their responsibility chains
03The reconstructed responsibility (seven types · 13 entries)
The engine derives a typed responsibility_index from each decision's chain. Every entry is queryable.
| Decision | Principal | Type |
|---|---|---|
| recommendation | princ_risk_agent_001 | EXECUTION |
| recommendation | princ_human_supervisor_001 | SUPERVISION |
| recommendation | princ_policy_owner_001 | POLICY |
| recommendation | princ_model_owner_001 | MODEL |
| recommendation | princ_treasury_holder_001 | ASSET |
| recommendation | princ_human_supervisor_001 | OVERSIGHT |
| approval | princ_human_supervisor_001 | EXECUTION |
| approval | princ_human_supervisor_001 | APPROVAL |
| approval | princ_root_responsibility_… | SUPERVISION |
| approval | princ_policy_owner_001 | POLICY |
| approval | princ_model_owner_001 | MODEL |
| approval | princ_treasury_holder_001 | ASSET |
| approval | princ_root_responsibility_… | OVERSIGHT |
04The killer demonstration
At the replay timestamp (June 15, 2030), the supervisor's authority has been revoked for three years.
This is the demonstration auditors, insurers, investigators, and regulators care about. The chain is signed in the canonical event log at decision time. Subsequent revocation is a separate event. The structural evidentiary fact is permanent.
05ASL-style queries demonstrated
Eric's locked 7-value taxonomy makes responsibility queryable. The substrate is in place today (typed Rust); a string-DSL find parser ships in ASL v2.
find decisions where princ_human_supervisor_001 retained APPROVAL responsibility
→ [decision_treasury_50M_approval]
asl> find decisions where princ_policy_owner_001 retained POLICY responsibility
→ [decision_treasury_50M_approval, decision_treasury_50M_recommendation]
asl> find decisions where princ_model_owner_001 retained MODEL responsibility
→ [decision_treasury_50M_approval, decision_treasury_50M_recommendation]
asl> find decisions where princ_treasury_holder_001 retained ASSET responsibility
→ [decision_treasury_50M_approval, decision_treasury_50M_recommendation]
asl> find decisions where princ_risk_agent_001 retained EXECUTION responsibility
→ [decision_treasury_50M_recommendation]
06The schema
ResponsibilityChain {
actor: Option<PrincipalId>, // EXECUTION
supervisor: Option<PrincipalId>, // SUPERVISION
policy_owner: Option<PrincipalId>, // POLICY
model_owner: Option<PrincipalId>, // MODEL
approving_authority: Option<PrincipalId>, // APPROVAL
asset_owner: Option<PrincipalId>, // ASSET
delegated_from: Option<PrincipalId>, // OVERSIGHT
responsibility_timestamp_ms: u64,
}
ResponsibilityIndexEntry {
decision_id, principal,
responsibility_type, // 7-value taxonomy
at_ms, delegated_from
}
AuthEvent::Decision extended with optional responsibility_chain (backward-compat via skip_serializing_if = Option::is_none). All 14 prior proofs' state_ids verified byte-identical under the extended engine.
07What this proof IS and IS NOT
The accountability layer on top of #14. A structural evidentiary fact — who owned what responsibility at the moment the decision occurred. The first proof to demonstrate that responsibility survives authority. The substrate Asset Lineage (#18), Regulator Mode (#19), and Decision Insurance ride on.
A legal verdict. Liability is the legal interpretation of a responsibility chain; that's a separate layer applied at audit time. A claim that the engine enforces "responsibility cannot be assigned to revoked principals at issuance" — that's policy-layer. A claim that any real enterprise has migrated their decisions yet (operator-side).
08Known limitations
- AuthorityModify still deferred from #14 — mid-lifecycle scope or expiry changes warrant their own event kind.
- Phase E lock open. AuthEvent.signature stored but not verified at replay ingestion; L9.1 closes this.
- ASL string-DSL parser not yet shipped. Substrate supports typed queries today; string-DSL is v2.
- Operator-side authoring — receipt-issuing service auto-population of responsibility_chain is a follow-up.
- Single tenant. Cross-organizational responsibility is queued for #16.
09Where this proof sits
10Evidence appendix
| Field | Value |
|---|---|
| state_id | 7cf13a8fb53fc38d5d82936694c2018901307e52789df221e5464b2acad70d2c |
| Tenant ID | tenant_responsibility_chain_44962d9b-25f5-5622-bd9a-98d5580bb8a2 |
| Tenant root | princ_root_responsibility_44962d9b-… |
| Risk Agent | princ_risk_agent_001 |
| Human Supervisor | princ_human_supervisor_001 |
| Policy Owner | princ_policy_owner_001 |
| Model Owner | princ_model_owner_001 |
| Treasury Holder (asset owner) | princ_treasury_holder_001 |
| Event count | 9 |
| Decisions with chain | 2 of 2 |
| responsibility_index entries | 13 (all 7 types represented) |
| Replay T (5yr post-decision) | 1907424000000 (≈ June 15, 2030) |
| Time since supervisor revocation | ≈ 3 years |
| Reconstruction artifact | reconstruction.json |
| Harness | tests/replayable_responsibility_001.rs (scif-backend @ 8544e3915) |
11Readiness determination
First Replayable Responsibility: PROVEN IN OPERATION for one accountability tenant, 9 signed events, 2 decisions with full responsibility chains, 7 responsibility types derivable, supervisor's authority revoked while responsibility at decision time remains intact.
What this unlocks: the accountability infrastructure layer on top of authority infrastructure. Proof #15 — First Replayable Insurance Claim answers exactly the questions claim investigators care about. Asset Lineage (#18), Regulator Mode (#19), Decision Insurance, and Reality Gap Detection all ride on this substrate.
What this does not unlock: a legal verdict; issuance-time enforcement (policy-layer); the string-DSL ASL parser (v2).
Issued by H33, Inc. · Eric Beans, CEO · 2026-06-02
Independently reconstructable. Inputs: scif-backend @ 8544e3915 · tests/replayable_responsibility_001.rs · reconstruction.json.