Trustless Submission Status · Phase C

✓ Package Integrity

✓ Independent Replay

• Signature Verification Not Applicable for this case (pre-L9.1)

✓ Evidence Package Included

✓ Responsibility Chain Present

✓ Lineage Present

✓ Vendor Independence

VerdictReviewable

A regulator should never need to understand H33 to verify H33 evidence. The verdict above answers one question: can I trust this package? Drill into findings + raw evidence below — see [ Findings ] · [ Verifier runbook ].

Regulator workflow · the four steps

1Open submissionsubmission.tar.gz · 8 files including the embedded evidence package
2Verify packageRun h33-independent-canonical-replay — confirm state_id byte-identical
3Review findingsfinding_report.json — the assertion the regulator is investigating, with evidence_refs
4Drill into evidenceWalk decision/asset/loss/claim lineage from the canonical event log

Evidence Binder · Case File · auto-generated

What caused this loss?

Case: Claim #84711

Investigation typeInsurance Claim Defense·Primary buyerInsurer · Reinsurer · Claims Counsel

Case ID

claim-84711

Claim ID

claim_84711

Asset

loan_84711_borrower_principal

Loss

4200000.00 USD

Loss event ID

loss_loan_84711_default

Primary decision

decision_loan_84711_approval

Tenant

tenant_insurance_claim_44962d9b-25f5-5622-bd9a-98d5580bb8a2

Tenant root

princ_root_claim_44962d9b-25f5-5622-bd9a-98d5580bb8a2

Replay state_id

e72d3c0e71a11ce0aaf1e8c9eb5c720aff49a6238c76976b4f4435b50e43bee2

Replay T (ms)

2064614400000

Binder Tabs · click to drill in

Show decision lineage

What happened, in causal order

Tab 3

Show responsibility chain

Responsibility ≠ Liability

Tab 4

Show asset lineage

Asset ≠ Responsibility

Tab 5

Show model influence

Influence ≠ Causation

Tab 6

Reproduce decision

Reproducibility ≠ Justification

Export evidence package

Institutional Memory ≠ Legal Truth

Tab 9 · Phase C

Review findings

What the evidence demonstrates

Tab 10 · Phase C

If H33 disappears mid-review

Package still verifies. Without us.

Tab 11 · Showcase

Ask The Evidence

Natural-language queries over signed evidence

Three Confidence Scores · honest, not aspirational

Reconstruction Confidence

85 / 100

Verification Confidence

Pre-L9.1 EvidenceNot Applicable for this case

Reproducibility Confidence

63 / 100

About this case

This claim was filed on 2026-03-20 for a $4.2M credit-default loss recognized on 2026-03-15. The underlying credit decision was approved by princ_credit_officer_001 on the recommendation of princ_credit_risk_agent_001 (an AI risk agent) on 2025-04-15. The enterprise that filed this claim was dissolved on 2031-04-01. This case file is reconstructed from signed canonical events four years after dissolution — the entire chain remains independently verifiable without H33 and without the dissolved firm.

Tab 1 · Verify evidence

Verify the evidence — independently, without H33

Replay at target T produces state_id e72d3c0e71a11ce0aaf1e8c9eb5c720aff49a6238c76976b4f4435b50e43bee2 byte-identically. Confirm by running the verifier binary against the exported package (see Tab 8).

state_id at T	e72d3c0e71a11ce0aaf1e8c9eb5c720aff49a6238c76976b4f4435b50e43bee2
Replay verdict	Valid
Replay confidence	85 / 100
Verification confidence	Pre-L9.1 Evidence · Not Applicable for this case

Verified ≠ True.Verification proves integrity. Not correctness.

Tab 2 · Show decision lineage

Decision lineage — what happened, in causal order

       1744675200000  decision_loan_84711_recommendation
      actor:      princ_credit_risk_agent_001
      capability: recommend_credit
      outcome:    recommend_approve
      policy:     pol_credit_underwriting:1
      model:      model_credit_underwriting:1
      parents:    []

       1744678800000  decision_loan_84711_approval
      actor:      princ_credit_officer_001
      capability: approve_credit_application
      outcome:    approved
      policy:     pol_credit_underwriting:1
      model:      model_credit_underwriting:1
      parents:    ["decision_loan_84711_recommendation"]

Lineage shows what happened, not whether it was right.

Tab 3 · Show responsibility chain

Responsibility chain — bound at decision time

Chain on the primary decision (decision_loan_84711_approval):

actor	princ_credit_officer_001
supervisor	princ_root_claim_44962d9b-25f5-5622-bd9a-98d5580bb8a2
policy_owner	princ_policy_owner_underwriting
model_owner	princ_model_owner_credit_underwriting
approving_authority	princ_credit_officer_001
asset_owner	princ_borrower_loan_84711
delegated_from	princ_root_claim_44962d9b-25f5-5622-bd9a-98d5580bb8a2
responsibility_timestamp_ms	1744678800000

Responsibility ≠ Liability.Naming a principal in the chain establishes a structural role at decision time; liability is a court / regulator determination.

Tab 4 · Show asset lineage

Asset lineage — Asset → Decisions → Loss → Claim

Asset:            loan_84711_borrower_principal
  Decisions:      2
  Loss events:    1
  Claim events:   1
  Consequences:   0
  Aggregate $:    $4200000.00

Asset ≠ Responsibility.The walk shows what touched the asset. Ownership is a legal status; responsibility is structural fact.

Tab 5 · Show model influence

Model influence — why the model produced the score it produced

No ModelInfluenceRecord bound to the primary decision.

Influence ≠ Causation.The chain establishes that the model's score depended on these features; it does NOT establish that those were the right features or that the threshold was fair.

Tab 6 · Reproduce decision

Reproduce decision — measured confidence, not aspirational

Reproducibility confidence: 63 / 100. Five components × 20 points each. See #167 Decision Reproducibility for the rubric.

Reproducibility ≠ Justification.The chain runs again the same way. That's not the same as saying it should have.

Tab 7 · Show reality gaps

Reality gaps — what changed without authorization?

0 gaps detected

0 Shadow Authority
0 Undocumented Delegation
0 Policy Bypass
0 Responsibility Drift

Policy ≠ Reality.The report names structural disagreements between policy and recorded reality. It does NOT name fault, intent, or correctness.

Tab 8 · Export evidence package

Export the evidence package — and verify it without H33

The evidence package bundles the canonical event log, the case summary, and a verification runbook. No H33 access required at any step.

events.json	The signed canonical AuthEvent log for this tenant
manifest.json	tenant_id · tenant_root · target T · expected state_id · expected verdict
public_keys.json	Per-principal PQ public-key bundles (L9.1; empty for pre-L9.1 cases)
signatures.json	Per-event full PQ signature archive (L9.1; empty for pre-L9.1 cases)
case_summary.json	Three-confidence scores at export time + case metadata
verifier_runbook.md	Step-by-step runbook for independent verification

How to verify independently (regulator runbook)

Download the evidence package (button below).
Build the verifier: cargo build --release --bin h33-independent-canonical-replay from the open-source scif-backend repo.

Extract the tarball, then run:

h33-independent-canonical-replay \
    --events-file events.json \
    --manifest-file manifest.json \
    --verbose

Confirm the binary's reported state_id matches e72d3c0e71a11ce0aaf1e8c9eb5c720aff49a6238c76976b4f4435b50e43bee2.
For cases under L9.1, add --public-keys-file public_keys.json --signatures-file signatures.json --verify-signatures. Pre-L9.1 cases (verification = Not Applicable) skip this step.

⬇ Download submission.tar.gz (8.4 KB) · the canonical Phase C artifact · 8 entries including the embedded evidence_package.tar.gz

⬇ Or download evidence-package.tar.gz alone (4.3 KB) — the component, not the full submission

Phase C submission carries 8 entries: case_summary · verification_report · reconstruction_report · lineage_report · responsibility_report · finding_report · evidence_package.tar.gz · verifier_runbook. The submission is what a regulator opens; the evidence package is the substrate it cites.

Institutional Memory ≠ Legal Truth.The package preserves structural memory. Legal truth — verdicts, judgments, damages — remains the courts' determination.

Findings · what the evidence demonstrates

Findings — not raw evidence, the assertions the evidence supports

Regulators investigate FINDINGS — not raw evidence. Each finding below carries a claim, the evidence that supports it, the responsibility chain it implicates, the lineage it walks, and a status (open / investigating / substantiated / dismissed).

finding_claim_84711_loss_recognitionfinancialsubstantiated

Loss of 4200000.00 USD recognized on asset loan_84711_borrower_principal was the subject of claim claim_84711 filed against the reinsurer. The credit decision approving the loan was made by the credit officer on the AI risk agent's recommendation. The chain that produced the loss is independently replayable from canonical events.

Evidenceloss_event:loss_loan_84711_default
claim_event:claim_84711
primary_decision:decision_loan_84711_approval

Responsibilityactor:princ_credit_officer_001
approving_authority:princ_credit_officer_001
asset_owner:princ_borrower_loan_84711
policy_owner:princ_policy_owner_underwriting

Lineagedecision_chain
asset:loan_84711_borrower_principal
loss_lineage

finding_claim-84711_reality_cleaninformationaldismissed

Reality Gap Detection (#177) returns 0 gaps on tenant tenant_insurance_claim_44962d9b-25f5-5622-bd9a-98d5580bb8a2 across all four classes. Clean baseline.

Evidencereality_gap_report

Responsibility—

Lineage—

Asset ≠ Responsibility. Responsibility ≠ Liability. A finding's status is structural — what the evidence demonstrates — not a legal verdict. Liability is determined by courts and regulators.

Killer demo · the package was already in your possession

What happens if H33 disappears before you finish reviewing?

Imagine: this submission was delivered to your office. You begin review. Mid-review, H33 disappears — vendor failure, dissolution, sanctions, supply-chain compromise, whatever. The reviewing regulator has no live H33 channel, no support phone, no auth endpoint. The package on disk is all that remains.

✓
Package still verifies.
h33-independent-canonical-replay is a standalone binary built from the open-source scif-backend repo. No DB. No SCIF. No network. The reported state_id matches e72d3c0e71a11ce0aaf1e8c9eb5c720aff49a6238c76976b4f4435b50e43bee2 byte-identically.
✓
Responsibility still reconstructs.
responsibility_report.json was derived from canonical events and ships inside submission.tar.gz. The chain at decision time is permanently recoverable.
✓
Lineage still reconstructs.
lineage_report.json walks Decision → Asset → Loss → Claim directly on the events. Reproducible without any vendor surface.
✓
Verdict still reproducible.
finding_report.json carries the claim + evidence_refs + responsibility_refs + lineage_refs. Every reference walks back to a signed canonical event in the embedded package.

Distinct from #12 (vendor dies) and #184 (customer dies) — here the package was already in the regulator's possession when H33 disappeared, and the entire review still completes. The category claim Phase C earns: a regulator can consume H33 evidence without H33.

Tab 11 · Ask The Evidence · showcase

Ask The Evidence — natural-language queries over signed evidence

The search architecture that ties the corpus together: Ask The Evidence → Search Twin → Replay Engine.

ask> Who approved this?

→ Look in responsibility_chain.approving_authority on the primary decision.

ask> Who could have approved this?

→ Filter active_grants at decision T by the required capability.

ask> What policy governed this?

→ Decision.policy_version_ref. Amendments after-the-fact don't apply.

ask> What changed after approval?

→ Walk the event timeline post-decision: PolicyAmend, ModelRegister, Loss, Claim, Consequence, Dissolution.

ask> Who retained responsibility?

→ Decision.responsibility_chain — recorded at decision time, permanently recoverable.

ask> Were there reality gaps?

→ show_reality_gap_report(tenant) returns the four-class breakdown (Shadow Authority / Undocumented Delegation / Policy Bypass / Responsibility Drift).

ask> What evidence survives?

→ The canonical event log. Replayable byte-identically without H33 and without the dissolved firm.

Ask The Evidence is a translator. The verifier is the truth.The LLM never decides authority / responsibility / verification — those come from replay and signed indexes.

Page generated by h33-evidence-case-generator from case_descriptor.json. Reproducible: same descriptor + same database state = same case file.