H33
L5 · Reconstructed · June 2, 2026

First Time Travel Replay.
What did it look like when this decision was made?

Five replays. Five distinct moments in time. Five distinct state_ids. Each one captures the exact model versions, policy versions, and decisions in effect at that moment. Replays at the same T are byte-identical forever. Decision lineage reconstructs the causal chain: the human's approval consumed the AI's recommendation. Replay confidence is scored honestly — including the open Phase E signature gap (L9 closes it).

Auditors Regulators Insurance Underwriters Transfer Agents Fund Administrators AI Governance Buyers CISOs · Risk
What was proven · 10-second read

Five state_ids at five moments in time. Byte-deterministic forever.

01
Replay at distinct T values reconstructs substantively distinct state.
02
Replays at the same T are byte-identical, forever.
03
Decisions carry causal lineage. Replay confidence is scored honestly.
Reading any H33 proof · the six questions

Same six answers. Different scope. The reader recognizes the machine.

  1. 1What happened?

    11 signed canonical events recorded a small underwriting flow: one policy registered + amended (v1→v2), one model registered + retrained (v1→v2), two AI recommendations, two human approvals consuming them. Replays at 5 different T values reconstruct 5 distinct snapshots.

  2. 2Who had authority?

    Human supervisor princ_customer_9 via the underwriting policy (approve_underwriting); AI princ_ai_underwriter_001 delegated by the human (recommend_underwriting only).

  3. 3How was authority reconstructed?

    replay_until(events, T, …) processes the three new event kinds (policy_amend, model_register, decision) in time order; the snapshot now carries active_policy_versions, active_model_versions, decisions_up_to.

  4. 4What state was produced?

    Five distinct state_ids at T5/T6/T8/T10/T∞. Each captures the exact policy & model versions and decisions present. The full table is in section 03 below.

  5. 5What artifact was returned?

    reconstruction.json — full timeline, all five snapshots, decision lineage examples, replay confidence with named checks.

  6. 6How can a third party verify it?

    Run scif-backend tests/time_travel_replay_001.rs at SHA d29cc7c33. Expect identical state_ids at each T, byte-deterministic; expect decision_002.parent_decision_ids = [decision_001]; expect replay confidence at T10 = 72/100.

01The timeline (11 signed events)

Underwriting tenant · tenant_time_travel_44962d9b-…
1780440000000
policy_register
pol_underwriting declared
1780440001000
model_register
model_underwriting v1 (weights + training fingerprint)
1780440002000
grant
root → princ_customer_9 (approve_underwriting + delegate)
1780440003000
grant
human → princ_ai_underwriter_001 (recommend_underwriting only)
1780440004000
policy_amend
pol_underwriting v1 pinned with content_hash
1780440005000
decision_001
AI recommends application_001 (bound to model_v1, policy_v1)
1780440006000
decision_002
Human approves application_001 · parents=[decision_001]
1780440007000
policy_amend
pol_underwriting v1 → v2 (tightened threshold + EU jurisdictions)
1780440008000
model_register
model_underwriting v2 (retrained)
1780440009000
decision_003
AI recommends application_002 (bound to model_v2, policy_v2)
1780440010000
decision_004
Human approves application_002 · parents=[decision_003]

02The schema extensions (the L5 substrate)

New event kindRecords
policy_amend(policy_id, prev_version, new_version, content_hash, amended_by, at_ms)
model_register(model_id, version, weight_hash, training_fingerprint, registered_by, at_ms)
decision(decision_id, actor_principal, capability, subject, model_version_ref, policy_version_ref, outcome, parent_decision_ids, at_ms)

Three new snapshot fields are added with #[serde(default, skip_serializing_if = "Vec::is_empty")]. Backward compatibility verified: every prior proof's published state_id was reverified under the extended engine and matched byte-for-byte (V101 96a29047…, Tokenized cc0d4369…, L1 1cbd6979…, L2 b52fe565…, L3 5aefda52…, L4 2a4bf5f6…).

03The time travel · five distinct state_ids

T5after decision_001 · T = 1780440005000
1890b20c61daa91ed6079b0215f3c99c5b61d1e7031f9d8fc3ddeb91b72b0025
policy versions: 1 (v1) model versions: 1 (v1) decisions: 1 (decision_001)
T6after decision_002 (lineage visible) · T = 1780440006000
70fdc855447623918c2ee1b5cdfd3550ca20273a4116a3cd4bd33892508b91e8
policy versions: 1 (v1) model versions: 1 (v1) decisions: 2 (decision_002.parents = [decision_001])
T8after model v2 register · T = 1780440008000
deb7f04a0e6bbf48cf3e68817a1aece75c7849da0380b3cc992eaa62f928eb60
policy versions: 2 (v1 + v2) model versions: 2 (v1 + v2) decisions: 2 (3 & 4 not yet)
T10after decision_004 · T = 1780440010000
b07974aed797856dc47ca07f423124804a1096cb892294c57fb902db149cde50
policy versions: 2 model versions: 2 decisions: 4 (decision_004.parents = [decision_003])
T∞far future · T = 1800000000000
0f0e51dd8c35d13d53de9b49c7e72f1926160b19f8f5d5e1b55f0c7cd1770c97
policy versions: 2 model versions: 2 decisions: 4 (same content as T10, distinct T → distinct state_id)

T10 and T∞ share the same content but have distinct state_ids because timestamp_t_ms is a field in the snapshot. Asking "what was true at T10" returns a different artifact than asking "what is true now" — even when nothing else has changed. That's the right behavior.

04Decision lineage (the causal chain)

decision_002 — why was application_001 approved?
▸ princ_customer_9 approved · capability=approve_underwriting · model_v1 · policy_v1
   ↑ consumed
▸ princ_ai_underwriter_001 recommend_approve · capability=recommend_underwriting · model_v1 · policy_v1

→ Encoded as decision_002.parent_decision_ids == ["decision_001"]

Three years from now a regulator asks "why was application_001 approved?". The chain is reconstructable from the event log: human approved (decision_002) ← AI recommended (decision_001) ← (model v1 + policy v1 in effect at T) ← (grant chain to root). Every output has a causal chain.

05Replay Confidence Score (honest completeness metric)

Replay confidence at T10 72/100
authority_chain
All grants traced to root.
Critical
signatures_verified_at_replay
Phase E lock: AuthEvent.signature is stored but not verified at replay ingestion. Standalone L9 verifier will close this.
Critical
policy_versions_present
Every decision's policy_version_ref resolves to a registered policy version.
Warning
model_versions_present
Every decision's model_version_ref resolves to a registered model version.
Warning
decision_lineage_resolves
Every decision's parent_decision_ids resolve to earlier decisions.
Info

The honesty is the feature. A regulator/insurer/auditor needs to know HOW COMPLETE a replay is — not just that it ran. The 72/100 score reflects the one open critical gap (Phase E signature verification) directly. L9 Evidence Survivability closes that gap with a standalone verifier. Weighted scoring: Critical = 3× · Warning = 2× · Info = 1×.

06Why this matters

Eric, June 2, 2026

"Most companies can tell you what their organization looks like today. H33 can prove what it looked like when a decision was made five years ago and replay why that decision occurred."

That's the sentence this proof earns. Time travel + decision lineage + replay confidence + ASL (from L4) = the substrate. The next proof — First Replayable Organization — composes the L1-L5 substrates and names what they describe: a company as a reconstructable graph.

07Known limitations

  1. Phase E lock — signatures stored but not verified at replay ingestion. Honestly reported in the Replay Confidence score. L9 closes it.
  2. Policy content referenced by hash, not stored. Full policy text lives in a content-addressable store.
  3. Model weights referenced by hash, not stored. Same pattern.
  4. Reconstruction-only, not live decision issuance. Decisions seeded via the signing CLI for demonstration.
  5. ASL time-travel syntax (descendants(p) AS OF T_past) not in v1. Engine supports time travel; ASL gets the syntax in v1.1.

08Where this proof sits in the ladder

L1
Agent Recommendation. first-ai-assisted-transfer
proven
L2
Agent Authority Envelope. first-agent-authority-envelope
proven
L3
Agent Supervisor Chain. first-agent-supervisor-chain
proven
L4
Tenant-Scoped Agent Hierarchy + ASL v1. first-tenant-agent-hierarchy
proven
L5
Time Travel Replay — 3 new event kinds, decision lineage, replay confidence. This proof.
proven now
#12
First Replayable Organization (meta) — composes L1-L5 substrates and names what they describe. Next.
next
L6
Counterfactual Replay (authority-counterfactual on the L5 substrate).
roadmap
L7
Authority Drift Detection (graph diffs across T).
roadmap
L8
Blast Radius Live API.
roadmap
L9
Evidence Survivability + standalone verifier (closes Phase E).
the moat
L10
Regulator Mode (depends on L5).
roadmap
L11
Organizational Memory (depends on L5).
roadmap

09Evidence appendix

FieldValue
Tenant IDtenant_time_travel_44962d9b-25f5-5622-bd9a-98d5580bb8a2
Tenant rootprinc_root_time_travel_44962d9b-…
Humanprinc_customer_9
AIprinc_ai_underwriter_001
Policypol_underwriting (v1 → v2; content hashes in reconstruction.json)
Modelmodel_underwriting (v1 → v2; weight + training fingerprints in reconstruction.json)
Event count11
Time-travel snapshots5
Decisions (with lineage)4 (002→001, 004→003)
Reconstruction artifactreconstruction.json
Harnesstests/time_travel_replay_001.rs (scif-backend @ d29cc7c33)

10Readiness determination

Determination

First Time Travel Replay (L5): PROVEN IN OPERATION for one underwriting tenant, 11 signed events, 2 model versions, 2 policy versions, 4 decisions with causal lineage, 5 distinct deterministic state_ids at 5 distinct T values, Replay Confidence reported honestly.

What this unlocks: L6 (Counterfactual), L7 (Drift Detection), L10 (Regulator Mode), L11 (Organizational Memory), and the next proof — #12 First Replayable Organization, the meta-proof that frames the corpus's capability as describing a company as a reconstructable graph.

What this does not unlock: live decision issuance; ASL time-travel syntax; policy/model content payload storage; standalone independent verification (L9).

Issued by H33, Inc. · Eric Beans, CEO · 2026-06-02

Independently reconstructable. Inputs: canonical event log access · scif-backend @ d29cc7c33 · harness tests/time_travel_replay_001.rs.