# E34 — Forensic Baseline & Reuse Map

**Status: PARTIAL.** One real new primitive was implemented (ADPO + candidate classification + replay
divergence); the rest reuses or is recorded as not implemented. No fabricated capability.

| E34 request | Authoritative component | Status |
|---|---|---|
| Autonomous Decision Provenance Object (ADPO) | **NEW** `cain/govdecision/core.py` (binds candidate set, selection, roots, epoch, refs) | IMPLEMENTED+TESTED |
| Candidate classification (CONSIDERED/REJECTED/SELECTED/INADMISSIBLE/UNKNOWN/NOT_EVALUATED) | **NEW** `classify_candidate` — enforces *unevaluated ≠ rejected* | IMPLEMENTED+TESTED |
| Governed selection proof | E18 `ControlSelectionProof` + **NEW** ADPO selection binding | IMPLEMENTED+TESTED |
| Counterfactual engine | E18 candidate search + futures + causal graph | IMPLEMENTED+TESTED |
| Byzantine decision evaluation | E18/E27 independence-aware quorum | IMPLEMENTED+TESTED |
| Decision consistency / mutation detection | E31 proof objects + E30 race engine | IMPLEMENTED+TESTED |
| Model/tool/runtime substitution detection | E27 context binding (agent/model/tool/runtime/resource/policy/environment) | IMPLEMENTED+TESTED (as context binding) |
| Decision drift detector | parts: E30 dependency invalidation; a named drift classifier | NOT_IMPLEMENTED |
| Outcome reconciliation | E16/E18 reconciliation + E23 performance fabric | PARTIAL (prediction/performance, not a decision-outcome object) |
| Decision memory / time machine / historical-state reconstruction | E22 replay + E31 chain (single-decision); full decision store | PARTIAL |
| Decision dispute engine | E32 dispute (trust), E30 dispute (cross-domain) | PARTIAL |
| Negative decision proofs | E31 `make_non_authorization` + E22 negative verification | IMPLEMENTED+TESTED |
| Decision replay + divergence | **NEW** `replay_decision` (REPRODUCED/DIVERGED/INCONCLUSIVE) | IMPLEMENTED+TESTED |
| Clean-room verifier for ADPO | not built; E22/CATCP/cross-domain Go verifiers exist | PARTIAL |
| Physical-AI decision record | E30/E33 abstraction only | DESIGN_ONLY |

## Genuinely new here
The **ADPO** (single canonical object binding the whole governed decision context) and the explicit
**candidate-classification semantics** (`NOT_EVALUATED`/`UNKNOWN` are never `REJECTED`) are the incremental
contribution. Everything else is E18/E30/E31 reuse.

## What is NOT differentiated (honest)
Decision provenance, counterfactual selection, and Byzantine evaluation are **known concepts** (E18 already
implemented counterfactual governance; E31 implemented proof chains and negative proofs). E34's value is
integration + the ADPO/classification semantics, not a new algorithm. A professional prior-art review is
required before any IP claim; none is made.

## Blockers (unchanged)
1. Enforcement is in-process (no real customer tool boundary).
2. No genuine third-party verifier (three same-project Go verifiers exist: E22, CATCP, cross-domain).