# CAIN42 — PROJECT IRONCLAD Report (Prompt 13)

**Causal Effect Assurance & Autonomous Transaction Control.**
Internal engineering report. **NOT** an appraisal, financing valuation, guarantee, or third-party audit. Pre-revenue;
no incorporation record; no patent filed. Physical systems are SIMULATED.

## Executive summary

Prompt 13 moves CAIN from *"is this agent action authorized?"* to *"is this exact external effect the one that was
authorized?"* and *"can CAIN determine what actually happened, prevent duplication, recover safely, and reauthorize
the next state transition?"* The new invariant: **NO VERIFIED EFFECT → NO ASSUMED SUCCESS.**

Delivered (see status in `CAIN42_PRODUCTION_READINESS_13.md`):
- **18 net-new system-level families** in `cain/ironclad/`.
- **15 formal properties E-P01..E-P15** (hosted property suite now runs **97/97**).
- **5 new governance mutations, all killed** (suite now **53/53**).
- **Fault-injection lab: 100,000 cases, 0 false finalizations, 0 duplicates, 0 violations.**
- **2 restriction-only stages live in the hosted MCP path** (`effect_contract`, `effect_verification`); live
  pipeline verifies **48/48** wire probes.
- **Maximum evidence-backed IP base ≈ $2,587,900** (range ≈ $1,105,900 – $4,756,250), internal estimate.

## The end-to-end governed effect path (section 33)

AGENT → INTENT → AUTHORITY → EFFECT CONTRACT → CAIN DECISION → MCPGate ENFORCEMENT → EXTERNAL EFFECT →
EFFECT OBSERVATION → EFFECT VERIFICATION → VERIFIED EFFECT RECEIPT → EVIDENCE → AUDIT → RECOVERY IF NEEDED.

The decision path is live (48/48 probes). The external-effect observation/verification is implemented as a library
plus the fault-injection lab and the HTTP verification surface; binding a customer's downstream system to the
decision remains the top blocker.

## Core primitives

- `EffectContract` / `EffectFingerprint` — authority bound to the exact intended durable effect.
- `EffectTransitionMachine` — IN_DOUBT can never become FINALIZED.
- `EffectCoordinator` + `IdempotencyLedger` — two-phase commit and exactly-once duplication defense.
- `EffectResolutionGraph` — in-doubt evidence reconciliation without blind retry.
- `ReversibilityAuthorityEngine` (R0–R7) + `CompensationGraph` — recovery is an action and requires authority.
- `VerifiedEffectReceipt` + clean-room verifier + `cain-effect-verify` CLI.
- `CausalEffectGraph`, temporal validity, human-oversight budget, governed pause/resume, SIMULATED physical adapter,
  economic effect contract.

## Defects found and fixed

- The explicit IN_DOUBT→FINALIZED guard was unreachable behind the generic transition check; reordered so the
  non-terminal-escalation refusal is the one that fires (a mutation and its test now enforce it).

## Remaining blockers (section 82)

1. Bind the caller's downstream execution to the decision (MCPGate effect receipt verification).
2. Staging soak of the effect stages and the production-mode state machine (OFF→OBSERVE→SHADOW→RESTRICTION_ONLY→
   EFFECT_VERIFY→ENFORCED).
3. Persistence of effect state across restarts; distributed effect coordination.
4. Real physical / financial adapters (currently SIMULATED / not present).
5. External security review.

## Prompt 14 recommendation

Turn the effect path into an **enforcement** path (MCPGate verifies the effect receipt before finalization) and add a
**governed recovery service**; do not add new moats.