# CAIN-42 — The Governed Path to Vertical AGI

**2027 category position · FIRME AI CORPORATION · 2026-11-17**
**Status:** canonical positioning. It is a claim to be earned with published evidence, not an assertion that AGI exists today.

---

## 1. The position

> **CAIN-42 is the governor of Vertical AGI — the governed path to vertical AGI: one intelligence architecture, many domains, thousands of agents, every consequential action authorized, enforced, recorded, and independently verifiable.**

This is deliberately **not** the claim "CAIN-42 is an AGI." That claim is both unprovable today and strategically weaker. The scarce thing in 2027 is not another model; it is the ability to let autonomous intelligence touch money, health data, infrastructure, contracts and customers **without losing control of it**. CAIN owns that layer.

**Category we create:** *Governed Vertical AGI Infrastructure* (GVAI).
**One-line:** *One intelligence. Many domains. Every action proven.*
**Proof line:** *Don't trust the agent. Verify the receipt.*

---

## 2. Why this is the right claim

Vertical AGI is an **outcome** — expert-level work in a domain. Three things are required to deliver it responsibly at scale:

1. an intelligence core that reasons, plans, learns and transfers across domains;
2. a workforce fabric that runs thousands of specialists for long missions;
3. a governance and evidence layer that authorizes, enforces, contains and proves every consequential action.

The market has (1) and parts of (2). It does not have (3) as a product. CAIN has been building (3) since day one — Byzantine evidence, cryptographic receipts, delegated authority, kill switch, portable verification — and is now adding (1) and (2) on top of it. That combination is the first-mover position.

**The defensible sentence:** *Other companies sell intelligence and ask you to trust it. CAIN sells governed intelligence and gives you the receipt.*

---

## 3. Competitive map

| Player class | What they are strong at | What they cannot give a regulated buyer | CAIN's difference |
|---|---|---|---|
| Frontier labs (OpenAI, Anthropic, Google, DeepSeek) | Raw model capability | Per-action authority, containment, portable evidence, multi-vendor neutrality | Model-agnostic; governs any model; proves the action |
| Agent frameworks (LangGraph, AutoGen, CrewAI, MCP) | Orchestration, integrations | An enforcement boundary, a kill switch with evidence, an offline verifier | Enforcement is outside the agent; receipts outlive the run |
| Security / AI control planes (Zenity, HiddenLayer, Bay, Microsoft) | Visibility, posture, policy | A general intelligence core and multi-domain workforce | Intelligence + governance in one architecture |
| Consultancies / systems integrators | Bespoke delivery | A product, a standard, repeatable evidence | Productised engines + signed artifacts |
| **CAIN-42** | **Governance + evidence + execution boundary** | **Raw frontier capability today** (addressed by model-agnostic routing and partnerships) | **The only one combining all three layers** |

**First-mover assets:** a defined evidence format (E84 receipts), an offline verifier that imports no CAIN code, a hierarchical kill switch with signed shutdown evidence, and an installed base of audit workflows. These are hard to retrofit onto a framework that started as an orchestrator.

---

## 4. The proof stack (what is actually true today)

| Layer | Evidence | Classification |
|---|---|---|
| Byzantine containment | E83 disposable 4-node equivocation run, 30/30 verifier checks, 3/3 honest proofs | **MEASURED / LOCAL** |
| Live failover | E83 paused-primary drill, 8/8 verifier checks | **MEASURED / LOCAL** |
| Verifiable action receipt | E84 one-action receipt, 24/24 attacks blocked, `VERIFIED_COMPLETE` | **LOCAL PROTOTYPE** |
| Hierarchical kill switch | E85 `HierarchicalKillSwitch`, 15/15 tests, ten scopes, durable epochs | **IMPLEMENTED / TESTED** (in progress) |
| Governed execution | E85 `GovernedExecutor` + `GuardedResource`, exact-request + kill-switch epoch | **IMPLEMENTED / TESTED** (in progress) |
| Reasoning core / evaluation | E85 work in progress; local 0.5B models only | **FOUNDATION / NO AGI CLAIM** |
| Independent witness | 0 independent cosignatures; operator witness labelled | **BLOCKED** |
| Production deployment | No production service changed | **NOT AUTHORIZED** |
| Hardware attestation | No TEE on host | **BLOCKED** |

Every row above has an artifact path or a test. None is a slide.

---

## 5. The 2027 claim ladder

The phrase "closest thing to vertical AGI" is comparative. We earn it by publishing the comparison, not by asserting it.

### Today (Q4 2026) — the foundation
- One governed action receipt, verified offline.
- Hierarchical kill switch with durable state and signed shutdown evidence.
- Governed execution path with exact-request enforcement.
- One measured vertical pilot (audit/evidence workflow).
- **What we say:** "the governed execution fabric with cryptographic receipts."

### Q1 2027 — the intelligence layer
- Model-agnostic reasoning core (local + remote providers).
- Persistent, provenance-aware memory; verified-only retrieval; contamination quarantine.
- Held-out multi-domain evaluation suite; single-agent vs multi-agent comparison.
- Three narrow vertical prototypes with domain evaluation sets.
- **What we say:** "a governed multi-domain agent system with measured capability."

### Q2 2027 — the workforce
- 100–1,000 active workers; hierarchical swarms; dynamic specialist selection.
- Mission recovery, leases, fencing, idempotency.
- 10,000 verified productive agent-hours across ≥10 fields.
- Multi-tenant evidence isolation; ClawX workforce/evidence interface.
- **What we say:** "the largest governed agent workforce with per-action proof."

### Q3 2027 — generalization
- Cross-domain transfer benchmarks; unfamiliar-task evaluation.
- Measured learning improvements on held-out tasks (no contamination).
- Independent witness network (2+ organizations); customer pilots.
- **What we say:** "the closest thing to vertical AGI you can actually deploy."

### Q4 2027 — the category
- Public multi-domain benchmark: **≥80% on a 500-task, 10-domain held-out suite** (per-domain reported honestly; no single aggregate hiding a weak domain).
- **1,000 concurrent agents, 10,000+ verified productive hours.**
- Kill switch **<1 s** with signed evidence and adversarial tests.
- 2+ independent witnesses, 3 paying customers, one third-party audit.
- Every consequential action carries a portable receipt.
- **What we say:** "CAIN-42 is the closest thing available to vertical AGI — and the only one that can prove it."

If any criterion is not met, the sentence is not used. The claim is gated by the matrix, not by the calendar.

---

## 6. The five messaging pillars

1. **One intelligence, many domains.** A single reasoning/planning/memory core; domains are packages, not separate products.
2. **Authority is a lease, not a permission.** Identity, delegation, policy, single-use grants, revocation, kill switch.
3. **Every action leaves a receipt.** Exact request, quorum where used, enforcement, execution, observed effect, log inclusion, signed checkpoint, witness.
4. **Proof, not promises.** The verifier imports no CAIN code. A stranger can check the artifact without trusting us.
5. **Scale with a kill switch.** Thousands of agents are only useful if one signed command can stop them — and prove it.

---

## 7. Claims discipline (non-negotiable)

**Allowed today:** governed execution fabric; cryptographic action receipts; hierarchical kill switch; governed multi-domain agent architecture; measured local prototypes; model-agnostic by design.

**Allowed only when the gate is PASS:** closest thing to vertical AGI; multi-domain expert capability; 1,000-agent production workforce; independently witnessed; third-party audited; hardware-attested.

**Never allowed without evidence:** "AGI achieved", "human-level", "production-ready", "independently verified" for our own clusters, "SOC 2 certified", "SLSA Level 3", "notarized" for the vendor quorum, "court-admissible."

The full gate list is `docs/evolution-84/E84_ACCEPTANCE_MATRIX.md` plus the E85 matrix. A public claim that cannot cite a gate is removed.

---

## 8. Why first mover — and why now

- **The evidence format is being created now.** RFC 6962 + C2SP witnesses + a portable receipt is a standard-shaped artifact. Whoever defines the shape first becomes the default.
- **Governance is hard to retrofit.** A framework that starts as an orchestrator cannot bolt on an enforcement boundary, a kill switch with epochs, and an offline verifier without re-architecting.
- **Trust is the buying barrier.** Regulated buyers do not ask "which model?" first; they ask "who can prove what happened?" CAIN answers that.
- **The audit business creates distribution.** Every audit produces a signed evidence pack and a customer reference; that is a wedge into the platform.
- **The 2027 market will consolidate.** The winner is likely the trust layer that every vertical AGI must pass through, not the nth vertical app.

---

## 9. The one metric

**Verified productive agent-hours under governance, with a portable receipt.**

Not modules. Not evolutions. Not parameter counts. Today the number is single-digit verified hours. The 2027 target is 10,000+. Every evolution must move it or it is decoration.

---

## 10. What "fully done" looks like

CAIN-42 is not "an AGI." When fully done it is:

> **the trust, control and evidence fabric for the autonomous economy — the layer that lets any vertical AGI, in any domain, act with delegated authority, be contained in under a second, and leave a receipt a third party can verify offline.**

That is a larger and more defensible outcome than being one vertical AGI. It is also the only one consistent with the evidence we can actually produce.
