# Evolution 85 — Roadmap, November 2026 to December 2027

**FIRME AI CORPORATION · CAIN-42 · Evolution 85** — branch `evolution-85`, baseline Evolution 84 (`2c419a042`). Development result. Not deployed. Not released. Built by the three-person team.

A target roadmap built from capability gates. It is not a forecast and not a promise.

**2027 positioning target:** CAIN-42 known as the governor of vertical AGI — the layer vertical AI workforces run
under. That reputation has to be earned by published measurements that others re-check; the gates below are what
would support it.

| Quarter | Gate | Evidence that closes it | Today |
|---|---|---|---|
| Q4 2026 | E84 verified | clean-copy reproduction | done |
| Q4 2026 | E84 G38 resolved | kill switch in the path; bypass prevented | kill switch done; bypass detected, prevented only for sandboxed workers |
| Q4 2026 | Evaluation harness | held-out results with intervals | done on small local models |
| Q4 2026 | One governed external resource | receipt over a third-party resource | second host in the same admin domain only |
| Q4 2026 | Hosted MCP path emits receipts | deploy with owner approval | dev instance only |
| Q1 2027 | Frontier-model results | same suite on a remote provider | blocked: credential and spend approval |
| Q1 2027 | Learning gain that survives held-out | significant paired improvement | no statistically significant held-out improvement |
| Q1 2027 | Three verticals on customer data | precision on data the team did not write | fixtures only |
| Q1 2027 | First long autonomous missions | 6 h and 24 h runs with model calls | 1 h, no model calls |
| Q2 2027 | Larger real worker populations | 1,000+ processes across hosts | 500 on one host |
| Q2 2027 | Dynamic specialist selection | routing on reliability history | measured router, no history use |
| Q2 2027 | ClawX workforce interface | deployed, tenant-isolated | local store |
| Q3 2027 | Independent witness | a second organization cosigns | none |
| Q3 2027 | Unfamiliar-task evaluation | a public benchmark, disclosed | constructed tasks only |
| Q3 2027 | External pilots | a paid pilot with acceptance criteria met | none |
| Q4 2027 | Independent security assessment | a third-party report | none |
| Q4 2027 | 72 h reliability | a completed 72 h run | none |
| Q4 2027 | Public capability and limitation report | published with evidence | this bundle is the first input |

## What moves the roadmap most, in order

1. A remote model credential with a spend cap (unblocks every intelligence gate).
2. One paid $500 assessment using the vertical prototypes.
3. An independently run witness.
4. Owner-coordinated receipt emission on the hosted MCP path.
