# Evolution 18 — Top-100 delta

**The registry is unchanged.** `artifacts/commercial/TOP100_PRODUCT_REGISTRY.json` still lists 100 products; nothing
was added, merged, retired, renumbered or regraded. Evolution 18 built a mechanism that belongs inside products
already listed, and it is not wired to any product's live path, so no readiness grade can move.

## Where the mechanisms land

| Mechanism | Existing product (rank) | What it adds | Maturity of the addition |
|---|---|---|---|
| separate-process sensors, per-effect levels | Epoch 6 — World-State Integrity & Postcondition Reconciliation (78) | observation of an action's local file effects by something other than the executor, with stated gaps | tested and adversarially tested locally; not wired in |
| object-bound authorization, pinned execution | CAIN-45 Agent Hypervisor (21) | the step from "admitted" to "this object, not another" | same |
| typed causal evidence graph | Trust Runtime Kernel (6) | who-says-so and causal-vs-temporal for one action's evidence | same |
| clean-room re-deriving verifier | Independent Verifier (8), E38 Clean-Room Mission Verifier (88) | a verifier for execution receipts that rejects overclaims | runs standalone today |
| adversarial corpus, mutation run | Compound Adversarial Simulator 2.0 (64/65) | 82 executed scenarios, 26 mutants | evaluation method |

Note for the registry's owner: the ranks and names above are the registry's. `artifacts/evolution17/top100_delta.md`
uses product names for ranks 8, 21, 23, 64 and 78 ("CAIN Action Proof", "CAIN Agent Passport", "CAIN Runtime Bridge",
…) that do not appear in the registry file.

## Offerings considered

| Offering in the specification | Decision | Why |
|---|---|---|
| Execution Evidence Verification; Independent Action Receipt Validation | **one thing; merge** into rank 8 / 88 | same verifier, same buyer |
| Agent Execution Integrity Monitoring | **defer** | needs sensors operated apart from the enforcer and outside the user's uid; neither exists |
| Runtime Conformance Testing; Cross-Runtime Security Assessment | **already sellable as a service, no product change** | the corpus and mutation run are a method a consultant can apply to a client's agent tool path (revenue path Tier 1: audits, red-teaming) |
| Enterprise Governance Evidence API | **reject for now** | no hosted path produces these receipts |
| Agent Workflow Incident Reconstruction | **defer** | the graph is per action; reconstruction across actions and services is not built |
| Ongoing Agent Governance Assurance | **reject** | duplicates ARGUS (continuous assurance) |

## The one offer the evidence supports today

**Agent tool-path audit with an effect-observation harness** (a service, under the existing audit / red-team revenue
path, not a new product).

- Customer: a team running an agent that edits files through MCP or function tools (coding agents, ops agents).
- Problem: they do not know whether a tool does only what its description says, or whether an approval survives a
  path swap.
- Deliverable: their tools run under the Evolution 18 harness in a scratch copy of their workspace; a report of
  undeclared reads/writes, postcondition failures and race outcomes per tool, with receipts they can verify themselves
  with the 400-line verifier.
- Deployment: their machine or a VM; unprivileged; Python 3 and `cryptography`.
- Evidence level: what was demonstrated here is E4 (local adversarial tests). Nothing customer-side.
- Price hypothesis: within the existing audit range in `docs/strategy/CAIN42-REVENUE-PATHS.md` ($5k–$25k per
  engagement). Unvalidated: nobody has been asked.
- Cost to deliver: a few days of one engineer per engagement; no infrastructure.
- Liability boundary: findings are about what was observed under one directory on one host during the test; not a
  certification, not an audit opinion, no claim about network or process effects.
- Validation plan: run it on two open-source MCP file servers and publish the findings; offer one pilot to an
  existing contact; count whether anyone pays.

No legal certification, independent audit or production readiness is claimed for any of this.