# Regenold governed workflow — v1 pilot contract

## The bounded result

EcoKure will sit around one AntifragileAI/Regenold compliance workflow for 30
days in shadow mode. AntifragileAI continues to propose an answer. EcoKure
checks the declared control pack, returns `PASS`, `BLOCK` or `ABSTAIN/REVIEW`,
and preserves a signed EvidencePack for a named customer reviewer.

The pilot is successful when the customer can see which answer was proposed,
which rule and source version was used, why the gate reached its verdict, and
what must be revisited when one source or rule changes.

## One workflow and one owner

- **Workflow:** AI Act compliance answer review for one agreed Regenold use case.
- **Owner:** Customer AI governance or compliance owner, named in the kickoff record.
- **Boundary:** Customer-approved staging account or private environment; synthetic or customer-approved shadow data only.
- **Mode:** No production actuation. A held decision waits for the named reviewer.
- **Reviewers:** Customer risk owner plus technical assurance reviewer. AAF can provide the OIDC identity route for the reviewer; it is not a validator.

## Control pack baseline

`regenold-ai-act-v0.1` covers Article 6 classification, Article 50
transparency, Article 4 AI literacy and evidence lineage. The baseline freezes
the model identifier, prompt or request shape, policy, source references,
control versions and expected outcomes. The customer approves the pack before
the first shadow run.

## Demonstration and pilot checks

The synthetic walkthrough has three deliberately different outcomes: a
classified request passes, a missing transparency notice is blocked, and an
incomplete accessibility interpretation abstains for review. During the pilot
we introduce one controlled Article 50 source revision. EcoKure classifies the
dependency graph, replays only the affected replayable request and leaves the
uncertain interpretation in the review queue.

The acceptance record measures decision coverage, false pass/block and
abstention rates against the customer's labelled sample, replay correctness,
reviewer time, p95 latency, evidence completeness, independent verification
and cross-tenant isolation. The customer records `GO`, `GO WITH LIMITATIONS`
or `NO-GO`.

## Roles after the pilot

AntifragileAI owns the proposed compliance workflow and its domain content.
EcoKure owns the assurance runtime, control-pack versioning, evidence chain,
change impact and replay tooling. The customer owns its data boundary, keys,
reviewers, acceptance decision and production operations. A second phase can
add DCLA/DELA dependency graphs, subgroup monitoring and vigilance routing
only after the first workflow has earned that scope.

## Current evidence and limits

The AWS customer-owned staging smoke demonstrates runtime identity, KMS
signing, PostgreSQL persistence, recovery, headers and one synthetic request;
it is not customer production validation. The independent local reproduction
is ready for a second reviewer. The image scan still has six critical and
twelve high findings to remediate or explicitly accept before production.

This contract is a pilot offer, not a Regenold commitment, legal advice,
regulatory approval or certification. No email or partner invitation is sent
by publishing this document.