Environmental evidence is the strongest implemented vertical slice.
Public recovery sequence · evidence before claims
A roadmap with acceptance criteria.
This is not a feature wish list. It is the sequence for turning the current FairMind workbench into a defensible assurance system without allowing mock, missing, or insufficient evidence to look complete.
Subject, threshold, artifact, limitation, reviewer, remediation, and rerun.
Changed content under the same run identity must be rejected.
Current state and recovery priority are different things.
The labels below describe the July 2026 recovery decision record. They do not silently promote a planned capability into a shipped one.
Useful evidence paths exist, but the assurance spine is incomplete.
- Environmental governance is the strongest bounded vertical slice.
- Governance records, approvals, risks, remediation, and evidence exist across overlapping models.
- The LLM judge depends on a named external provider; AI BOM is a metadata transformation.
- Workbench evidence records are still mutable and are not yet immutable Evidence Passports.
Make every visible result tell the truth about its evidence.
- Enforce organization scope across systems, evidence, mappings, approvals, and reports.
- Add immutable framework versions, source hashes, evidence-run identities, and reviewed mappings.
- Remove mock-pass and browser-generated results; unavailable means unavailable.
- Make artifacts integrity-aware and produce reports from a pinned canonical manifest.
Close one complete evaluation-to-decision loop.
- Make the assurance router required and consolidate the active governance data path.
- Connect environmental evidence and one real tabular or LLM adapter to a canonical ingestion port.
- Derive readiness from applicability, accepted mappings, freshness, findings, policies, and approvals.
- Preserve compatibility gates until the replacement path proves closure.
Reduce surface area and widen the adapter ecosystem.
- Consolidate the dashboard into six assurance tasks with tested redirects.
- Add artifact adapters for Giskard, MLflow, Fairlearn/AIF360, ModelScan, and garak/PyRIT.
- Connect AI BOM and environmental exports to the same canonical manifest.
- Retire mock provenance, duplicate endpoint families, and placeholder services after closure tests.
The recovery is credible when one real run closes the loop.
Run or import a real evaluation, preserve its identity and artifacts, expose the limitation, route the mapping to an independent reviewer, retain the failure, link remediation and a clean retest, then reproduce the same framework-pinned manifest at the same evidence cutoff.
Missing evidence never becomes a pass.
Bring the evaluation output you already have