METTLE 0.5.0

[0.5.0]

Credential schema

Credential schema 1.1 and suite policy 2026-08-14 remain unchanged. This
release changes the agent control surface and does not alter credential scoring,
tier eligibility, signature verification, or portable credential acceptance.

Suite policy

Challenge generation, scoring, and tier semantics remain unchanged from 0.4.8.
The MCP adapter can now complete authenticated multi-round novel-reasoning
sessions through explicit round submission and feedback operations.

Public key changes

No signing key, discovery format, or verification-key rotation is included.
Production key identity remains independently verifiable through the published
discovery surface and deployment receipts.

Compatibility

• Publish mettle-control-v1 output schemas, structured content, typed session
  snapshots, valid next actions, mutation receipts, and bounded structured errors.
• Add session inspection, authenticated cancellation, novel-reasoning round
  submission, and completed-round feedback. The original seven MCP tool names
  remain available, giving eleven reviewed public tools in total.
• Preserve concise text fallback for older MCP hosts and keep the quick-session
  bearer in a caller-isolated capability vault. Repeat result reads remain
  available for the capability lifetime.
• Preserve the prohibition on an automatic challenge solver. Callers still
  supply their own answers.
• Continue Python 3.10 through 3.14 compatibility.

Safety and recovery

• Map upstream failures to fixed, non-reflective error codes without returning
  raw response bodies, exception strings, participant content, or credentials.
• Distinguish repeatable reads from mutations with ambiguous commit state.
  Session mutations require refresh before retry; ambiguous resource creation
  requires operator action.
• Validate session, suite, challenge, and round identifiers before constructing
  credentialed request paths.
• Return post-action state without introducing a follow-up read whose failure
  could misrepresent a successfully committed mutation.

Known limitations

METTLE produces probabilistic behavioral evidence. A passing result does not
prove identity, substrate, consciousness, freedom, agency, safety, governance,
operator identity, or authorization suitability. Relays, source-aware solvers,
model-assisted humans, evaluator error, and imitation remain possible. Human
accessibility, rights-cleared fairness evaluation, independent protocol and
cryptographic review, independent agent-control review, and destructive recovery
drills retain their separate evidence and authority requirements. Later control
plans, revisions, idempotency keys, budgets, and cross-transport semantic
generation remain roadmap work.