feat(trajectories): promote verifiable agent computations from experimental to usable

[mnemonik-dev]

Goal

Turn the existing feature-gated step, verdict, trajectory, mandate, evidence, and compliance work into a coherent, usable roadmap workstream for verifiable agent computations.

This issue does not restart schema design. It audits and completes the existing implementations and specifications under work/verifiable-trajectories/ and work/agent-provenance/.

Required outcome

A producer can client-sign ordered computation steps containing declared tool inputs/outputs and results; independent evaluators can sign verdicts; a verifier can reconstruct the trajectory, validate signatures and lineage, recompute its order-preserving root, and assess correspondence to a signed mandate.

Work

• Inventory implemented versus specified behavior behind trajectory-experimental.
• Freeze privacy/redaction rules for prompts, tool arguments/results, environment data, and secrets.
• Require client signing for producers and distinct evaluator identities for verdicts.
• Define runtime adapter boundaries so Codex, Claude Code, Hermes, A2A, and other agents normalize to the same fixtures.
• Complete chain-based reconstruction without requiring the historical SQL store.
• Publish golden vectors and cross-runtime verification.
• Define honest trust statements: cryptographic authenticity/correspondence is not automatic semantic truth.
• Establish performance, batching, payment, and selective-anchoring policy.
• Define the GA gate; keep the feature experimental until it passes.

Acceptance criteria

• An implementation/status audit maps every current task and deviation.
• Canonical schemas cover step, tool input/output digest, result, verdict, trajectory root, mandate, evidence, and compliance.
• Producer and evaluator signing ownership is enforced.
• Secret redaction and privacy policy run before canonicalization.
• A trajectory reconstructs and verifies from portable anchored artifacts without SQL history.
• Golden vectors verify byte-for-byte in Rust and TypeScript.
• At least two agent runtimes emit conformant fixtures.
• Tampered order, missing step, altered tool output, self-verdict, foreign mandate, and replay fail deterministically.
• UX explains what is and is not proven.
• GA criteria and feature-flag removal decision are documented.

Related

• MCP-to-MCP delegation attestation (MCP_DELEGATION_V1) #65 — MCP delegation attestation
• ERC-8004 Phase 2 — overall integration plan (umbrella) #69–erc8004-5: Ethereum anchor end-to-end #75 — ERC-8004 integration
• Research: integrating Mnemonic with decentralized language model (DeLM) frameworks #169 — DeLM research
• work/verifiable-trajectories/
• work/agent-provenance/