Morphz Agent Trajectory Reference Implementation Verification v0.1

Status: Draft implementation evidence

Steward: Newvar

Last updated: 2026-08-21

Chinese translation: zh-CN

1. Purpose

This document records executable evidence for the Morphz reference implementation of the Agent Trajectory Specification v0.1. It is not a conformance certificate. The Event Store remains authoritative; export, verification, Reward interpretation, and Episode derivation never rewrite source execution facts.

The implementation closes one practical loop:

authoritative Event History
  -> bounded Agent Trajectory export
  -> structural and integrity verification
  -> immutable Verifier Result
  -> separate immutable Reward Record
  -> permission-checked Training Episode

2. Implemented surfaces

Surface	Reference implementation behavior	Evidence
Deterministic export	Selects a bounded Context/Object/Activation scope through indexed Event queries; orders by sequence, time, and identity; emits stable Bundle identity	exporter_preserves_causality_redacts_secrets_and_seals_integrity
Causal projection	Emits typed edges for declared parent fields, preserves out-of-scope parents, and derives ordered Plan-effect edges	exporter and verifier tests in trajectory::tests
State boundaries	Projects Context before/after/snapshot revisions as stable State references and validates every Node-State reference	exporter and structural verifier tests
Disclosure and rights	Recursively redacts credential-shaped fields, redacts user content by default, records omissions, and denies training by default	exporter and permission tests
Integrity and untrusted input	Seals Bundles with a declared SHA-256 serialization digest; verifies identity uniqueness, cross-references, causal acyclicity, scope consistency, and digest without executing payloads	verifier_rejects_tampering_and_causal_cycles and cross-reference tests
Verifier Result	Commits a deterministic immutable Event only after each Evidence reference is found in the same Context; exact replay is idempotent	trajectory_verifier_and_reward_facts_are_durable_idempotent_and_exportable
Reward Record	Commits a separate deterministic interpretation whose sources must be existing Outcome, Verifier Result, or Reward Record facts in the same Context	Runtime integration test and training-loop test
Training Episode	Requires both the AT-Training Profile and explicit rights.training=true; emits explicit model-input, supervised-target, environment-output, and loss-mask roles	verifier_reward_and_training_episode_form_a_permissioned_loop
Administrative API	Exposes Bundle export/validation, fact commit, Episode derivation, and pure Episode validation through the Rust SDK	SDK compile and library tests
Operator interface	Exposes trajectory export, trajectory verify, and trajectory episode through the CLI	trajectory_commands_preserve_scope_rights_and_input_file

The current reference JSON forms are described by:

• Agent Trajectory Bundle schema;
• Training Episode schema.

3. Recovery and authority properties

• Verifier and Reward identities are content-derived. Repeating the same commit returns the existing Event; an identity occupied by different content is rejected.
• Verifier Evidence and Reward sources are resolved through Context-scoped Event queries before persistence.
• Reward Records never mutate Outcome or Verifier facts and cannot silently become source truth.
• Bundle verification is pure and never executes embedded payloads, follows external references, restores capabilities, or writes Runtime state.
• Episode derivation is refused when the Bundle lacks the Training Profile or explicit training permission.

4. Reproducible gates

The focused evidence can be reproduced with:

cargo test -p morphz trajectory --lib --offline -- --nocapture
cargo test -p morphz typed_context_proposal_commits_once_and_recovers_the_commit_window --lib --offline
cargo test -p morphz objective_wait_proposal_uses_authority_and_replays_without_a_second_transition --lib --offline
cargo test -p morphz objective_completion_proposal_consumes_committed_outcome_and_replays_intent --lib --offline
cargo test -p morphz trajectory_verifier_and_reward_facts_are_durable_idempotent_and_exportable --lib --offline

Repository release gates additionally run the complete Yao and Morphz unit/integration suites, format checks, Clippy, JSON parsing of the reference schemas, and git diff --check.

5. Current limits

• The exporter preserves external-parent declarations but does not recursively fetch an unbounded causal closure outside the requested selection.
• Context State is currently exported primarily by exact version reference and optional delta, not by automatically disclosing a complete Context snapshot.
• AT-Evaluation environment and model bindings are best-effort projections from represented facts; a deployment must declare unavailable bindings instead of overstating reproducibility.
• The current integrity Profile is a declared deterministic digest, not a canonical signature or proof that represented Outcomes are true.
• Dataset sharding, consent revocation workflows, trainer-specific adapters, independent implementation interoperability, and a normative conformance suite remain future work.

These limits keep the current claim narrow: Morphz provides a tested reference pipeline for portable structured experience and permissioned Episode derivation, not complete Agent Trajectory v0.1 conformance.