Finding
arXiv:2608.23623, announced August 26, 2026, evaluates Evidence Carrying Termination for tool using agents. COMPLETE is permitted only when every required claim is bound to valid in scope trace evidence and deterministic closed replay reconstructs the claimed value.
The originating team reports 0 of 66 premature unsupported completions for ECT versus 40 of 66 for its faithful controller in a fresh held out study, with supported completion remaining within the paper's stated noninferiority margin. These results are not yet independently reproduced in Dream Machine.
RuV mapping
Reusable across Dream Machine, MetaHarness, Ruflo, RVM, RVF, and Core Memory.
Implementation
Branch sota/evidence-carrying-termination adds the primitive to @dream-machine/witness with deterministic canonical hashing, trace binding, claim scope validation, closed replay, duplicate evidence rejection, trace mutation detection, tests, ADR 0004, dependency and security review, benchmark plan, and rollback plan.
Required validation before merge recommendation
- Full repository CI is green on supported Node versions.
- Deterministic fault matrix produces zero unsupported COMPLETE decisions.
- Matched baseline versus ECT benchmark reports supported completion, unsafe completion, recovery, latency, memory, trace size, and failures.
- No more than 2 percentage point decrease in clean supported completion.
- Security review confirms evidence labels cannot create authority.
- Human review remains required. No autonomous merge or deployment.
Key limitation
ECT certifies support in a recorded trace, not external truth. A compromised tool can emit a false value that closed replay faithfully reconstructs. Production use therefore requires independent evaluators plus RVM or equivalent authority enforcement and signed provenance where identity matters.
Finding
arXiv:2608.23623, announced August 26, 2026, evaluates Evidence Carrying Termination for tool using agents. COMPLETE is permitted only when every required claim is bound to valid in scope trace evidence and deterministic closed replay reconstructs the claimed value.
The originating team reports 0 of 66 premature unsupported completions for ECT versus 40 of 66 for its faithful controller in a fresh held out study, with supported completion remaining within the paper's stated noninferiority margin. These results are not yet independently reproduced in Dream Machine.
RuV mapping
Reusable across Dream Machine, MetaHarness, Ruflo, RVM, RVF, and Core Memory.
Implementation
Branch
sota/evidence-carrying-terminationadds the primitive to@dream-machine/witnesswith deterministic canonical hashing, trace binding, claim scope validation, closed replay, duplicate evidence rejection, trace mutation detection, tests, ADR 0004, dependency and security review, benchmark plan, and rollback plan.Required validation before merge recommendation
Key limitation
ECT certifies support in a recorded trace, not external truth. A compromised tool can emit a false value that closed replay faithfully reconstructs. Production use therefore requires independent evaluators plus RVM or equivalent authority enforcement and signed provenance where identity matters.