Skip to content

SOTA: anchored trace replay for causal multi-agent failure reproduction #41

Description

@ruvnet

Finding

SymTrace, arXiv:2608.25920, submitted 2026-08-26, shows that whole-trajectory reruns confound causal repair with sampling luck. It records execution traces, establishes intervention anchors, reconstructs the exact prefix during replay, and regenerates only the downstream suffix. The originating team reports 67.97% failure reproduction and 6.90% repair for unguided reruns across three MAS frameworks. Symptom-driven anchored intervention repaired 20.15% of failures on the 536-trajectory SymFail dataset.

Evidence class: originating-team report with a human-annotated dataset, not yet independently reproduced by RuV.

Weighted Opportunity Score: 4.60/5.

RuV gap

Dream Machine already has deterministic witnesses, held-out promotion gates, trace-driven optimization work, and bounded mutation plans. Recuris experiment #36 localizes memory changes, but there is no reusable primitive that proves an experiment restarted from the same causal prefix before comparing a repair.

Without this invariant, a candidate can appear to repair a failure simply because the model resampled a different upstream trajectory.

Proposed primitive

Add AnchoredTraceReplay to @dream-machine/witness:

  1. canonicalize and hash the immutable trace prefix
  2. bind an intervention anchor to an event id and step
  3. verify a replay prefix before any downstream regeneration
  4. reject duplicate ids, non-monotonic steps, malformed anchors, prefix mutation, and suffix collisions
  5. produce a deterministic replay receipt containing prefix digest, anchor digest, regenerated suffix digest, and combined trace digest

The digest is evidence of trace identity, not execution authority.

MetaHarness roles

Use independent research, implementation, review, security, testing, reproducibility, and release roles. The reviewer and reproducibility roles must not share the implementation agent's generated repair rationale.

Benchmark

Compare under identical model, temperature, seeds, tool versions, task set, and evaluator:

  1. whole trajectory rerun
  2. anchored replay with no intervention
  3. anchored replay plus symptom-driven intervention

Report reproduction rate, repair rate, false-repair rate, wall time, tokens, API cost, prefix verification failures, variance, and per-failure-category results.

Security and falsification

Reject the primitive if replay can proceed after any prefix mutation, if regenerated events collide with immutable event ids, if the anchor can move without changing the receipt, or if apparent repair gains disappear after controlling the prefix.

Acceptance

The infrastructure implementation may land once deterministic tests and CI pass. Do not claim SymTrace-level repair gains until an independent matched reproduction shows at least 95% exact prefix verification, materially higher failure reproduction than whole reruns, no authority expansion, and zero acceptance of mutated prefixes.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions