Skip to content

Research / CAIRN-DAG-01: dynamic Fact/Intent exploration vs 007 baseline (kill-first) #160

Open

Description

@PhysShell

Classification / decision

P3 / RESEARCH / REUSE-FIRST / KILL-FIRST / PARKED / NOT ADOPTED / NO IMPLEMENTATION AUTHORIZED.

Source: oritera/Cairn, protocol, dispatcher, paper. Full source-backed research page.

Question: On genuinely long, branched engineering investigations, can dynamic fact/intent-driven branch selection improve time to an independently verified result relative to 007's existing and simpler fixed-branch workflow, without propagating unsupported claims as verified facts or creating a new orchestration stack?

This is not approval to run Cairn, create a new agent framework, start autonomous pentesting, or unpark 007. It is a bounded research candidate.

Source-confirmed mechanism and limitations

  • Cairn's server maintains a Fact/Intent/Hint DAG and claims/heartbeats/release; Dispatcher is the protocol writer and assigns Bootstrap / Reason / Explore to Claude Code/Codex/Pi workers. Fact is basically id + description; a DAG edge establishes referential/causal lineage, not truth. Protocol, server models, Reason prompt.
  • Server checks ID references and state/lease constraints, but there is no general independent oracle for Fact.description or complete. A syntactically valid completion must not promote a claim to VERIFIED.
  • Design documents state one Dispatcher instance is the intended/tested mode. Current claim status is not complete worker history; heartbeat/lease expiry does not establish idempotency or exactly-once side effects.
  • Local mode runs agents under current user without sandbox. Container config with Docker socket/host network must not be assumed safe for untrusted work.
  • Preprint, 26 Sep 2026 reports 76.5% faster solutions and up to 3.08x speedup on high-effort (≥1M token) tasks; overhead with no gain on low-effort tasks. Author-reported, not an independent replication or evidence of 007 applicability. Repository's 54/54 Tencent competition claim is not an engineering benchmark.
  • Repository code: AGPL-3.0 / separate commercial licensing option in README. Architectural study only; no code transfer.

Ownership / do not create duplicate subsystems

  • #42 owns action identities, verifier results, measured costs, replay-safe side effects.
  • #43 owns persisted run events, deterministic reducer and fail-closed semantics.
  • #46 owns source snapshots, evidence spans, atomic claims, contradictions and citations. Planner must consume this evidence, not define a second trusted Fact store.
  • #57 owns provider-neutral agent loop, typed tool results, observer boundary.
  • #74 owns change admission, permissions, independent verifier gates and final verdict.
  • #159 already considers durable agent orchestration/Memory V2, not dynamic research branch-selection.
  • This issue owns only dynamic planning-policy hypothesis, its bounded A/B/C evaluation and measured incremental value. It can be closed as DUPLICATE/NO-CHANGE after inventory.

Gate 0 — read-only inventory and kill-first

Gate 1 — synthetic A/B/C test design (requires separate GO)

Same authorized, frozen inputs; same model/provider, prompts except planner seam, tool/capability/network policy, context budget, reproducibility settings and time/token limits. Record NON-COMPARABLE whenever parity is unattainable.

  • A: existing 007/single-agent or existing standard workflow.
  • B: same baseline with fixed independent parallel research directions (if it exists; otherwise explicitly omit).
  • C: minimal external Fact/Intent planning overlay and deterministic scheduler adapter, not the full Cairn stack.

Suggested cases: short/simple control; complex C#/Go source investigation with known answer; cross-document contradiction or stale SHA; injected unsupported Fact and conflicting hints; interrupted/expired claim with attempted tool side effect. Use only synthetic/open authorized code and inert actions; never production/client source, credentials, DTs, real vulnerabilities or live targets.

Independent negative controls: invalid Fact provenance, stale source revision, contradictory evidence, hallucinated success, malicious Hint/prompt injection, duplicated intent after lease expiry, attempted side-effect retry, oversized graph and timeout/unknown oracle. A model's complete is advisory only; no VERIFIED unless evidence and external gate pass.

Measure / acceptance

  • Task success using external oracle, time-to-first verified solution, wall time and tokens/cost (unknown ≠ zero).
  • New independent useful branches, duplicated/wasted branches, FP claims, propagation of unsupported Fact, contradiction-handling and reviewer burden.
  • Context inflation as graph grows, reasoner overhead on short tasks, stop/cancel/resume behavior, duplicated side effects and policy bypass attempts.
  • All evidence tied to exact input/commit/config/model/verifier hashes with replayable outputs. Negative controls RED on broken assumptions, not false PASS.
  • GO-NARROW only for independently demonstrated incremental gain on nontrivial work with no lost trust invariants, acceptable overhead and a clear minimal seam. KEEP IDEAS/PARK for useful DAG visualization without net benefit. KILL if baseline is as good/better, if false claims leak into VERIFIED, or if complexity explodes.
  • If no current 007 gap, close as NO-CHANGE with a short cross-reference, not an implementation PR.

Hard stop

No dependency install, container spin-up, autonomous agent runs, untrusted repository execution, real-world pentesting, privilege/network changes, CI gating, push/merge/release, refactor of 007, new persisted schema or automatic conversion of model-generated facts into policy/approved knowledge. This issue alone does not authorize any implementation.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

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