Skip to content

Cairn v1.0.0

Latest

Choose a tag to compare

@STiFLeR7 STiFLeR7 released this 21 Oct 04:54
· 3 commits to master since this release
9a786c9

Cairn v1.0.0 — release notes

Released 2026-10-21 (owner approval: AP-0036). Cairn is a framework-agnostic reference implementation and
benchmark for recoverable long-horizon AI agents, built around the thesis "Checkpoints Are Compactions."
v1.0.0 follows the pre-registered live recovery gate (M5), which returned GO.

Highlights

  • BYOM recovery library (M4): the Re-grounding Recovery (RGR) protocol as public functions over World,
    CheckpointStore and EffectLedger seams, plus an opt-in Agent loop and reference implementations. You
    bring your own agent loop, model and keys.
  • P1–P6 proof ladder: P1 crash/restart; P2 compaction continuity (Continuation Contract v0); P3 external
    effects (Receipt/Reconciliation Contract v0); P4 Claude Code host integration; P5 two-host portability
    (Claude Code and OpenHands SDK 1.42.1); P6 independent conformance, admitted 2026-10-19 (Haiku-only
    clean-room candidate 10: Stage 6A 30/30, sealed holdout 30/30, uninvolved reproducer 30/30).
  • M5 v1.0 gate: GO (pre-registered before any live call; commit e1b4280;
    gate-verdict.json: "C1 (capability-matched) and C3 supported on both
    models"). Per model, Claude claude-sonnet-5-5 and Codex gpt-6-sol returned identical numbers:
    • calibration 10/10; 32/32 planned cells fired (B0 16, B3 16), 0 skipped, 0 lost; 520 calls, 0 errors;
    • C1 (capability-matched): B3 success 1.00 vs B0 1.00; recovery tax B3 mean 4.50 ± 0.50 (max 5.00) vs B0
      8.00 ± 0.00 (min 8.00); no-regression B3 0.42 vs B0 0.00;
    • C3: 0 duplicate effects with B3 vs 16 with B0.

Scope and limitations

Honest scope of the M5 result (claims registry, 2026-10-21):

  1. Both models were fully competent on the task, so the capability-matched rule was never stressed.
  2. The chain task's oracle advances one step per call, so recovery cost in work units is determined by task
    structure for a competent model. The evidence shows RGR's recovery mechanism cuts recovery tax (5 vs 8 work
    units at most/least) and preserves pre-crash work with near-zero variance; it is not a broad
    live-performance or model-robustness claim.
  3. C3's torn effect is injected by the harness at the crash step, so C3-live is "effect safety under live
    control flow", not model-dependent evidence.
  4. C2, C4 and C5 remain reference-harness only.
  5. One task (chain n=6), crash steps 2 and 3, 8 repeats, two models.
  6. Not an ecosystem standard; not exactly-once delivery.

P-ladder boundaries (README "Project status"):

  • No exactly-once delivery of external effects; P3 admits only the decision semantics proven for its reference
    effect (absent → retry; matching present → skip; unknown, mismatch, never-retry → escalate).
  • Not an ecosystem standard and not universal compatibility; P4 admits one host integration, P5 deterministic
    two-host portability, P6 one passing external chain.
  • P2 requires a fresh process and a clean, verified checkpoint; neither v0 contract establishes broad
    live-model performance or a framework API.
  • C2, C4 and C5 are reference-harness evidence only.

Reproduce

  • M5 evidence: results/m5/ (preregistration.json, protocol.json, claude.json,
    codex.json, gate-verdict.json); design: the
    M5 spec; re-run with python benchmarks/m5_study.py
    (needs the claude and codex CLIs). The pre-registration pins code at 168a6e6; verify the evidence at
    commit e1b4280.
  • Reference harness: python -m pytest -q; see REPRODUCE.md.

Install / cite

  • Apache-2.0. Python 3.12+, no third-party runtime dependencies.
  • Cite via CITATION.cff (version 1.0.0, released 2026-10-21).