Treebeard 0.2.0 takes the StateTree runtime line public. The runtime moves from the 0.1 line (b9624) to the gated RC8 line (b9743-c7091b65b): sequence-ragged KV attention plus state-I/O fusion (shared-prefix forks, ragged batching), and the live PCBT proof-carrying branch transaction surface (pcbt.receipt.v1 receipts, /transactions, /metrics counters).
Install (unchanged command):
curl -fsSL https://raw.githubusercontent.com/newjordan/treebeard/main/install.sh | bash
~/.local/bin/treebeard doctor
~/.local/bin/treebeard serveValidation (re-benched 2026-07-17/18 on the release bytes)
| Gate | Verdict |
|---|---|
| Agent Bench, Intel Arc Pro B70 | PASS — 94/100, all 69 verdicts identical to the 0.1 line |
| Agent Bench, NVIDIA GB10 (cross-host) | PASS — 94/100, exact verdict-vector match with B70 |
| Held-out probe pack v1.1 (24 scenarios, measurement-only) | PASS — candidate == 0.1 line, 43/46, zero behavior diffs |
| Seed sweep (42 / 1709 / 7) | PASS — identical output vectors |
| PCBT-12 wave benchmark | 97.8% vs 53.6%/63.6% single-pass (two arms, 30 tasks) |
| Deep-task quality gates (deep9 + deep4k, width 8) | 0 regret under the TQG-5 selector |
| Portable CPU runtime (ubuntu 22.04 container) | PASS — build + smoke, token parity with the GPU line |
| GB10 CUDA correctness | PASS — 1900/1900 backend ops; StateTree fragmented-restore smoke PASS |
GB10 token generation: investigated regression, cleared as host drift
A first cross-check measured tg128 6.8% below the 0.1 line on the GB10. A three-point bisect (0.1-line build / upstream merge-base / release build, interleaved ABA x5) shows all three at parity: tg128 ~52.5–53.8 tok/s and pp4096 ~2,020 tok/s within the host's ±5% run-to-run drift. The July GB10 numbers previously quoted on this page were measured with a co-resident process pinning clocks and are not reproducible on the idle host; the site and package manifest now carry the drift-controlled values. No Treebeard or upstream code regression exists on this path.
Wave selector (TQG-5)
The self-consistency selector now counts one vote per distinct prompt variant: at temperatures of 0, repeated diversity suffixes re-decode identical branches and previously counted as independent votes, which let a single prompt variant outvote six correct siblings on one deep-task gate (regret 0.111). The wave now overrules the single-pass baseline only on replicated consensus across independent variants, otherwise it abstains to the baseline — worst-case regret is bounded by construction. Confirmed on the release bytes: regret 0 on both deep-task gates. Trade-off, recorded honestly: a correct answer carried by a single branch (7 of 8 siblings stalled) no longer overrides the baseline.
Package (pkg5)
- SYCL (Intel, oneAPI 2026):
treebeard-0.2.0-rc.1-c7091b65b-pkg1-sycl-oneapi2026-linux-x86_64.tar.gz— sha256d844ff9a21a60f7f3283d396bf4a83ef092eaf4ffec781a40d830a17e27e3358 - CPU (portable x86_64, glibc 2.35+):
treebeard-0.2.0-rc.1-c7091b65b-pkg1-cpu-ubuntu22.04-linux-x86_64.tar.gz— sha256e91174c51f547344e35e5569542346fce35052b3dc213cf152abbfd45a777df5 - CUDA (NVIDIA GB10 ARM64, CUDA 13.3):
treebeard-0.2.0-rc.1-c7091b65b-pkg1-cuda13.3-linux-aarch64.tar.gz— sha256e7221ed3e1e61d0158fb58d6c9152cd1217d0eecb6ad0779c81f3bc5b1f3fdc0 - NVIDIA Blackwell Q8_0 12-column kernel: 33.0% on its designed workload (re-validated on the release bytes), MoE down 3.4%
- Launcher controls from 0.1.1 unchanged:
TREEBEARD_REASONING(off by default, matching the validated benchmark),TREEBEARD_SPECULATION(opt-in, still not claimed as a speedup)
Notes
- 0.1.1 installations are not auto-upgraded; rerun the installer for 0.2.0. The model file is reused by checksum, so the upgrade downloads only the runtime (~50 MB SYCL / ~11 MB CPU).
- The wide wave-bench sweep (R4+) completed below its preregistered claim bar and is deliberately not claimed here; it is recorded in the private evidence trail.
- Evidence index: https://github.com/newjordan/treebeard/tree/main/results