Arbiter 27B starts from Qwen3.8-27B and adds recurrent depth: three windows of the network run three passes over the prompt, with small per-pass adapters trained on operations a verifier can check exactly. The base weights never change. Measured on sealed sets: 16-hop graph reasoning 31.5% for the base model, 86.3% for Arbiter; proposing, repairing, eliminating and verifying rules at 96, 90, 78 and 86 percent; as an autonomous research agent on the Erdős–Straus conjecture, exact coverage of hard primes from 39.9% to 91.2% in 80 turns, every rule machine-verified. It is the included Arbiter route in Orrery, with hosted fallback when capacity is busy.
The workspace, selected route, and review requirement stay explicit before hosted work begins.
Request received · boundary visibleArbiter is a subscriber-only hosted route. Capacity, latency, quotas, and availability can change as capacity is added.
Route named · availability not assumedSummary, usage, proof, and boundary records remain distinct so a polished answer cannot masquerade as verification.
Evidence visible · success not fabricatedThe receipt gives you material to inspect. It does not decide whether the work is safe to keep, run, or ship.
Receipt presented · acceptance remains yoursArbiter is a hosted Orrery route for coding and research work where evidence matters. The experience is simple: ask Orrery to run work, review what happened, see the usage it consumed, and decide whether to trust the result.
Inside Orrery, Arbiter appears as a hosted model route named arbiter-flash-27b — Arbiter 27B, the looped Qwen3.8-27B. It can be selected from Nexus alongside Doubleword hosted routes. Subscribers see usage meters, run receipts, and proof/boundary summaries where the app supports them.
The route is designed for agentic coding and research workflows: planning, editing, follow-up fixes, and review. It does not replace your own testing, code review, security review, or backups.
Most models do their thinking out loud: the visible chain of thought is generated token by token, and every one of those tokens is billed. Arbiter thinks before it writes. Three windows of the network run three passes over the prompt during prefill, with small per-pass adapters trained on operations a verifier can check exactly; decode is then a single ordinary pass. The base Qwen3.8-27B weights never change — the loop is where the extra depth lives.
Two practical consequences for a subscriber. The reasoning is not a transcript you pay for by the token, so an Arbiter run is metered as compute rather than as reasoning tokens. And because the extra depth is in the prefill, the answer arrives as one ordinary completion: Nexus shows the prompt, route, run summary, usage, proof, and boundary receipts so you inspect the result instead of trusting a raw completion.
Arbiter runs on Ephemerent's own hardware and on a cloud lane, quantized on both. Every request is the same shape wherever it lands: three prefill passes through the loop windows, then a single-pass decode.
| Lane | Hardware | Precision |
|---|---|---|
| On premises | NVIDIA DGX Spark | NVFP4 |
| Cloud | AWS L40S | FP8 |
Both lanes answer to the same route id, and nothing about the loop or the review boundary changes between them.
Every number is a paired, same-day measurement on sealed sets built to be disjoint from training, with an executor or an exact verifier as the judge. Against the base Qwen3.8-27B: 16-hop graph reasoning 31.5% to 86.3% (n=600). The trained operations, versus the first served version: proposing a rule after refutation 78 to 96, repairing a refuted rule 72 to 90, eliminating candidates 52 to 78, denoising evidence 75 to 89; verifying a claim against evidence 86. Full held-out SuperGPQA (n=16,400, transfer across datasets): 36.25% to 45.48%, p=5e-79. As an autonomous research agent on the Erdős–Straus conjecture: exact coverage of hard primes from 39.9% to 91.2% in 80 turns, every rule machine-verified. The passes are the mechanism: with the pass adapters unreachable, accuracy drops 2.1 points (p=0.017); one adapter shared across passes is indistinguishable from a single pass.
Not measured yet, and therefore not claimed: Terminal-Bench 4.0 and Terminal-Bench Science, which are queued on the served model.
There is one way in: an Orrery subscription. Install Orrery, sign in, and select arbiter-flash-27b from Nexus for a workspace or a task. Every paid plan includes a monthly pool of Arbiter compute credits — 14M on Pro, 35M on Max, 70M on Ultra — alongside a separate Doubleword pool. Signed-out users, unsubscribed accounts, and canceled or unpaid accounts cannot use hosted Arbiter credits.
One credit is one millionth of a dollar of provider cost. Arbiter credits buy looped inference compute — the three prefill passes plus decode — rather than billed reasoning tokens, which is why the pool is quoted in compute and not in tokens. Pools reset monthly, stop hard at the limit, and never overage. Full plan detail is on Orrery pricing.
The exact capacity, latency, quotas, model mix, and availability may change as capacity is added.
By default, Orrery shares your Arbiter coding and research sessions with the Arbiter service so the next version is trained on this kind of work: the prompt and tool results sent to the model, the response, and the outcome of the turn, tied to an anonymous per-session id. Your account and email are never attached, and Doubleword turns are never captured. Turn on Zero data retention for Arbiter in Orrery settings and the service keeps counts and an access-log line only, from the next request on. The full description is in the Privacy Policy.
Treat Arbiter output like any other AI-generated code: inspect it, run tests, review security implications, and keep backups. Orrery can help organize evidence, but you remain responsible for what you accept, run, and ship.