Horopter

Horopter: the set of points that both eyes see as one.

Code: MIT (LICENSE). Weights, data recipe and docs: CC BY-NC-SA 4.0 (research, non-commercial). Weights: horopter-0.4B, horopter-1B (first release, 2026-10-20).

Horopter is an encoder-only decision model. Give it a state (a document, a record or a JSON object) and typed questions about it (yes/no, choose one, score on a scale). It answers them in one pass, and for every answer it also returns:

• evidence: whole-word spans in the state that support or refute each option;
• relation: whether the state bears on the question at all, and with what stance (support / refute / undetermined);
• abstention: a calibrated probability that the answer is wrong, with a reason (unrelated, model_uncertainty).

The state is encoded once, without seeing any question, and cached. Each question reads that cached state, so:

• adding a question never changes the answers to the others;
• asking many questions about one state costs about one encoding.

Built on Ettin encoders (400M and 1B), from Fight Fire with Fire Robotics.

Status

First release, 2026-10-20: horopter-0.4B (horopter-0.4b-20261020) and horopter-1B (horopter-1b-20261020). Router + adapters, one pass:

The full tables and limits are in the model cards. Grounding and relation are level with 4–9B decoders. Multi-step policy decisions on real text trail them by about 20 points. This release has rule-application data with always-determined answers, so it rarely says "more information needed" on real rule text.

Layout

horopter/     model, packing, losses, runtime (adapters), evidence, abstention, fields (typed extraction), api + server (wire format)
horopter/fast/  fast path: state cache, CUDA graphs, block-sparse state attention, all adapters in one pass
training/     joint trainer, adapter / evidence trainers, abstention head fit, data builders, recipes/beta2.md
eval/         benchmark scorers (E37 bench, E40 rule evals) and the HTTP client
scripts/      export_hf.py: training checkpoint -> HF layout
tests/        parity against the research code; fast path against the release path
docs/         release plan, porting notes (docs/PORTING.md), fast path (docs/FAST_PATH.md)
assets/       theme art (public domain)

The same code runs against a server or a model in your own process, and gets the same answers:

import horopter
h = horopter.connect("http://127.0.0.1:8420")     # a running server; or:
h = horopter.local("jmccardle/horopter-0.4B")      # the model in this process (first use downloads ~0.8 GB)
r = h.decide("Refunds within 30 days. The lamp was bought 12 days ago.",
             questions={"refund": {"type": "noul", "instructions": "Is a refund allowed?"}},
             fields={"days": {"type": "duration", "unit": "days", "instructions": "How long ago it was bought"}},
             evidence="chosen")
r["answers"]["refund"]["noul"], r["answers"]["days"]["value"]

Server: python -m horopter.server --model jmccardle/horopter-0.4B --port 8420 --fast (add --device cpu on a machine without a GPU). It speaks Jev's System One format (/v1/systemone: typed questions) and Cloudflare's Clairic format (/v1/clairic: typed field extraction), with both in one request at /v1/horopter. Walkthrough: docs/API.md; spec: docs/openapi.json.

--fast serves the same answers (up to bf16 rounding) 3–19× faster on a GPU, and caches states on a CPU: a state asked about again costs only its question tokens. Measurements, settings and memory are in docs/FAST_PATH.md.

Credits

Title page of Opticae thesaurus (Friedrich Risner ed., Basel 1572), Bayerische Staatsbibliothek; public domain (Wikimedia Commons, Thesaurus_opticus_Titelblatt.jpg).