v2.5.0-ce — Decisions in one call with the d1 decision model¶
Released 2026-11-11 · GitHub release
v2.5.0-ce adds a new endpoint, POST /v1/systemone, that returns a typed
decision instead of prose — a routing choice with a probability for every option, a
yes/no probability, or an urgency score — in a single forward pass with zero generated
tokens. It is backed by d1, a purpose-built decision model, and it accepts any causal
model you have pulled. Additive endpoint, drop-in upgrade, no data migration. Community
Edition, free, seamless upgrade from v2.4.x.
Why it matters
Most "should the agent do X?" questions don't need a chat completion — they need an
answer you can branch on: which team to route to, whether the customer is angry, how
urgent a ticket is. Until now you'd coax that out of a chat model with brittle
prompt-parsing. /v1/systemone gives you routing, triage, guardrail checks and
priority scoring as structured JSON you act on directly — faster, cheaper, and
deterministic to consume.
One call, a decision you can act on¶
Send the situation as state, plus one or more typed questions, and get a structured
answer back for each:
choice— give a set of named options; get the chosen option, a probability for every option, and a confidence. Routing, triage, classification.noul— ask a yes/no question; get the probability of "yes". Guardrail checks, gating.score— give ordered levels, lowest first; get the expected (fractional) level with per-level probabilities. Priority and urgency scoring.
Because it weighs the options rather than writing text, a decision comes back in one pass and
reports output_tokens: 0 — fast and cheap enough to put on a hot path.
See the full install → enable → call-it walkthrough on the Decision model (d1) page.
Proof: one message, three decisions¶
A real support message — "You charged my card twice and still have not shipped my order. I
need this sorted TODAY or I am disputing the charge." — answered by d1 in a single
/v1/systemone call:
route → billing(0.93)angry →0.78urgency → "today"(0.79)output_tokens: 0
You branch your workflow on the choice, on noul > 0.5, or on the score bucket — with the
full probability maps there when you'd rather threshold than take the argmax.
Backed by d1, open to any model¶
d1 is the calibrated, recommended model — its probabilities are tuned to mean what they say.
But the endpoint isn't locked to it: name any causal model you've pulled in the model field
and it works. d1 is pull-on-demand (not bundled in the image): pull it once (~2.7 GiB)
and the endpoint loads it on first use.
Also in this release¶
- AutoML cleanup is reliable.
DROP MODEL name;,DESCRIBE MODEL name;and the other model/experiment commands now handle a trailing;correctly, so teardown and re-runs no longer error on the name. - One default chat model. The embedded
localprovider now defaults toqwen2.5:7bacross chat,AGENT_RUN, and NL tooling — one consistent, budget-adaptive default. - Cypher
CREATEaccepts a comma-separated pattern list —CREATE (a), (b)-[:R]->(c)— matching standard Cypher, so multi-node creates no longer need separate statements. - Faster batch embeddings against Ollama via
/api/embed, with a graceful fallback on older Ollama servers.
Proven before shipping¶
- The state-asserting feature validator is fully green on the release build (316 assertions, 0 failures), including vector search and the bundled UI.
/v1/systemoneis parity-matched to the reference d1 across the decision suite — argmax agrees on every case, worst-case probability difference 0.025.- Clean boot plus a d1 decision and an
EMBEDrun on a non-AVX-512 CPU with no GPU — no illegal-instruction faults on older hardware.
Upgrade¶
Drop-in. No on-disk format change — existing data and vectors read as-is.
Then pull the decision model once (synapcores --config <cfg> pull d1:latest) and name
"model": "d1:latest" on each /v1/systemone call. Full walkthrough:
Decision model (d1).