MIMIR

mimir-decisions

Decisions your agents can act on. MIMIR is a non-generative decision model: give it a context, a question and the options, and get back a typed answer with calibrated probabilities, the evidence behind it, and a certified verdict on whether to act or escalate. No text generated. Nothing to parse. Nothing to hallucinate.

It beats Laya and GLiNER2.5-Decide head-to-head on six of ten tasks — by 54.7 points on Banking77, 42.3 on MASSIVE, 36.3 on typed decisions — and where it cannot back an answer, it abstains instead of guessing.

pip install "mimir-decisions[local]"        # the local engine (CPU and CUDA)

pip install mimir-decisions # data models and HTTP client only

Python 3.11+. Documentation: https://abderahmane-ai.github.io/mimir/

---

Why MIMIR

Most agents route, classify, and verify using a general-purpose language model: slow, expensive, and impossible to audit. MIMIR is built for structured decisions. It runs locally on Torch, on CPU or GPU, returns calibrated probabilities with every answer, and issues a certificate — measured evidence that answers passing its threshold stayed at or below the risk level you ask for, on held-out data.

• No generation. Answers are drawn from the options you supply, not synthesised. The model cannot hallucinate an answer that wasn't on the list.
• Calibrated confidence. Probabilities are not softmax scores; they are calibrated to match realised accuracy on held-out data.
• Answers, always. Every decision returns the model's prediction with its probabilities. In threshold and certified modes, answers below the floor come back deferred for review — never withheld.
• One typed contract. Seven decision types — choice, multi-choice, yes/no, verify, rank, rate, estimate — all returning the same result shape, over any context.
• Portable. The same Python interface works locally on CPU or GPU, over HTTP, and over MCP. Framework adapters exist for eight agent SDKs.

---

Quickstart

from mimir import Context, Field, Mimir, Passage

model = Mimir.from_pretrained("Mythologic/MIMIR-1")

The short form: the chosen option id, or None when the model abstains or defers

team = model.pick( "Hi, invoice INV-8821 was billed twice. Please refund the duplicate.", ["billing", "technical", "sales"], ) # "billing"

Context pairs prose passages with typed structured fields (amounts, IDs, metadata)

context = Context( passages=[ Passage( title="Ticket #4091", text="Hi, we were billed twice ($2,400 total) on invoice INV-8821. Please refund the $1,200 duplicate today or we will cancel our plan.", ) ], fields=Field.from_json({ "invoice_id": "INV-8821", "duplicate_amount": 1200, "customer_plan": "enterprise", }), )

The question is optional: model.choose(context, options) asks "Which option applies to this context?"

result = model.choose( context, "Which department should handle this request?", options={ "billing": "Billing: invoices, payments, refunds", "technical": "Technical: bugs, outages, system errors", "sales": "Sales: pricing, new contracts", "other": "Other: everything else", }, )

result.status # Status.DECIDED, Status.ABSTAINED or Status.DEFERRED result.answer # "billing", or None when no option applies result.probabilities # calibrated probability of each option id result.certificate # mathematical proof of risk bound on held-out data

answer is always the model's prediction. status says whether it cleared the operating floor:

• DECIDED — act on answer.
• ABSTAINED — no listed option applies (clean OOD rejection).
• DEFERRED — the answer came in below the certificate floor; route to a human reviewer.

The Risk Dial: From Mission-Critical SLA to High-Throughput Agents

The release certifies thresholds across four finite-sample risk levels (model.info().risk_levels = (0.005, 0.01, 0.02, 0.05)):

• `risk=0.01` (99.0% SLA) — Mission-critical aerospace / financial SLA: Sets an ultra-strict statistical floor. Only near-certain answers execute automatically; anything uncertain is deferred for human review rather than risked.
• `risk=0.05` (95.0% SLA, the default) — High-throughput web agent / customer workflow: The sweet spot for autonomous agents, unlocking automated execution on legitimate requests while still mathematically bounding error rates on held-out data.
• `mode="standard"` — Returns the calibrated argmax without finite-sample risk deferral; risk then only selects the certificate attached to the result. mode="certified" also defers below the threshold certified at risk.

Give the model the ticket as a person wrote it paired with typed fields — the same options over a one-line summary can come back ABSTAINED. The Decisions guide covers the question and option shapes that decide.

The first call downloads the model from the Hugging Face Hub at the revision this package version pins, verifies its Sigstore signature, checks every file against the manifest's SHA-256, and loads it.

---

Decision types

from mimir import Context, Field, Passage, Rate, Table

context = Context( passages=[Passage(title="Ticket #4412", text="The export has failed every night this week.")], tables=[Table.from_rows([["2026-03-02", "failed"]], header=["date", "status"])], fields=Field.from_json({"customer": {"plan": "enterprise", "seats": 240}}), ) result = model.decide(context, Rate("How urgent is this?", ["low", "medium", "high"]), risk=0.01)

A context can be a string, a list of strings, a dict read as a JSON state, or a Context of typed passages, tables, and fields. Table.from_dataframe(frame) reads a pandas or polars DataFrame. decide_many batches multiple decisions, and every method has an async counterpart (adecide, adecide_many, …).

---

Certification

decide answers every request in standard mode. threshold mode defers answers below your min_confidence; certified mode defers answers below the release's threshold at risk (model.info().risk_levels), and attaches the certificate when the answer passes. Dial risk=0.01 (99% SLA) for mission-critical operations where false actions carry heavy penalties, or risk=0.05 (95% SLA) for high-throughput autonomous agents. decide_uncertified returns the raw model answer with no policy applied.

A certificate covers one exact configuration: weights, Torch version, device, and hardware. On hardware not listed in the certificate, the first load runs the release's equivalence set and requires every decision to match. To certify thresholds on your own labelled data:

mimir calibrate labelled.jsonl --risk 0.01 --confidence 0.95 --out policy.json

model = Mimir.from_pretrained("Mythologic/MIMIR-1", policy="policy.json")

---

Remote use

from mimir import MimirClient

remote = MimirClient("https://mimir.internal", api_key="...") remote.choose( "Hi, we were billed twice for March. Please refund the duplicate today or we will cancel our plan.", "Which department should handle this request?", options={ "billing": "Billing: invoices, payments, refunds", "technical": "Technical: bugs, outages, system errors", "sales": "Sales: pricing, new contracts", "other": "Other: everything else", }, )

MimirClient has the same interface as Mimir, so all code, decision tools, and framework adapters accept either. It requires only the base install. Connection errors, timeouts, and 429/502/503/504/529 responses are retried with exponential backoff that honours Retry-After.

---

Decision tools

from mimir import Choice

route_ticket = model.tool( "route_ticket", Choice( "Which department should handle this request?", { "billing": "Billing: invoices, payments, refunds", "technical": "Technical: bugs, outages, system errors", "sales": "Sales: pricing, new contracts", "other": "Other: everything else", }, ), description="Route a support ticket to the team that owns it.", ) route_ticket( "Hi, we were billed twice for March. Please refund the duplicate today or we will cancel our plan." ) route_ticket.input_schema, route_ticket.output_schema

Tools can also be declared in a YAML file, which the HTTP and MCP servers load:

tools:

• name: route_ticket description: Route a support ticket to the team that owns it. decision: type: choice question: Which department should handle this request? options: billing: "Billing: invoices, payments, refunds" technical: "Technical: bugs, outages, system errors" sales: "Sales: pricing, new contracts" other: "Other: everything else"

---

Tool-call checks

A tool-call check decides, against rules you write, whether an agent's pending tool call may run. A confident yes allows it, a confident no denies it, and anything else escalates to a person.

check = model.tool_call_check(

["Refunds above 500 dollars need a manager's approval."], tools=["issue_refund"] ) outcome = check("issue_refund", {"order": "4412", "amount": 900}) outcome.permission # Permission.ALLOW, Permission.DENY or Permission.ESCALATE outcome.reason # one sentence for the agent or the approver

---

Benchmarks

benchmarks/mimir_vs_laya_enterprise_benchmark.ipynb (run it in Colab) runs ten scenarios, each a written policy plus a pending tool call: five legitimate and five that violate the policy. Every system sees the same prompts, tools, arguments and policy text. Recorded run: Colab CPU, mimir-decisions 1.2.0, laya 0.4.0.

The suite is ten scenarios written by the MIMIR authors, so the figures describe this suite rather than a general rate. MIMIR-1's four escalations came at the pick step; its policy check allowed or denied, and denied the two violating calls the router did route to their requested tool (4B and 5B). MIMIR-1 ran in threshold mode at 0.50. Latency is a Colab CPU runtime with two model passes per decision.

benchmarks/sysone-bench records the run on the independent sysone-bench suite.

---

Agent frameworks

Each adapter turns decision tools into the framework's native tool type and wires a tool-call check into that framework's own approval hook.

| Py