Janus

Janus sends each decision to a small model or to a larger one, according to how confident the small model is. It measures where that line sits on your data before it routes anything. Janus ships no default threshold: it measures one.

_{Real output, replayed from the raw JSONL committed in this repository. No model is called.}

pip install janus-decide

Quickstart

Five minutes, on your own data. Janus ships no default threshold: it measures one.

1. Install

pip install "janus-decide[typesafe,deepseek]"

The bare package depends only on numpy; each backend is an extra. DeepSeek speaks the OpenAI protocol, so [deepseek] and [openai] pull the same client.

2. Prepare a labelled JSONL — one object per line, three fields:

{"id": 0, "text": "I lost my card", "gold_label": "lost_or_stolen_card"}
{"id": 1, "text": "when does my card arrive", "gold_label": "card_arrival"}

and a label file naming every class you allow:

{
  "instructions": "Which banking intent does this customer query express?",
  "labels": {
    "lost_or_stolen_card": "The card has been lost or stolen.",
    "card_arrival": "Chasing a card that was already ordered and has not arrived."
  }
}

Start with a few hundred labelled rows; larger samples generally give more stable estimates. Ours were 500.

3. Measure

janus measure \
  --dataset mydata.jsonl --labels mylabels.json \
  --primary typesafe:jev-latest \
  --fallback deepseek:deepseek-v4-pro \
  --out janus.json

Try --sample 20 --seed 1 first: it checks the wiring and the real cost per call before you spend on the full set. --budget 2.00 stops the run if the projected cost goes over. An interrupted run resumes by id without re-paying for a completed call.

4. Read the table. This is the real output for the Banking77 data in this repository:

  rule                         thr     cov     acc       cost      p50
  always_primary                 -  100.0%   77.8%     0.0507    296ms
  always_fallback                -    0.0%   78.8%     0.2207   2269ms
  ... 33 more thresholds, written to the report
  primary_if_confidence_ge    0.67   88.4%   80.2%     0.1033    302ms
  ... 30 more thresholds, written to the report

VERDICT: ROUTE
  threshold   :     0.67
  accuracy    :    80.2%    +1.4%  vs best single model
  cost        :  $0.1033     -53%  vs fallback only
  latency p50 :    302ms     -87%  vs fallback only
  escalation  :    11.6%  of traffic
  ceiling     :    83.2%  this pair of models, not the task

The full sweep is written to the measurement report.

The verdict compares each number against the baseline the decision is actually made against. accuracy is read against the better of the two models alone. cost and latency p50 are read against the fallback, since routing exists to avoid calling it: here that is less than half the money and a median decision that still answers in about 300 ms, against 2.3 seconds. If you are building anything interactive, that last line matters before the other two.

measure can also conclude DO NOT ROUTE, which is what it does on the second dataset in this repository: no threshold beat the better single model, so the policy runs that model alone rather than paying for an escalation that buys nothing.

5. Route from Python

from janus import Router, question_from_json, resolve

question = question_from_json("mylabels.json")
router = Router.from_file(
    "janus.json",
    primary=resolve("typesafe:jev-1.13.0"),     # the resolved version, not the alias
    fallback=resolve("deepseek:deepseek-v4-pro"),
)

decision = router.decide(input="I lost my card", question=question)
decision.label       # "lost_or_stolen_card"
decision.source      # "primary" | "fallback"
decision.escalated   # False
decision.cost_usd    # from real tokens; None when the rate is unknown

Pin the resolved version here rather than an alias: an alias moves when a release ships, and a threshold measured on one version does not transfer to the next. Aliases are fine in janus measure, which records whatever the API actually answered.

When the models do change under you, janus check says so, and Router raises instead of quietly applying a threshold measured on something else. janus explain shows why one input escalated and another did not.

Why measure at all?

The same pipeline was run on two labelled datasets, 500 examples each, and no routing parameter carried over. The optimal threshold moved from 0.67 to 0.37. The sign of the accuracy gap between the two models reversed. On one dataset routing beat both models on its own; on the other it matched the better one while costing 47% more, so the honest answer there was not to route.

A default threshold would therefore be wrong roughly as often as it was right, which is the whole reason this tool measures instead of assuming.

Full measurement, raw data and limitations: RESEARCH.md.

Reference

janus measure

Runs both models over your labelled data, sweeps every observed confidence level, and writes a policy plus its report.

flag	meaning
--dataset	JSONL with id, text, gold_label
--labels	JSON label file; --labels-module loads a Python module instead, and executes it
--primary, --fallback	provider:model, e.g. typesafe:jev-latest
--out	policy path, janus.json by default
--raw-dir	where the raw JSONL goes; written next to --out otherwise
--target-accuracy	pick the cheapest point reaching it; reported as unattainable rather than revised
--max-cost	discard points above this total
--sample N --seed S	smoke test on N random rows; both flags are required together
--budget	stop if the projected cost goes over
--estimate	print the plan, call nothing
--top N	show only N routed thresholds around the chosen one; the full sweep still goes to the report
--replay	re-measure from raw JSONL already recorded; calls nothing
--task	replay one of the two measurements committed in this repository; implies --replay

--color auto|always|never sits on janus itself. auto follows the terminal, and NO_COLOR wins over FORCE_COLOR.

The run writes one line at a time, flushed and fsynced, and resumes by id. It refuses to resume a file whose statement has changed rather than mix two prompts.

--replay re-runs the whole measurement offline, from raw JSONL a previous run wrote. It is how you sweep to a different --target-accuracy without paying for a single new call. The providers it installs raise if they are asked to answer, so a replay that is missing a row fails loudly instead of quietly calling a model; and it writes no policy, because nothing new was measured.

janus measure --replay --task banking77 --top 3   # the demo above, offline

--task reads the datasets and raw runs committed in the repository, which the installed package does not carry, so it needs a clone. It says so plainly if the files are not there. --replay on its own works anywhere, on your own --dataset and --raw-dir.

janus check

Says whether the policy still applies: same statement, same resolved model versions. --offline validates the file's structure without calling anything.

janus explain

Shows why one input escalated and another did not.

$ janus explain janus.json --input "I lost my card"

Primary: jev-1.13.0
Prediction: lost_or_stolen_card
Confidence: 0.61

Threshold: 0.67
Decision: ESCALATE

Fallback: deepseek-v4-pro
Final label: lost_or_stolen_card
Source: fallback

Python API

from janus import Router, question_from_json, resolve

router = Router.from_file("janus.json", primary=..., fallback=..., on_drift="raise")
decision = router.decide(input="...", question=question)
decisions = router.decide_many(["...", "..."], question=question)

Decision carries label, source ("primary" or "fallback"), escalated, confidence, cost_usd and answers. confidence and a provider's distribution are None when it exposes neither, and cost_usd is None rather than 0.0 when the rate for the returned model is unknown — an unknown must not disappear into a sum.

Router raises NoPolicyError without a policy, and StalePolicyError when the statement or the model versions no longer match the measurement. on_drift takes "raise" (the default), "warn" or "ignore".

Providers

spec	needs	notes
typesafe:jev-latest	[typesafe]	exposes a confidence and a full distribution
deepseek:deepseek-v4-pro	[deepseek]	constrained by a strict tool call
anthropic:claude-opus-5	[anthropic]	constrained by structured output
openai_compat:<model>	[openai]	any OpenAI-compatible endpoint, with base_url

Only a provider that reports a confidence can serve as the primary of a confidence threshold. Register your own with janus.register.

File formats

Dataset — one JSON object per line:

{"id": 0, "text": "I lost my card", "gold_label": "lost_or_stolen_card"}

Labels — labels maps each class to a description, or is a bare list of names:

{"instructions": "Which intent is this?", "labels": {"a": "…", "b": "…"}}

Two real ones are in the repository: data/banking77.labels.json (77 classes) and data/wos.labels.json (145). Both are exact exports of what the measurement sent, checked by recomputing the prompt_hash recorded in the raw runs.

janus.json — the report of a measurement, not a preference. It records the statement's hash, the resolved model versions, the dataset fingerprint, the rule and threshold, the operating point with its baselines, the oracle ceiling, and the full table of observed confidence levels so the sweep can be replayed at another target without a single new call.

Layout

src/janus/     the package
experiments/   the measurement it is built on, unchanged
tests/         invariants, including that the package still reproduces
               the published numbers from the committed raw JSONL
data/          frozen datasets + provenance
results/       raw JSONL and figures
docs/METHOD.md protocol, measured constraints, related work

Research

Two datasets, 500 examples each, protocol frozen before any result, raw JSONL and figures committed. The headline is that nothing measured on the first dataset predicted the second.

• RESEARCH.md — results, calibration, ECE and Brier, agreement between the models, limitations, related work.
• docs/METHOD.md — the protocol, what was fixed before each run, and the constraints measured on the APIs themselves.
• experiments/ — the scripts that produced it.

License

MIT — see LICENSE.

Banking77 is distributed under CC BY 4.0; see data/README.md for provenance and citation.