tydex

Typed decision primitives for LLMs — choice, score, and noul (yes/no) — returned as data with probabilities, not prose. Built on top of a single model call with greedy decoding, plus calibration, confidence-based escalation, and an iterative re-calibration loop.

tydex is designed as a decision engine: consume it via the Python API or the HTTP server. There is no chat UI.

Why

A decision workflow built on chat completions has to parse free text and guess how sure the model is. tydex instead asks the model for a single typed token, reads its probability mass directly, and lets you push that raw distribution through:

• Temperature scaling (p^(1/T)) to trade confidence against calibration.
• Isotonic recalibration against your own labeled samples (with a blend toward identity when data is scarce).
• Routing/escalation — cheap model by default, escalate to a frontier model (or a human) when confidence is too low.
• Feedback loop — label predictions and auto-refit the calibration as samples accumulate.

tydex vs Jev

tydex is an open-source behavioural clone of Jev, the decision engine that inspired it. The primitive semantics (choice/score/noul) are modelled after Jev; the engineering tradeoffs favour openness and self-hosting.

Capability                          tydex                          Jev (original, closed)
-------------------------------------------------------------------------------------------
Source code                         open (GitHub)                  proprietary / closed
Install                             pip install . / py -m tydex     closed SaaS
Models & providers                  any: OpenAI, Anthropic,        fixed internal vendor set
                                    Ollama, OpenAI-compat, mock
No vendor lock-in                   yes                            no
Probabilities                       raw token logprobs or          internal estimate, not
                                    self-estimate JSON; auditable  documented
Calibration                         temperature + isotonic PAV     internal, not open
                                    + auto temperature, ECE report
Confidence escalation               tiered routing + human         yes (clone target)
                                    fallback
Feedback / auto-refit               Recorder + AutoCalibrator      internal
Interfaces                          Python API, HTTP, CLI, bench   limited HTTP/SDK
Self-hosting & data                 yes, all files on your side    no, data on their servers
Verification                        89 tests + CI + e2e            not published

Implementation openness (illustrative, 0–10):

tydex      ██████████  10  — your code, data, and calibration pipeline
Jev (orig) ██           2  — black-box SaaS

Primitive behaviour parity vs Jev:

choice  ██████████  100%  (logprobs + self mode, temperature scaling)
score   ██████████  100%  (ordered levels, per-primitive calibration)
noul    ██████████  100%  (probability + bool_value)

The Jev column describes the behaviour that tydex clones when it applies to primitives or features that are not publicly documented — it is not a claim about Jev's internal implementation.

Install

pip install tydex

From a source checkout: pip install .. Python >= 3.10. No third-party runtime dependencies (HTTP backends use stdlib urllib).

Quick start

from tydex import Tydex, MockBackend

tdex = Tydex(MockBackend(), model="mock")

r = tdex.choice(
    {"topic": "refund"},
    ["refund", "replace", "no_action"],
    question="How should we resolve this ticket?",
)
print(r.choice)              # "refund"
print(r.probabilities)       # {"refund": 0.6, "replace": 0.3, "no_action": 0.1}
print(r.confidence)          # 0.6
print(r.source)              # "logprobs"

The three primitives

Primitive	Returns	When
choice	one of your options + full probability distribution	single/multi-class selection
score	a level from an ordered set (wraps choice)	prioritize/triage (critical → low)
noul	probability of a statement being true + bool_value	yes/no gates, claim validation

n = tdex.noul({"priority": "P0"}, "This incident violates the SLA.")
print(n.probability)   # 0.9
print(n.bool_value)    # True

Modes

Two ways to obtain probabilities; chosen automatically via mode="auto":

• logprobs — read the actual token probability from the model (supports_logprobs backend). Exact and cheap, but needs backend support.
• self — ask the model for JSON ({"choice": ..., "probability": ...}) and parse it. Works with every backend, including the Ollama cloud API that exposes no logprobs.

tdex.choice(state, options, mode="self")          # force JSON estimate
tdex.choice(state, options, temperature=2.0)      # sharpen/soften via p^(1/T)

Backends

All backends speak plain HTTP to a chat-completions-style or Messages-style API.

Backend	Family	logprobs	CLI provider
MockBackend	—	yes	mock
OpenAIBackend	OpenAI	yes	openai
OpenAICompatibleBackend	any /v1/chat/completions	toggle	openai_compatible_cloud / openai_compatible_local
LocalOpenAIBackend	OpenAI-compatible	toggle	—
OllamaBackend	Ollama /api/chat	no	ollama
AnthropicCompatibleBackend	Messages /v1/messages	no	anthropic

openai_compatible_cloud defaults to https://api.openai.com/v1 (override with OPENAI_BASE_URL or --base-url) and reads OPENAI_API_KEY/OLLAMA_API_KEY; openai_compatible_local targets a server on your machine — http://localhost:8000/v1 by default (override with OLLAMA_HOST) — and accepts the same two key vars.

from tydex import AnthropicCompatibleBackend, Tydex

tdex = Tydex(
    AnthropicCompatibleBackend(
        model="claude-3-5-haiku-latest",
        api_key="...",
        base_url="https://api.anthropic.com",
    ),
    model="claude-3-5-haiku-latest",
)

The OpenAI SDK is not required; every backend uses stdlib HTTP.

Calibration

The model's raw probabilities are rarely calibrated. tydex provides two layers.

Temperature scaling

from tydex.calibration import tune_temperature

probs = [{"refund": 0.95, "replace": 0.05}]   # list of predicted distributions
labels = ["refund"]                            # ground truth labels
result = tune_temperature(probs, labels)       # sweeps T to minimize ECE
print(result.temperature, result.metrics.ece)  # e.g. 2.1, 0.060

Calibration system (isotonic + temperature)

Fit per-primitive isotonic (PAV) recalibration with a blend toward identity for small samples, plus per-primitive tuned temperature:

from tydex import CalibrationSystem, Recorder

system = CalibrationSystem(Recorder("tydex-feedback.jsonl"), min_samples=8).fit()
calibrated = system.apply_to(tdex)

r = calibrated.choice({"x": 1}, ["a", "b"])     # source becomes "logprobs+cal"

Results persist via system.save("tydex-calibration.json") / system.load(...).

Escalation routing

Route across tiers of cost: a cheap local model first, escalate to frontier when confidence is too low, and finally to a human:

from tydex import RoutedTydex, Tier, Tydex
from tydex.backends import MockBackend

weak = Tydex(MockBackend({"0": 0.55, "1": 0.45}), model="local")
routed = RoutedTydex([
    Tier(weak, threshold=0.8, label="local-llama", cost=0.1),
    Tier(Tydex(MockBackend({"0": 0.92, "1": 0.08}), model="gpt-4o"), threshold=None, label="frontier", cost=1.0),
])

rr = routed.choice({"q": 1}, ["a", "b"])
print(rr.tier, rr.escalations, rr.total_cost)   # e.g. "frontier" ["local-llama"] 1.1

A tier with action="human" raises RequiresHuman (with the full request context) when confidence never reaches the threshold.

Feedback loop & auto-calibration

Recorder logs every prediction (tydex-feedback.jsonl); label them and AutoCalibrator re-fits and saves the calibration live.

from tydex import AutoCalibrator, Recorder

auto = AutoCalibrator(Recorder(), refit_every=20, min_samples=8)
calibrated = auto.apply_to(tdex)

entry = auto.recorder.choice({"x": 1}, ["a", "b"], tdex.choice({"x": 1}, ["a", "b"]))
auto.label(entry.id, "a")          # logs the ground truth
auto.maybe_refit()                 # refits when refit_every pending samples accumulate
print(auto.status())               # pending, temperatures, history, labeled_total

HTTP server

Run the decision engine behind a tiny JSON API. The server shares the same AutoCalibrator, so labels submitted over HTTP update the live calibration automatically.

python -m tydex.server --backend mock --port 8000
python -m tydex.server --backend anthropic --model claude-3-5-haiku-latest
python -m tydex.server --backend openai --model gpt-4o-mini
python -m tydex.server --backend ollama --model llama3.1:8b
python -m tydex.server --api-key <key>          # require auth (env: TYDEX_API_KEY)
python -m tydex.server --cors                    # allow cross-origin browser calls

Endpoints:

• POST /evaluate — run a batch of questions:

  {
    "state": {"ticket": "wrong item shipped"},
    "questions": [
      {"id": "q1", "type": "choice", "options": ["refund", "replace"]},
      {"id": "q2", "type": "noul", "statement": "This needs agent approval"},
      {"id": "q3", "type": "score", "levels": ["critical", "high", "low"]}
    ]
  }
  

  Each result carries an id, type, its value, probabilities/probability, confidence — and a log_id when auto-calibration is attached.
• POST /label — attach ground truth to a log_id; triggers maybe_refit().
• POST /refit — force a manual recalibration.
• GET /calibration — current status: pending count, temperatures, per-primitive meta, refit history.
• GET /health — liveness.
• GET /openapi.json — OpenAPI 3.0 spec of the API.

Security: with --api-key (or env TYDEX_API_KEY) every route except /health and /openapi.json requires Authorization: Bearer <key> or X-Api-Key: <key>. --cors enables browser preflight (OPTIONS) and Access-Control-Allow-Origin headers.

Run with Docker

docker build -t tydex .
docker run -p 8000:8000 --env-file .env tydex

Environment-driven defaults (--backend auto): OPENAI_API_KEY, ANTHROPIC_API_KEY, OLLAMA_HOST (plus OPENAI_MODEL, ANTHROPIC_MODEL, OLLAMA_MODEL). A .env file is loaded automatically (supports both KEY=value and KEY: value); a filled-in template lives at .env.example.

CLI

pip install . provides a tydex command (also runnable as py -m tydex). It connects to any AI provider from the shell and runs a single typed decision:

tydex providers                      # show providers, endpoints, key vars, configured status
tydex ask --type noul --statement "A week has seven days"
tydex ask --provider openai --type choice \
       --options "refund,replace" --state '{"ticket": "wrong item shipped"}'
tydex ask --provider openai_compatible_cloud --base-url https://ollama.com/v1 \
       --model gemma4:31b --no-logprobs --type noul --statement "..."

--provider defaults to auto and picks from the environment (OPENAI_API_KEY, ANTHROPIC_API_KEY, OLLAMA_API_KEY, OLLAMA_HOST); --mode self/--no-logprobs force the JSON self-estimate path when the endpoint exposes no token logprobs. Add --json for machine-readable output.

CLI tooling

• demo.py — end-to-end walkthrough: mock execution, calibration, escalation, calibration system.
• bench.py — run the labeled dataset (data/tickets.jsonl: 40 samples — 16 choice, 14 noul, 10 score) against a real backend and report accuracy + ECE before/after temperature tuning:

  python bench.py --model gemma4:31b --base-url https://ollama.com/v1
  python bench.py --model gemma4:31b --base-url https://ollama.com/v1 --reset-recorder
  

  Requires an Ollama cloud API key in OLLAMA_API_KEY (or OPENAI_API_KEY).

Development

python -m unittest discover -s tests     # 90+ tests, incl. local HTTP fakes for both API families
python -m ruff check .                   # lint (ruff config in pyproject.toml)
python -m mypy tydex                     # optional static typing

CI (.github/workflows/ci.yml) runs lint + the full suite on Python 3.10–3.13.

Project layout

tydex/
  core.py          # Tydex + choice/score/noul primitives, logprobs & self modes
  backends.py      # OpenAI-compatible, Anthropic-compatible, Ollama, Mock
  calibration.py   # ECE/Brier metrics, temperature tuning
  calibrated.py    # CalibrationSystem, isotonic PAV + identity blend
  routing.py       # RoutedTydex tiers, escalation, RequiresHuman
  feedback.py      # Recorder (thread-safe, atomic writes) + reports
  autocal.py       # AutoCalibrator live refit loop
  server.py        # HTTP API + CLI
  config.py        # .env helper
data/tickets.jsonl # labeled ticket-classification benchmark