Nudge

Canonical repository: dossier-labs-ai/nudge. The website source lives in site/.

Nudge learns action utilities from feedback and readable questions about the state. It is an experimental contextual-bandit library: no model fine-tuning, one SQLite file, and an explicit learn() call to activate updates.

Requires Python 3.11+. Install with pip install nudge-policy, or clone and run uv sync. The distribution is nudge-policy; the import is nudge. The PyPI name nudge is held by an unrelated, inactive project (Evite, last released 2012); a PEP 541 transfer request is the path to it.

from nudge import Answer, Choice, Nudge

choice = Choice(
    name="support",
    situation="A support ticket and conversation so far",
    actions={"answer": "Reply now", "escalate": "Hand to a person"},
    feedback={"resolved": "1 if resolved, otherwise 0"},
)

class Baseline:
    identity = "example-baseline-v1"  # pin model, prompt, and decoding revisions

    def ask(self, state, questions):
        # A runnable fixed baseline, not a model integration.
        return Answer({"answer": 0.7, "escalate": 0.3})

with Nudge(choice, store="support.db", model=Baseline(), seed=42) as policy:
    decision = policy.ask(state="My order is damaged")
    policy.feedback(decision.id, resolved=1)
    policy.learn()
    print(policy.report())

Your answerer must return answers for every requested feature question. Add a proposer to enable discovery; without one, Nudge still learns action intercepts and weights on baseline confidence. Caller-supplied initial Questions also work.

Semantics

• ask(state=...) makes one answerer call, samples an action, and records it before returning. It does not execute the action. Feedback must refer to the sampled action actually taken; overrides require separate treatment.
• feedback(id, **fields) requires all declared fields as finite numbers. Every report is retained; the latest report replaces the previous one. Missing reports never mean zero reward. Booleans are rejected; use numeric 0/1.
• learn() fits from a fixed latest-feedback snapshot and immediately activates on success. By default it trains from the retained question bank without discovery. It blocks its caller; other threads/instances can continue asking. It is not scheduled automatically. Concurrent fits use atomic compare-and-swap.
• discover_questions() updates the question bank and evidence without fitting or changing the serving critic or serving questions. Bounded replay samples inform discovery. Call this frequently as feedback arrives, then call learn() when ready to train the routing policy. learn(discover_questions=True) explicitly combines both operations. report().question_bank and .question_evidence expose retained hypotheses separately from active .questions.
• Learning visits labeled decisions newest first. By default a decision's weight halves every seven days relative to the newest labeled decision. Set learn(half_life_seconds=3 * 86400) for a three-day half-life, or None for unit weights. Fitting and question selection use the same decay weights. Late/corrected feedback retains the original decision's age.
• Expert(max_training_rows=1024) caps each learning/discovery update, including feature hydration and evidence, at that many unique labeled decisions. Feedback maintains recent, uniform-history and action-balanced surprise pools in SQLite; learning does not scan the audit log. training_sample_count on learn results and reports distinguishes fitted sample size from total feedback count. Schema v2 requires a fresh store and refit when upgrading; preserve old stores for auditing. Full audit reporting still reads history. See design.
• learn(reward={...}) explicitly changes weights and refits the logged fields. Reopen with those live weights. Changing the action/feedback/model contract requires a new choice name or store in v0.
• decision(id) reads the exact original sampled distribution and explanation.
• report() makes no model calls. It shows observed action outcomes, missing feedback, logged probability changes, model coefficient uncertainty, and each action's learned_weight (sum of weights in the last fitted snapshot).

Baseline probabilities are input features, not constraints on the chosen action. Expert(posterior_draws=64, explore_floor=0.05) averages posterior winner votes and mixes a uniform action distribution. Exploration starts before any feedback; every action receives at least explore_floor / number_of_actions probability, including baseline-zero actions. Coefficients describe predictions, not causes. Discovery supplies action-by-answer outcome comparisons to the LLM and supports atomic question bundles. Answerable proposals are retained as hypotheses without a regressor-utility gate. Later decisions provide separate evidence for refinement.

This policy breaks the previous bounded-tilt contract. Start a new choice name or SQLite store; preserve old stores for audit. The separate-bank contract also requires a new choice/store when upgrading earlier unrestricted experiments. Removed max_drift and sample_after settings are not accepted. Historical analysis tools retain the archived rule.

See the design, benchmarks, adapters, and contributing.

Development

uv sync --group dev
uv run python -m unittest discover -s tests -v
uv run ruff check src tests
uv build
uv run nudge-benchmark --steps 1000 --seeds 0,1,2 --output benchmark-results/run.json

The core has only NumPy as a runtime dependency and imports no host application code. Tests use synthetic states and local SQLite; no credentials are required. Database files contain verbatim states and feedback. Keep them out of source control and distribute only synthetic fixtures.

RouterBench

The RouterBench integration compares Rune alone with Nudge + Rune on a proportionally sampled online stream. It scores both policies before revealing only the chosen action's feedback, learns every 200 prompts, and exports rolling utility and cumulative advantage as CSV, JSON and PNG. It includes TensorZero gateway configuration, bounded concurrent replay, checkpoint recovery, and a dedicated VM runner with provider cleanup.

uv sync --frozen --group dev --extra routerbench --extra operations
uv run python -m nudge.routerbench --help
uv run python -m nudge.routerbench_report --help
uv run python -m nudge.routerbench_vm --help

Real inference requires the explicit gateway and Secret Manager configuration described in the runbook. No credentials are needed for unit tests. See the RouterBench runbook for the live pilot and its artifact locations.

MIT licensed. This is research software, with exploratory selection scores and no automatic statistical certification of policy improvement.