recension 0.5.0

Measured optimization of the text layer around a language model: versioned artifacts, held-out evaluation, leakage detection, and a complete audit record.

recension

Measured optimization of the text layer around a language model (prompts, context templates, skill and instruction files) with the rigor normally reserved for weight training: a held-out objective, a baseline, versioned artifacts, and a complete audit trail.

The name comes from textual criticism. A recension is the revision of a text by collating variant readings and keeping the best-supported one. That is the loop this library runs: propose multiple candidate edits, test each against held-out evidence, commit only what measurably improves, and record why.

Why

The usual way to improve a prompt is to edit it, eyeball a few outputs, and ship. There is no held-out measurement, no record of why a change was made, and no defense against overfitting to the handful of cases you inspected. recension replaces that loop with a measured one:

• No edit is accepted without a held-out score that beats the incumbent. Failures are diagnosed on a train split; acceptance happens only on a validation split, and can require the gain to be statistically significant rather than above an epsilon. An optional locked test split gives an unbiased final estimate.
• Every accepted version carries provenance: the failures that motivated it, the diagnosis, every sibling candidate considered (with scores), and the diff. A reviewer who didn't run the optimization can reconstruct every decision.
• One number doesn't hide regressions. Optional per-slice scores, non-regression guard objectives, and a token-cost ledger surface what an aggregate averages away.
• Leakage is checked, not assumed away. Heuristics flag candidates that embed validation content or show implausible validation gains.
• Records are built to be acted on. Gate a prompt in CI with recension check, detect tampering with recension verify, and share a standalone HTML audit with recension report.
• Compute is a dial. Candidates per round, rounds, diagnosis depth, and a hard ceiling on model calls are all caller-controlled.

Real-world use cases

• Production classification and extraction (support-ticket triage, invoice fields, moderation): improve a labeling prompt on labeled data with measured, regression-safe edits.
• RAG context templates: tune how retrieved chunks are assembled into the prompt with the model held fixed, so the metric move is attributable to the text.
• Agent and skill instructions: optimize longer instruction files judged by an LLMJudge rubric when there is no gold answer.
• Governance and audit: ship a replayable, tamper-evident RunRecord for every prompt change; gate merges in CI with recension check, and hand reviewers a standalone HTML audit with recension report.

Full write-ups, plus a "how it works" walkthrough, are on the documentation site.

Prior art, honestly

DSPy and GEPA own the optimization mechanics this library's ReflectiveOptimizer performs; if you want state-of-the-art prompt optimization algorithms, look there. recension's contribution is the measurement and governance shell around a text artifact: versioned artifacts with provenance, leakage detection, the complete audit record, and budgeted update-time compute. That delegation is a real seam, not a promise: the Proposer protocol lets an external engine supply the candidate edits while recension keeps owning the artifact, the measurement, and the record (see the Bring your own optimizer guide).

Install

pip install recension              # core: zero provider dependencies
pip install "recension[anthropic]" # adds the Anthropic backend

Python 3.12+. The core (and the whole test suite) runs against a deterministic MockModel with no API key and no network.

Quickstart

from recension import (
    Budget, EvalSet, ExactMatch, MockModel, ReflectiveOptimizer, TextArtifact,
)

artifact = TextArtifact.from_text("Label the sentiment of the message.")

# Held-out examples, split into train (for diagnosis) and validation (for
# acceptance). Load from a JSONL file with EvalSet.from_jsonl(path) instead.
evalset = EvalSet.from_records([
    {"id": "t1", "input": "Absolutely love this", "expected": "positive", "split": "train"},
    {"id": "t2", "input": "Broke after a day", "expected": "negative", "split": "train"},
    {"id": "v1", "input": "Terrible support", "expected": "negative", "split": "validation"},
    {"id": "v2", "input": "Exceeded expectations", "expected": "positive", "split": "validation"},
])

optimizer = ReflectiveOptimizer(
    artifact=artifact,
    evalset=evalset,
    objective=ExactMatch(),
    model=MockModel(),          # offline mock; see below for the real backend
    budget=Budget(candidates_per_round=4, rounds=3, max_model_calls=200),
    seed=7,
)
record = optimizer.run()

print(record.summary())         # baseline → accepted versions → final score
record.save("run_record.json")  # the complete audit artifact

To run against a real model, install the extra (pip install "recension[anthropic]"), set ANTHROPIC_API_KEY in your environment, and use the Anthropic backend:

from recension.models.anthropic import AnthropicModel  # kept off the top-level import so the core needs no provider deps

optimizer = ReflectiveOptimizer(..., model=AnthropicModel())

API keys are read from the environment only, never from code or config.

CLI

recension run --config run.yaml      # execute an optimization, write the record
recension show run_record.json       # baseline, accepted diffs, score progression, integrity
recension diff run_record.json vA vB # diff between two artifact versions
recension check --config run.yaml --baseline run_record.json  # CI guard: exit non-zero on regression
recension verify run_record.json     # detect tampering (content-addressed version chain)
recension report run_record.json -o report.html  # standalone HTML audit page

A runnable, fully commented config and dataset live in examples/cli/ (recension run --config examples/cli/run.yaml); the config schema and a GitHub Actions recipe for recension check are in the CLI guide.

Documentation

Full docs, API reference, and three worked examples (each reproducible offline against MockModel): https://anthonynystrom.github.io/recension/

License

MIT