Gatecheck

(formerly evalgate, and before that llm-eval-harness)

A small, provider-agnostic harness for evaluating LLMs the way you'd test any other system: define task suites, grade outputs with deterministic checks and an LLM judge, gate regressions in CI, and track cost and latency per model.

Why this exists. Shipping with LLMs without evals is shipping on vibes. An eval harness is the testing layer for model-backed systems: it turns "does the new prompt/model feel better?" into a number you can diff, gate a pull request on, and watch over time. This repo is a compact, readable reference implementation of that idea.

Why this and not promptfoo / Inspect AI / Braintrust?

Those are all good tools — and all heavier than many projects need. promptfoo brings a Node toolchain; Inspect AI is a full research framework; Braintrust is a SaaS with an account and a data pipeline. This harness occupies a deliberately small lane: a dependency-light, self-hostable, vendor-neutral Python harness you can read end-to-end in an afternoon and gate CI on — three runtime dependencies (requests, pyyaml, jsonschema), no SaaS, no Node, no account, no telemetry. Local models via Ollama are first-class (free, key-less), hosted providers are uniform raw-HTTP adapters (including a hand-rolled, test-vectored AWS SigV4 signer for Bedrock), and every artifact it produces is a plain file you can diff and commit. If you need distributed eval infrastructure, use the big tools; if you need honest numbers on a model swap wired into CI by this afternoon, this is the one.

(Not affiliated with EleutherAI's lm-evaluation-harness — that project is a benchmark battery for academic evals; this is a CI regression harness for your own product's tasks.)

A real run

Every bundled suite, two local models and one hosted model, one command (hosted spend: ~$0.001):

gatecheck run --models ollama:llama3.1:8b ollama:qwen2.5:7b-instruct anthropic:claude-haiku-4-5

suite	model	pass (95% CI)	judge	cost	p50
classification	anthropic:claude-haiku-4-5	1.00 [0.44–1.00]	—	$0.0002	0.78s
classification	ollama:llama3.1:8b	1.00 [0.44–1.00]	—	$0	2.26s
classification	ollama:qwen2.5:7b-instruct	1.00 [0.44–1.00]	—	$0	2.18s
structured-output	anthropic:claude-haiku-4-5	0.50 [0.09–0.91]	—	$0.0006	1.26s
structured-output	ollama:llama3.1:8b	0.50 [0.09–0.91]	—	$0	2.80s
structured-output	ollama:qwen2.5:7b-instruct	0.50 [0.09–0.91]	—	$0	2.63s
summarization	anthropic:claude-haiku-4-5	1.00 [0.21–1.00]	4.0	$0.0002	0.81s
summarization	ollama:llama3.1:8b	1.00 [0.21–1.00]	4.0	$0	2.71s
summarization	ollama:qwen2.5:7b-instruct	1.00 [0.21–1.00]	4.0	$0	2.54s

What it caught. All three models failed the same structured-output task in the same way — each returned valid JSON with an invented container key (details, details, parameters), because the task demanded schema-conformant JSON without stating the required shape (params). A three-way identical failure is the signature of an under-specified task, not a bad model. The suite was fixed to name the expected keys, all three models pass the corrected task, and the committed baselines/default.json snapshots the post-fix pass rates that CI now gates against.

Full artifacts from this run — the report, the browsable HTML dashboard, and the per-task transcripts, failures included — are committed in docs/example-run/.

Features

• Task suites — eval cases grouped by capability, defined in plain YAML (suites/*.yaml). No code needed to add a test.
• Two kinds of graders
  • Deterministic: exact, contains, regex, json_schema. Reproducible, free, and safe to run in CI.
  • LLM-as-judge: a second model (or several, ensembled) scores the output 1-5 against a rubric, for subjective qualities (faithfulness, tone) that string matching can't see.
• Regression gating — snapshot a baseline, then fail a run (non-zero exit) if any suite's pass rate drops below the baseline (with optional tolerance) or an absolute floor. Wired into GitHub Actions.
• Cost & latency tracking — token counts, USD cost per model (from an editable pricing table), and p50/p95 latency, aggregated per suite and model.
• Provider-agnostic — one interface, five adapters: Ollama (local, zero-config, free), OpenAI, Anthropic, Google Gemini, and Amazon Bedrock (Anthropic models, SigV4-signed). Models are addressed as provider:model, e.g. ollama:llama3.1:8b or openai:gpt-4o-mini.

How it fits together

suites/*.yaml ─► suite loader ─► runner ─► providers (ollama|openai|anthropic|gemini|bedrock)
                                   │
                                   ├─► graders (deterministic + llm-judge)
                                   ├─► pricing (tokens ─► USD)
                                   └─► results ─► report (md / csv / json)
                                                     │
                                                  gate (vs baselines/*.json) ─► exit code

Quickstart (local, no API keys)

pip install -e ".[dev]"

# requires a local Ollama with the models pulled:
ollama serve
ollama pull llama3.1:8b
ollama pull qwen2.5:7b-instruct

gatecheck run                       # run every suite against the config defaults
gatecheck run --only classification # just one suite
gatecheck run --models ollama:qwen2.5:7b-instruct

Each run prints a summary table (pass rate with a 95% Wilson confidence interval, mean judge score, cost, and p50/p95 latency) and writes five artifacts to results/: a machine-readable run-*.json, summary-*.csv and summary-*.md, a transcripts-*.jsonl with one self-contained record per task (the exact input sent, the model output, and the grades), and a self-contained dashboard-*.html you can open in a browser to scan and filter every result. The summary-*.md also carries a Failures section listing each failing task with its prompt/output preview and the specific graders that failed. When tasks carry tags (set per-task or per-suite in the YAML), the reports add a Tags breakdown — pass rate per (tag, model) across suites, so you can see how a model does on, say, all reasoning tasks.

Regression gating

# 1. record where you stand today
gatecheck baseline --name default

# 2. later -- after a prompt edit, model swap, dependency bump -- gate against it
gatecheck gate --name default                 # fails (exit 1) on any pass-rate drop
gatecheck gate --name default --tolerance 0.1 # allow up to a 10-point drop
gatecheck gate --only classification --min-pass-rate 0.8

In CI (.github/workflows/evals.yml) the test job always runs the unit tests. The gate job runs the deterministic suites against a cheap hosted model and fails the build on a regression — but only if an OPENAI_API_KEY repo secret is configured, so forks and key-less runs skip it cleanly. LLM-judge graders are left out of CI (--deterministic-only) because they aren't reproducible enough to gate on; run those locally or on a schedule.

Providers and cost

provider	id prefix	cost	needs
Ollama	ollama:	free	local server
OpenAI	openai:	per token	OPENAI_API_KEY
Anthropic	anthropic:	per token	ANTHROPIC_API_KEY
Gemini	gemini:	per token	GEMINI_API_KEY
Bedrock	bedrock:	per token	AWS creds + region (env)

Cost is computed from gatecheck/pricing.py, which ships current OpenAI, Anthropic, Google Gemini, and Amazon Bedrock (Anthropic) rates (verified 2026-06-27 against each provider's pricing page; Bedrock matches first-party Anthropic list price). Provider pricing drifts over time, so re-check it periodically. Any model not listed there — including every local Ollama model — is counted as $0.

LLM-judge calls are priced too: each judge invocation (including every member of an ensemble) is added to the task's cost, so --max-cost and the reported totals account for judge spend. The judge_cost_usd column breaks it out.

Pass --max-cost <USD> to any command to cap spend: the run stops before starting the next task once cumulative cost reaches the budget. It's a soft cap (actual spend can exceed it by at most the one task that crosses the line, or by up to --concurrency tasks when running in parallel), and free local models never trip it.

Pass --concurrency <N> to run up to N tasks in parallel (default 1). Each task is a single HTTP call, so this is I/O-bound work that parallelizes well; results are still reported in a stable suite/model/task order.

Pass --repeat <N> to run each task N times; the verdict becomes a majority vote across the runs, and the report shows each task's pass fraction and flags any whose result flips. Set a non-zero temperature in config for this to surface real variance — at temperature: 0 the runs are identical.

Pass --cache to reuse the result of any identical (model, prompt, options) call within a run instead of paying for it twice — useful when the same prompt recurs across suites. Cached tasks report $0 cost and ~0 latency and are marked cached; the cache is skipped under --repeat (repeated sampling must make real calls to measure variance). Dedup is exact when running sequentially; under --concurrency it's best-effort.

Pass --cache-dir DIR to persist that cache to disk, so re-running a suite after editing one task skips re-paying for all the unchanged ones. The key is (model, messages, options): editing a task's prompt/system or swapping the model forces a real call, editing a grader does not (grading re-runs on the cached output). Entries are one JSON file per key, written atomically; a fully-cached re-run costs $0.

Adding a suite

Drop a YAML file in suites/:

name: my-suite
description: what this checks
models: [ollama:llama3.1:8b]      # optional; falls back to config defaults
tags: [smoke]                     # optional; applied to every task in the suite
tasks:
  - id: my-task
    tags: [reasoning]             # optional; merged with the suite's tags
    system: optional system prompt
    prompt: the user turn
    graders:
      - type: contains
        any_of: ["expected", "word"]
        ignorecase: true
      - type: llm_judge
        judge_model: ollama:llama3.1:8b   # or a list to ensemble: [modelA, modelB]
        pass_threshold: 4
        rubric: |
          Score 1-5 ...

Grader reference

type	passes when …
exact	output equals value (optional ignorecase)
contains	output has any_of (or every all_of) substring
regex	pattern matches the output
json_schema	output is valid JSON, optionally matching schema
llm_judge	a judge model scores >= pass_threshold against rubric
pairwise	a judge prefers the output over a reference in both orderings (position-bias-controlled)

llm_judge takes either a single judge_model or a list via judge_models; with several, each scores independently and the task passes on the mean score — judge ensembling, to reduce single-model bias.

pairwise compares the output against a fixed reference answer and runs the comparison both ways (output first, then reference first); a preference only counts if it survives the swap, so a position-biased judge collapses to a tie rather than a spurious win. require: win demands a strict preference; the default (tie) passes unless the output is judged worse than the reference. Like llm_judge it's non-reproducible and excluded from CI gating. Any judge grader's model is swappable at run time with --judge-model (no YAML edit).

The bundled suites

The suites in suites/ are meant to show what's worth measuring, not just that grading runs. Each pairs a deterministic guardrail (reproducible, gate-able in CI) with a judge where the real question is subjective.

suite	what it actually measures	grading
classification	single-label sentiment/intent	deterministic — gates CI
structured-output	strict JSON conforming to a schema	deterministic (json_schema) — gates CI
summarization	faithful, concise summaries	LLM judge
faithfulness	RAG grounding: abstain when the answer isn't in the context; follow the context over parametric priors on counterfactual input	judge + a deterministic abstention/grounding check
tool-use	correct tool selection + argument grounding; decline when no tool fits; ask when a required arg is missing	json_schema (const pins the right tool) on the calls — gate-able; judge on the decline/clarify cases
red-team	refuse genuinely harmful / jailbreak-framed asks and don't over-refuse benign lookalikes (kill a process, SQLi for defense)	LLM judge, tagged refuse vs over-refusal
preference	is the output at least as good as a reference answer?	pairwise LLM judge, position-swap-controlled

The design bias throughout: deterministic where a failure is cheaply and reproducibly checkable, LLM-judge only where it genuinely isn't — and CI gates only on the deterministic graders (--deterministic-only), because a judge's score isn't reproducible enough to fail a pull request on. The judge suites are for local or scheduled runs; the harmful red-team prompts describe disallowed categories generically and contain no operational detail (a passing model produces nothing harmful — the suite tests the guardrail).

Running these against three models made the point concrete: re-grading the same faithfulness outputs with a stronger judge (--judge-model) flipped which model looked worst (qwen under a local 8B judge, llama under Claude) — the weak judge even failed a demonstrably correct answer that the deterministic grader passed. That's why the judge is swappable and why CI trusts only deterministic graders. Write-up and artifacts: docs/example-run/hard-suites.md.

Understanding the codebase

Read the files in this order.

1. gatecheck/types.py — the five dataclasses everything else passes around (Task, Suite, Completion, GradeResult, TaskResult). The whole repo is functions over these. Note TaskResult.verdict: the pass/fail rule.
2. gatecheck/cli.py — the three subcommands (run, baseline, gate) and how a run is wired together. This is the front door.
3. gatecheck/runner.py — the core loop: every (suite × model × task) becomes one TaskResult. The single most important file.
4. gatecheck/providers/ — base.py is the one-method interface; ollama.py, openai.py, anthropic.py, gemini.py, bedrock.py are uniform raw-HTTP adapters (hosted ones retry transient failures via _http.py; Bedrock adds SigV4 signing); __init__.py holds the registry and the provider:model id parser.
5. gatecheck/graders/ — deterministic.py (exact, contains, regex, json_schema) and llm_judge.py (a second model scores 1–5 against a rubric).
6. gatecheck/report.py then gatecheck/gate.py — aggregation into per-(suite, model) summaries, then comparison against a committed baseline.
7. gatecheck/pricing.py — the token→USD table and the $0 fallback.

The one path that matters (a run). Config + suites/*.yaml are loaded into Suite/Task objects → runner iterates suites × models × tasks → for each task, parse_model_id picks the provider, provider.complete() returns a Completion (text + token counts + latency), each grader scores the text (an llm_judge grader is handed a judge_fn that calls a second model), cost_usd prices the tokens, and it all lands in a TaskResult. report.summarize() then groups results by (suite, model) into pass rate, mean judge score, token/cost sums, and p50/p95 latency; gate compares those pass rates to a baseline and sets the process exit code.

Concepts worth understanding:

• provider:model addressing. Models are openai:gpt-4o-mini, ollama:llama3.1:8b, etc. A bare id uses default_provider. The parser is careful about the tag colon: llama3.1:8b splits to provider + 8b only if llama3.1 is a registered provider (it isn't), so the tag survives intact.
• Two grader families. Deterministic graders are reproducible and free, so they gate CI. The LLM judge captures subjective quality (faithfulness, tone) but isn't reproducible — which is why CI runs --deterministic-only.
• The verdict rule. A task passes only if every deciding grader passes; it's None when no grader produced a definite verdict, and False on error. Pass rate is computed over deciding tasks only.
• Two ways the gate fails. A relative drop below the baseline (beyond an optional --tolerance), or falling under an absolute --min-pass-rate floor. A (suite, model) with no baseline entry is reported but never fails the gate.
• Cost model. Local Ollama models and any model not in the pricing table are counted as $0 rather than guessed. temperature: 0 keeps runs as reproducible as the providers allow.

Roadmap & status

Planned work and known limitations are tracked in ROADMAP.md; the current implementation status is in PROGRESS.md.

Layout

gatecheck/     the package (providers, graders, runner, report, gate, cli)
suites/        eval suites in YAML
baselines/     committed baseline snapshots for gating
tests/         pytest unit tests + a stubbed-HTTP end-to-end test
.github/       CI workflow

License

MIT — see LICENSE.