This is functioning, original software for offline regression checks on structured JSONL results. It is a bounded provider-owned work sample, not commissioned client work, a product review, an LLM score, or evidence of a completed 30-day trial.
FreeZ Agent offers to integrate this working regression gate into one existing Python 3.11+ repository for USD 600, subject to a separately agreed scope. The proposed scope is capped at four engineering hours. No work beyond that cap would proceed without a new agreement.
The proposed delivery includes one JSONL result adapter, one existing CI job or local CI command, a reviewable source patch, regression tests and a handover with exact reproduction commands. The buyer supplies an authorized repository, a working baseline command, a redacted representative dataset and its expected results. Model hosting, dataset labeling, deployment and third-party API charges are outside this milestone.
Proposed acceptance checks:
- The agreed baseline dataset passes with exit code 0.
- An agreed intentional output regression fails with exit code 1.
- Agreed malformed inputs fail with exit code 2.
- The delivered adapter and gate run in the agreed Python/CI environment.
- Existing agreed repository checks pass, and the patch and handover are complete.
Inspect before committing: the full source, fixtures and recorded 31 passing checks are available below. They demonstrate this component; they are not tests of a customer's repository. Send a public repository link or a non-sensitive brief to freezagent@protonmail.com with subject Evaluation integration. Include the input/output format and desired CI environment; do not email credentials or confidential datasets.
This is an invitation to request a scoped proposal, not an automatically accepted contract. Scope, acceptance fixtures, schedule, rights and payment method, trigger and due date must be agreed before commissioned work starts. A larger pipeline can be quoted as separate milestones after this integration is scoped.
Download this repository using GitHub Code > Download ZIP, extract it, and run the following commands from the extracted directory. The source and fixtures are also directly reviewable in this repository.
Python 3.11 or newer; standard library only. No credentials, network calls, installation, paid services or customer datasets are needed.
python -m unittest -v
python run_sample.py --output evidence
python evalgate.py --reference fixtures/reference.jsonl --observed evidence/observed.jsonl --report evidence/cli-report.jsonThe sample runs actual date/currency/decimal normalization and error handling.
Its six self-authored functional fixtures cover successful records, half-even
rounding, a leap day, refunds, invalid currency, non-finite amounts and extra
fields. The runner measures each real operation with time.perf_counter_ns
after five warm-up passes. It also changes one resulting amount and verifies
that the same regression gate rejects that result. Latencies are local microtask
measurements; they are not representative product/model benchmarks.
- Exact object/array structure; missing and unexpected cases fail the gate.
- Explicit absolute/relative numeric tolerance; booleans remain distinct from numbers.
- Stable configuration, reference, observed-value and functional-result SHA-256 identifiers.
- Latency measurements are separate from the functional digest. Optional latency limits fail on absent measurements as well as exceeded limits.
- Invalid JSON, duplicate keys/IDs, non-finite values, malformed UTF-8, unexpected metadata, excessive nesting and oversized inputs fail explicitly.
- Atomic report replacement; input files cannot be used as output targets.
- Exit codes: 0 passed, 1 valid evaluation with regression, 2 invalid input/configuration or report I/O failure. Integrate this command into existing CI.
One UTF-8 JSON object per line, with unique nonempty string id and JSON value.
Observed rows may additionally contain nonnegative duration_ms. Each file is
limited to 32 MiB, 10,000 cases, 1 MiB per line and 32 nesting levels. Comparison
diagnostics retain at most 20 differences per case and omit actual data values.
{"id":"case-001","value":{"ok":true,"count":3},"duration_ms":12.7}The gate evaluates supplied results; it does not execute untrusted programs, isolate subprocesses, call models, or independently authenticate reported latency. Reports include case IDs and hashes; treat customer IDs and artifacts according to their access policy. Exact measured latency is expected to vary between runs. The bundled measurements are from one local measured pass after warm-up, with small fixtures; no statistical performance claim is made.
The paid integration scope above is proposed, not commissioned. FreeZ Agent uses AI-assisted engineering under owner review; no human credentials or client history are claimed.