Release Gate

The checklist that must be complete before the version is bumped from 0.x alpha to 1.0.0.

All accuracy gates are currently unproven

Version 0.1.0a1 is alpha. The private holdout corpus does not exist yet, the conformant runtime has not been retained, and no gate measurement has been taken. Every item below marked as unmet ([ ]) is currently unproven.

Accuracy gate

#	Status	Requirement
1	`[ ]`	Frozen private holdout corpus authorized — ≥ 300 cases, ≥ 100 per app (blinkit, swiggy, zepto), collected with explicit consent
2	`[ ]`	Pre-OCR labels verified — ground-truth labels created before any OCR run; ≥ 10% double-labelled and adjudicated
3	`[ ]`	Image deduplication complete — SHA-256 exact dedup and perceptual near-duplicate review
4	`[ ]`	Exact total + currency ≥ 95% — `exact_match_rate >= 0.95` on the frozen corpus under `runtime_policy="conformant"` on the retained reference runtime
5	`[ ]`	False-strong count = 0 — zero cases where `evidence_grade == "strong"` and `total + currency` is incorrect
6	`[ ]`	Confirmation rate = 100% — every result has `requires_confirmation=True`, confirmed in the holdout run

Runtime and performance gate

#	Status	Requirement
7	`[ ]`	Retained reference runtime available — immutable instance of `ubuntu-24.04 / amd64 / tesseract 5.3.4` (baseline `linux-noble-amd64-2026-08`) provisioned and locked
8	`[ ]`	p95 latency < 5 000 ms — measured on the retained reference runtime across all holdout cases
9	`[ ]`	600 MiB memory ceiling — peak per-process memory ≤ 600 MiB on the reference runtime
10	`[ ]`	Deterministic repeat probe — same image run twice produces identical `total`, `currency`, `evidence_grade`, and `matched_label`
11	`[ ]`	Model and runtime SHA-256 hashes recorded — `runtime-baseline.toml` checksums match the hashes measured during the holdout run

CI gate

#	Status	Requirement
12	`[x]`	Portable tests green — `pytest -m 'not integration'` passes
13	`[ ]`	Integration tests green — `pytest -m integration` passes with Tesseract installed
14	`[ ]`	Conformance CI pipeline green — CI run on the retained reference runtime with `runtime_policy="conformant"`
15	`[x]`	Package build and twine check — `python -m build && twine check dist/*` succeeds
16	`[x]`	Artifact privacy scan — `tools/check_artifacts.py` reports no receipt images or private data

Privacy and security gate

#	Status	Requirement
17	`[ ]`	Privacy review complete — documented review covering: no real data in repo, diagnostics default off, `source.filename` handling, Tesseract temp-file behaviour, no network requests
18	`[ ]`	Security review complete — review of input validation, error message content, and dependency versions
19	`[ ]`	No private data in public repository — confirmed that no holdout images, OCR text, or per-case labels appear in git history, CI artifacts, or published docs

Documentation gate

#	Status	Requirement
20	`[x]`	Schema reference published
21	`[x]`	Architecture documented
22	`[x]`	ADRs 0001–0008 complete
23	`[ ]`	Migration notes verified — at least one consumer has been tested against the guide
24	`[ ]`	README updated for v1

Gate thresholds in detail

The evaluator checks all four accuracy/performance thresholds simultaneously in a single run:

{
  "gates": {
    "eligible": true,
    "exact_amount_currency_rate_min_0_95": true,
    "false_strong_count_zero": true,
    "confirmation_rate_eq_1_0": true,
    "p95_duration_ms_lt_5000": true
  },
  "gate_passed": true
}

A run is eligible only when:

runtime_policy == "conformant"
All cases ran on the retained reference runtime (conformant=true for every case)
All cases matched the expected baseline ID
No unexpected errors occurred

An ineligible run does not count toward the gate regardless of metric values.

Statistical caveat

Passing the gate (≥ 95% exact on ≥ 300 cases, 0 false-strong) establishes that the algorithm met the threshold on a specific frozen corpus under a specific retained runtime. It does not prove:

The true population error rate is below 5%.
Future images from the same apps will not produce false-strong results.
Performance will be identical on different runtime environments.

The gate is a necessary condition for v1, not a sufficient proof of production-readiness.

Supported environment non-guarantees

The v1 conformance environment is ubuntu-24.04 / amd64 / tesseract 5.3.4 with the pinned model files. Other configurations (other Linux distributions, macOS, ARM, other Tesseract versions) are supported for development but are not the conformance target and are not covered by the accuracy gate.

On this page