# Public API and streaming command contract ## PreparedBank ```python PreparedBank(value_lower, value_upper=None, *, backend='auto', base_numerator=0, base_denominator=0) bank.evaluate(mass_lower, mass_upper, threshold, *, relation='<=', base_numerator=None, base_denominator=None, witness=False, exact_margins=False) ``` Prepare static values once. Shapes are `(M,)` for scalar values or `(M,P)` for P coordinates. Absent `value_upper` means fixed values. Mass bounds are vectors of length M. Scalar thresholds broadcast to all coordinates; vector thresholds must have exactly P entries. A length-one vector is not a scalar for P>1. Python numeric iterables and ordinary real numeric NumPy arrays are accepted; booleans, strings, complex and nonfinite numeric inputs are rejected. Numeric values convert to finite binary64 before the mathematical contract applies. Masses and the base denominator are nonnegative, lower bounds must not exceed upper bounds and every evaluated box needs `D0 + sum(mass_lower) > 0` exactly. Signed values and base numerators are allowed. Base numerator can be a scalar or length-P vector. A zero-row bank requires a strictly positive base denominator; its coordinate count is inferred from the base numerator. Python `None` base overrides retain bank defaults; per-call overrides never modify those defaults. | Backend | Selection and guarantee | |---|---| | `exact` | Standard-library integer arithmetic over the exact converted binary64 values | | `filtered` | Requires NumPy; accepted separated signs require the explicit IEEE/NumPy premises in FILTER_THEORY; unsafe/ambiguous cases fall back to integer arithmetic | | `auto` | Chooses filtered if NumPy is available, otherwise exact, once at construction; inspect `bank.backend` | Importing `qzguard` does not import NumPy. Applications with a fixed numerical contract should choose a backend explicitly. The CLI default is `exact`. ## Evaluation outputs | Field/method | Meaning | |---|---| | `guaranteed` | Tuple of universal inclusive-predicate booleans, one per coordinate | | `all_guaranteed` | All coordinate predicates are guaranteed | | `counterexample_exists` | Coordinatewise negation of `guaranteed` | | `backend` | Selected bank backend | | `relation`, `thresholds` | Requested relation and converted threshold snapshot | | `modes` | Per-coordinate decision routes, `filtered` or `exact` | | `support_interval` | Optional outward floating enclosures of centered supports; not weighted-average intervals or exact margins | | `exact_result` | Exact result if already computed or explicitly requested; otherwise `None` | | `.exact(witness=False)` | Explicitly compute exact rational supports and optional separate-coordinate counterexamples | False means a violating configuration exists in the declared independent box. It does not mean every configuration violates the threshold. For relaxed boxes, a counterexample can be impossible under correlations omitted from the model. Different coordinates can require different witnesses. A support-attaining witness is not necessarily a ratio-extremizing witness. `witness=True` forces exact work. `exact_margins=True` also charges exact work; if the original sign was filtered, `modes` still describes that sign decision. If one coordinate needs fallback, the current filter computes all exact coordinates. Route counts are not operation counts. Each Evaluation retains its mass snapshot: release old results in a long-running stream rather than accumulating O(number_of_results * M * P) state. For one unchanged box and many thresholds, use `NormalizedBox.enclose()` once and `guard_from_bounds()` for subsequent thresholds. PreparedBank is designed for new mass boxes. No claim covers concurrency of lazy exact-cache requests; separate snapshots/processes provide an uncomplicated application boundary. ## CLI ```bash qzguard --version qzguard doctor --backend exact qzguard demo --backend filtered qzguard check --input examples/json/check.json qzguard check --bank examples/json/bank_exact.json \ --updates examples/json/updates.jsonl --output answers.jsonl qzguard check --bank examples/json/bank_fast.json --updates - \ --exact-margins --witness ``` The source equivalents are `python START_HERE.py ...` and `python -m qzguard ...` from the release root. Installed equivalents work outside that root. A bank is one strict JSON object with required `value_lower` and optional `value_upper`, `backend`, `base_numerator`, `base_denominator`. Each update is one strict JSON object with required `mass_lower`, `mass_upper`, `threshold` and optional `relation`, base overrides, `witness`, `exact_margins`, `id`. `relation` defaults to `<=`; flags must be actual booleans. `id` is returned unchanged for valid updates. With `--input`, combine bank and update fields in one object. `--backend` overrides the input choice. `--witness` and `--exact-margins` enable details for all updates. Duplicate keys, unknown fields, nonfinite JSON numbers, numeric strings and booleans in numeric fields are rejected. Omit optional fields to retain defaults; explicit JSON null in numeric fields is invalid. Blank JSONL lines are skipped. The bank is prepared once for the stream; each line replaces only masses and optionally bases. Standard input uses `-`; bank and updates cannot both consume stdin. Output cannot alias an input file. Successful records contain `status='ok'`, `backend`, `numerical_contract`, a feasible-set label, dimensions, relation, thresholds, booleans, route counts and the optional identifier. The feasible-set label is: ```json {"semantics":"universal_over_independent_box"} ``` JSONL records also carry the source line number. Filtered results include `support_interval`. Exact detail requests add `support`, `decision_margins`, counters and optionally `witnesses`. Decision margins are signed so nonnegative means the requested relation is guaranteed; `support` is the raw centered support. Arbitrary-size rational components use decimal **strings**: ```json {"numerator":"100","denominator":"101"} ``` This avoids silent integer loss in JavaScript consumers. Outward support infinities use the strings `"-inf"` and `"inf"`; inputs must still be finite. No NaN/Infinity JSON extension is emitted. Valid false decisions return exit code 0. Invalid streamed lines emit explicit `status='error'` records, processing continues, and the command returns code 2. A fatal input/import/file error reports a concise message on stderr and returns 2. Doctor returns 1 if its deterministic checks fail. Summaries go to stderr; JSON results alone go to stdout or the requested output. Fatal setup errors are checked before creating/truncating output; successful `--output` operations replace that file. A stream containing errors can therefore leave partial results, clearly labeled by status and exit code. `doctor` checks format and deterministic basic/boundary fixtures. It does not prove that an arbitrary CPU/compiler runtime fulfills every filter premise. The full test suite and research evidence are bundled separately from the installed wheel.