CATALOG PILOT / EVALUATION EVIDENCE

Test the contract
before trusting the claim.

The public example proves one report shape. A credible design-partner evaluation must test unpublished cases locally, fix its labels and thresholds before execution, and keep every artifact outside Andromalius.

01 / THREE PROOF LAYERS

Do not confuse a preview with generalization.

Each layer answers a different question. None creates an accuracy, customer, demand, certification, or delivery claim by itself.

  1. 01
    Fixed public preview

    Shows the stable report shape, finding code, severity, JSONPath, observations, and score-policy fields without a wallet. It proves one documented synthetic output—not general accuracy.

  2. 02
    Non-preview conformance corpus

    Exercises six materially different synthetic challenge and discovery cases with nine expected findings across nine stable codes. Exact outputs, registry versions, and zero network calls must remain deterministic.

  3. 03
    Customer-controlled blind evaluation

    Lets an evaluator seal one through 100 unpublished labeled cases and predeclare thresholds before running the checksum-pinned CLI locally. Cases, inputs, labels, and detailed results stay in the evaluator's environment.

02 / CURRENT SYNTHETIC EVIDENCE

Six non-preview cases. Zero network calls.

The private regression covers resource binding, timeout drift, hostile fallback options, duplicate handling, address normalization, and cross-operation discovery drift. It binds report, finding-code registry, and score-policy versions.

03 / BLIND EVALUATION

Evaluator-controlled cases. Predeclared acceptance.

The stronger test happens in the evaluator's environment after the exact private artifact, order, payment, checksum, recipient, and delivery gates are complete.

  1. 01
    Create and label locally

    The evaluator prepares bounded challenge or discovery cases and exact expected finding tuples under registry version 1.0.0. Andromalius does not receive the corpus.

  2. 02
    Seal before execution

    A canonical SHA-256 corpus identity, case count, contract versions, and exact acceptance thresholds are fixed in a separate protocol before evaluation.

  3. 03
    Run without a network

    The checksum-pinned customer CLI evaluates only explicitly named local JSON. Network use fails closed; there is no upload, telemetry, wallet, signing, payment, persistence, or automatic update path.

  4. 04
    Review aggregate evidence

    The report exposes exact-match, precision, recall, F1, blocking accuracy, false-block, and missed-block rates plus protocol, corpus, and result-set digests—without emitting cases or per-case findings.

04 / CLAIM BOUNDARY

Passing one sealed corpus means exactly that.

The evaluation is designed to make useful evidence possible without turning a synthetic result into a broad marketing claim.

05 / COMMERCIAL STOP

Evaluation interest is not an order.

There is no public download or self-serve checkout. A customer-local evaluation requires professional review, a written customer-specific order, reconciled payment, an exact checksum-pinned artifact, and a separately approved private recipient and delivery channel. Do not send an artifact with the initial inquiry.

Contacting us does not create an order, payment authorization, entitlement, software delivery, benchmark claim, or service-level commitment. Terms · Privacy · Acceptable use · Refunds