Exact with respect to what?
For a supported parent decision graph and its declared numeric domain, exact compilation preserves the parent graph’s decision. The runtime consumes finite float32 features and follows strict comparisons: a feature greater than the threshold takes the high branch; equality takes the low branch. The relation is to that parent and input contract, not to ground truth about the physical world.
A correct exact copy of a poor classifier remains a poor classifier. Accuracy measures agreement with labels on a specified cohort; equivalence measures agreement with the source semantics. Teacher distillation that learns a new student is another quality question and must not be described as exact reproduction of an arbitrary neural network. Unsupported operators or numeric contracts require an adapter, not a stronger marketing claim.
A path is an executable explanation
A Reflex trace records the actual input, feature index, threshold, comparison result, node and chosen branch, ending at a class. These tests are why this graph produced that class for this event. They are not a post-hoc saliency estimate. You can replay the same float32 event and inspect the same decisions without guessing which retained predicate mattered.
Structural predicate usage is not statistical feature importance. A feature used by many nodes may be rare on real paths; a single test can matter for a safety-critical cohort. Node identifiers are meaningful only within their bound graph. Store the Reflex hash with a path record before comparing two devices or firmware revisions.
The standalone C predictor returns only a class ID. It exports no path-ID function. Browser explanations can walk the available graph, but a device logger needs a reviewed trace adapter if it is to publish real path IDs. Do not substitute class ID for path ID.
Perturb and replay the complete graph
Change one finite feature in Results to observe how the graph’s answer changes. A “flip this test” or threshold-neighbor probe should replay the complete graph after perturbation; changing a displayed branch alone would not establish a valid event. Other downstream predicates may change too, and a different path can still end at the same class.
/* Both operands follow the package float32 contract. */
next = feature > threshold ? high : low;
/* Equality belongs to low, not high. */A boundary probe is not a causal intervention in the real process and not necessarily the globally nearest counterfactual. Whether an altered feature vector can occur physically depends on acquisition and preprocessing. Record constraints such as correlated axes or normalized sums before treating a numerical edit as a plausible operating event.
Read the receipt, not just the badge
| Check | What it does not establish |
|---|---|
| Sample replay agrees | Universal equivalence or future task accuracy. |
| Compiler exact-domain receipt | Correct sensing, labels, preprocessing or application policy. |
| Host C probe replay | Nordic target timing, linked memory or PPK2 charge. |
| On-board class parity | Prospective recognition quality or customer battery life. |
A release receipt should bind source, graph, numeric domain, toolchain and exported runtime. Probe counts and zero mismatches are implementation checks; they cannot silently expand the compiler’s declared scope. Hashes check content identity but do not authenticate an owner or release without a trusted signature mechanism.
Use catalog receipts and replay to explore the public examples. Some published demonstration graphs carry a variant note rather than a claim that every viewed node is the timed artifact. For your own project, preserve the exact parent, feature/class manifest, generated package and verification result. A successful compile still leaves quality, target resources and production approval to be established separately.
Last reviewed .
Source of truth: published facts in content/public.ts; CURRENT_MEASURED_RESULTS.md; CONNECTED_BUILD_WORKFLOW.md; standalone C emitter and v3 endpoint contract. Reference examples are not physical-board test receipts.