Worked example
Inside an NDC compile
Review one recorded compilation: its inputs, discovered structure, validation results and published report files.
Public download: eight report files. Input trajectories and executable model files are not included.
1 · What went in
The data that was compiled
48 training rows · 16 validation rows
A short multi-channel trajectory set across three channels (A, B, C). The compiler sees only the trajectories — no equations, no labels. The raw frames are redacted here; only their SHA-256 travels with the artifact so the input identity is fixed without exposing the signal.
The input has a verifiable identity
request 345d754f…c1b96f6
train dataset 3cc72b6f…d76f41e8
validation dataset 4763f9e2…069e1be4
2 · What the compiler did
From trajectories to a runnable model
observed trajectories → GP-native Channel A/B/C readouts → robust coefficient estimation → Passport / runtime
428 GP-native readouts · 409 active terms
The report lists 428 readouts and 409 active terms, with coefficients estimated over a 48×428 feature matrix. Structure and mixing choices also used validation information. This is not a coefficients-only selection on training data.
Exhaustive finite-grammar search
The report labels its search coverage exhaustive_finite_grammar. That describes the declared search space, not a universal completeness or reproducibility guarantee independent of configuration, version and input data.
| Channel | What was discovered |
|---|---|
| A | Hybrid local-slot policy (self-copies 3, lag 2, neighbour 1); self-power max degree 3; adaptive cascade depth 2; a learned rotor policy (objective 1.010495 → 1.009485 over 3 optimizer steps). |
| B | Grade-resolved, validation-aware edge discovery: 5 candidate edges, all 5 accepted against validation, 0 rejected; 5 multi-lag edges; one group wormhole. |
| C | Full-rank mixing, 4 modes, one connected global cascade. |
3 · Before → after
Measured in free-run, not one-step
Free-run feeds model predictions back after the initial history, with allowed inputs. In this example, validation free-run also selected the persistence blend. Its 0.9160 → 0.7674 improvement is selection-set performance, not an untouched independent test.
| Metric | Before | After |
|---|---|---|
| Free-run NRMSE (validation used in selection) | 0.9160 | 0.7674 |
| Rotor objective | 1.010495 | 1.009485 |
Residual whiteness 0.929
The report marks this diagnostic above its threshold. This is not proof of independent residual noise, calibrated uncertainty or general model accuracy.
Runtime parity 16 / 16
The report records agreement on 16 checked rows between exported runtime and compiler predictions. This parity result is not an independent accuracy test, and this public report subset cannot replay the predictions.
Refusal is on
Unsupported-extrapolation refusal is active: outside the support the model declines rather than inventing an answer. Identification status is honestly reported as UnobservableDirection (active numerical rank 48 / 409).
4 · The explanation
Every model ships a Passport
The Passport is the human-readable account of what the model is and where it applies. An excerpt, verbatim:
passport_version: ndc-gp-native-passport/v2 active GP-native readouts: 428 active terms: 409 estimator: robust linear estimator over GP-native readouts Validation and deployment unsupported extrapolation refusal: true runtime predictions match compiler predictions: true export-safe runtime: true Identifiability, support, residuals, uncertainty identification status: UnobservableDirection active numerical rank: 48 / 409 residual whiteness score: 0.928624 (white enough: true) runtime parity passed: true; checked rows: 16 Order, memory and graph semantics graph semantics: predictive_interaction_graph_not_causal_graph
The Passport states graph semantics plainly — a predictive interaction graph, explicitly not a causal one. Full file: passport.md.
5 · Proof, not faith
Hashes you can re-check
The original manifest describes the full delivery, including files not in this public subset. Use PUBLIC_SUBSET_CHECKSUMS.sha256 to verify the eight downloadable originals; hashes do not establish prediction quality.
| Artifact | SHA-256 |
|---|---|
| model (runnable) | 9aaea132…c2485d7 |
| passport.ndc.json | c69dd3a4…2fbbb739 |
| passport.md | 52aaf248…f6989261 |
| approximation_graph.json | 49b51cf1…b934e2af |
| manifest.json | 9b9add61…38fe8ebb |
Policy gate on this artifact: allowed, civil surface, closed-loop control API disabled — the compiled model is advisory, it does not actuate.
6 · Take the evidence
Download the public report files
The provenance and explanation files from this exact model — nothing mocked:
- passport.md — the Nonlinearity Passport
- compile_report.json — what the compiler emitted
- manifest.json · checksums.sha256 — provenance
- compile_request.redacted.json — the input identity (frames redacted)
- policy_gate.json · compiler_boundary.json — safety & export boundary
- approximation_graph.json — the discovered structure
The public download is a report/provenance subset, not a runnable-model replay. Model-file hashes below identify files from the original larger delivery. A customer compile has its own model, access rights, runtime requirements and validation record.
Download all eight report files together: NDC public report ZIP