Nexus Cerebri --- Testing, Visualization & Research Standard

Testing layers

1. Unit tests
2. Property-based tests
3. Fuzz tests
4. Scenario/golden tests
5. Integration/adapter contract tests
6. ML evaluation/regression tests

Safety invariants

• ValidatedPlan => HardConstraintViolations == 0
• ExecutableAction => PermissionGranted
• Execution => AuthorizedActionPlan
• Inference cannot replace Fact
• LearnedPreference cannot override ExplicitConstraint
• StalePlan => Revalidation
• Unknown != Assumed
• Planner != Executor

Invariant violations are critical defects.

Golden scenarios

Fixed medical appointment vs flexible activity; deadline pressure; external lock; participants/resources; timezone/DST; recurrence; dependency chain/cycle; ambiguous time; no solution; repair search; stale revision; unsupported adapter capability; permission denied.

From-scratch ML verification

Compare manual forward pass to reference, manual gradient to numerical gradient, and manual gradient to autograd. Keep deterministic seeded fixtures.

dL/dw ~= [L(w+eps) - L(w-eps)] / (2*eps)

Dataset quality

Measure class balance, duplicates/template leakage, vocabulary coverage, linguistic-family split, DE/EN distribution, hard negatives, ambiguity coverage, missing labels, dataset version/checksum.

Do not claim generalization from near-duplicate train/test templates.

Visualization requirements

Graphs must be readable, titled, labeled with units, interactive where useful, backed by inspectable raw data, accessible, usable in light/dark mode and not dependent on color alone.

Planner: timeline, free/busy, candidates, rejected candidates, constraints, score decomposition, repair path, trace.

ML: train/validation loss, F1/accuracy, confusion matrix, calibration, gradients, learning rate, latency, parameter/model comparison.

Dataset: labels, language, splits, lengths, hard negatives, versions.

Preferences: evidence/weights over time, ranking before/after, explicit/session/personal/global contribution.

Explanation modes

SIMPLE, TECHNICAL, RESEARCH.

Experiment record

Store run ID, date, hypothesis, Git commit, software/dataset/model versions, checksum, architecture, hyperparameters, seed, environment, metrics, plots, result, interpretation, limitations and next step.

Preserve baselines

Rules -> Mean Pooling -> MLP -> Sequence Encoder -> Attention -> Transformer should remain comparable. Complexity must earn its place through measured benefit.

Planner result semantics

Do not combine search proof and domain outcome in one enum.

PlanningOutcome: SOLUTION / NO_SOLUTION / NEEDS_RELAXATION / INSUFFICIENT_INFORMATION.

SearchAssessment: PROVEN_OPTIMAL / COMPLETE / BEST_FOUND.

PROVEN_OPTIMAL requires exhaustive evaluation/proof over the explicitly declared search space and objective. Wall-clock timeout results are never PROVEN_OPTIMAL.

Deterministic search tests

Prefer deterministic candidate/expansion/depth budgets in tests and research. Use stable tie-breakers: objective cost, mutation count, shifted duration, candidate start, stable ID.

Validator testing

Zero violations in validated plans is necessary but not sufficient. Include intentionally invalid plans and mutation/adversarial tests to verify the validator detects violations rather than merely withholding validation.

Foundation lifecycle tests

Verify that executor APIs cannot accept ProposedPlan, ValidatedPlan, or ActionPlan. Cover stale AuthorizedActionPlan, changed authorization, replay/idempotency, max_mutations = 0, scope None-vs-empty semantics, and multi-action partial failure.

COMPLETE requires full traversal of the declared search space without an optimization proof; PROVEN_OPTIMAL additionally requires proof under the declared objective/tie-break rules.