Nexus Cerebri --- Development & Learning Roadmap
Baseline: 2026-09-19
Initial planning horizon: 2026-10-01 through 2027-09-30
Dates are planning targets. Re-estimate the next 4--8 weeks at each monthly review. Reserve roughly 20--30% of development capacity for learning, debugging, refactoring, documentation and unexpected work.
October 2026 --- Foundation --- target v0.1.0
Learn: Rust workspace architecture, errors, serialization, CI/test basics.
Build: workspace/crates, IDs/version types, error model, CPIR skeleton, config/tracing, in-memory ports, API/Node placeholders, Cerebri Lab shell.
Docs: README, CONTRIBUTING, SECURITY, ROADMAP, CHANGELOG, ADR template, DE/EN structure, progress log, GitHub Pages skeleton.
Verify: CI build/test/lint/format and first property-test harness.
November 2026 --- Temporal Core --- target v0.2.0
Learn: interval algebra, timezone/DST behavior, recurrence boundaries.
Build: TimeRange, Duration, ZonedDateTime, relations, overlap/intersection, free/busy, PlanningHorizon, granularity, buffers, bounded recurrence.
Visualize: interactive timeline, free/busy ranges, DST/timezone debug.
Verify: interval properties, timezone/DST golden cases, malformed input fuzzing.
December 2026 --- Deterministic Planner --- target v0.3.0
Learn: constraint satisfaction, dependency graphs, objective functions.
Build: hard constraints, cycle detection, candidate generation, weighted scoring, ranking, ConflictSet, explanations, SearchBudget v0.
Visualize: candidates, rejection reasons, constraint status, score decomposition.
Verify: hard-constraint invariants, randomized schedules, golden scenarios.
January 2027 --- CPIR/API/Integration --- target v0.4.0
Learn: schema evolution, API contracts, idempotency, optimistic concurrency.
Build: CPIR v1, operations/scope, ProposedPlan/ActionPlan, action states, idempotency, stale checks, Rust API, REST skeleton, usable Node binding sample, adapter capabilities.
Verify: serialization round trips, schema compatibility, adapter contract tests, stale-plan scenarios.
February 2027 --- Semantic Baseline --- target v0.5.0
Learn: features, vectors, uncertainty, classification/calibration intuition.
Build: semantic vector, provenance/confidence, rule/signal extractor, source layers, ambiguity model, explanation modes, DE-first fixtures with EN equivalents.
Visualize: semantic bars, evidence tree, confidence, signal contributions.
Verify: hard negatives, negation/context, bilingual fixtures, ambiguity.
March 2027 --- Data & ML Foundations --- target v0.6.0
Learn: vectors/matrices, logits/probability, splits/leakage, precision/recall/F1.
Build: dataset schema/versioning, synthetic generator, linguistic variation, challenge set, experiment metadata, baseline rule classifier.
Visualize: class/split/language distributions, confusion matrix, example browser.
April 2027 --- Neural Network From Scratch --- target v0.7.0
Learn deeply: Linear, activations, Softmax, Cross-Entropy, derivatives, Gradient Descent, Backpropagation.
Build educationally: required matrix ops, embeddings, mean pooling, Linear, ReLU, Softmax, loss, SGD, backward pass, numerical gradient checks.
Visualize: activations, logits, probabilities, loss, gradients, weight updates.
Verify: manual vs numerical gradients and reference framework.
May 2027 --- Production Intent Intelligence --- target v0.8.0
Learn: framework/autograd, hyperparameters, overfitting, calibration.
Build: framework model, model registry, immutable artifacts, TrainingRun, production inference interface, UNKNOWN/clarification policy, regression suite.
Compare rules vs from-scratch vs framework.
June 2027 --- Preference Learning --- target v0.9.0
Learn: ranking, pairwise preferences, regularization, recency/decay.
Build: FeedbackEvent, PreferenceEvidence, interpretable PreferenceProfile, pairwise records, simple learning-to-rank experiment, global/personal/session separation.
Visualize: evidence and weights over time, before/after rankings, source contribution.
July 2027 --- Repair Planner --- target v0.10.0
Learn: heuristic search and complexity.
Build: Tier-2 repair candidates, minimal-disruption repair cost, bounded multi-event replanning, dependency propagation, BEST_FOUND/OPTIMAL semantics, benchmarks.
Visualize: repair paths, budget, moved-event cost, best-so-far progression.
August 2027 --- Attention Research --- target v0.11.0
Learn: sequence representations, Q/K/V, scaled dot-product and multi-head attention.
Experiment: sequence-aware baseline and attention; compare against mean pooling, especially word order/negation.
Visualize: token representations, attention matrices, failure comparisons.
September 2027 --- Small Transformer Research --- target v0.12.0
Learn: Transformer encoder, residuals, normalization, multi-task heads.
Experiment: small Transformer, intent head, additional semantic/category heads only where data supports them. Measure complexity vs benefit and document whether adoption is justified.
Monthly review
At month end: 1. compare completed vs planned work; 2. record learning outcomes; 3. review tests/metrics; 4. verify DE/EN documentation parity; 5. review ADRs and technical debt; 6. re-estimate next two months; 7. update README/ROADMAP; 8. release only if Definition of Done is satisfied.
Release gating
Monthly candidate versions are sequencing labels, not calendar promises. A release occurs only when:
- milestone acceptance criteria are met;
- required tests and docs are green/current;
- unresolved critical safety issues are zero;
- any scope reduction is recorded in ROADMAP/CHANGELOG.
Do not skip a version merely because a calendar month elapsed.
Search reproducibility rule
Research/test runs use deterministic search budgets whenever possible. Production may add wall-clock timeout limits, but any timeout-limited solution is recorded as BEST_FOUND.