Phases
24. Minimum Viable System¶
The first coherent implementation contains:
- one host architecture;
- one insertion region;
- one reversible growth slot;
- one universal residual growth envelope;
- Tolaria ordinary host training through one deterministic step engine;
- exact Tolaria snapshot, restore, branch and common-future replay under one Academy reference profile;
- one explicit three-axis
ScaffoldStateand scaffold manifest registry; - a calibrated-stochastic harness capable of measuring Field-to-Academy disagreement, even if full Field operation remains disabled;
- one fixed Tamiyo strategic envelope;
- one heuristic Narset tactical controller;
- one Nissa telemetry schema published directly to Narset and Momir;
- one narrow Narset GrowthIntent and deterministic request resolver;
- one versioned Sarpadian stock-reference bootstrap corpus;
- one small deterministic or latent-conditioned Momir designer that also runs with ancestry absent;
- one rule-driven Elesh verifier and canonicaliser;
- one eager-mode Tezzeret compiler with explicit manifests;
- one rule-driven Urabrask QA suite with mandatory no-op measurement;
- one fixed, provider-blind Augustin judge;
- fixed Kasmina maturation and blend schedules;
- fixed Emrakul safe-maintenance rules;
- Sarpadia retention of complete candidate pools, evidence and decisions;
- Oona flight recording and branch inspection;
- reference-seed bootstrap, scaffold-withdrawal, static and short-horizon counterfactual curricula;
- and random, analytic, retrieval, online-optimised and no-op controls.
The MVP does not require:
- learned Tamiyo;
- learned Augustin;
- learned Emrakul;
- arbitrary graph generation;
- asynchronous CUDA code generation;
- diffusion;
- multiple insertion regions;
- or image-scale tasks.
25. Implementation Sequence¶
Phase A — Namespec, Leyline and dependency boundaries¶
- record Namespec 1.0 in an ADR;
- define package ownership and forbidden authority;
- define all core contracts;
- encode lifecycle and warrant rules;
- implement budgets and spend records;
- establish versioning and compatibility;
- and add import-lint and authority tests.
Phase B — Tolaria host-training baseline and Academy profile¶
- move ordinary host execution behind one Tolaria engine;
- make data, optimiser, scheduler, precision and device state explicit;
- define the Academy-exact runtime profile;
- establish deterministic mainline traces;
- implement
ScaffoldManifestandScaffoldStaterecording; - and ensure Kasmina exposes a neutral host-runtime protocol.
Phase C — Kasmina, Elesh and Tezzeret mechanics¶
- implement the universal growth envelope;
- define raw and canonical graph IRs;
- implement structural validation and canonical hashing;
- implement eager compilation and manifests;
- and prove raw-to-canonical-to-artefact identity.
Phase D — Tolaria replay, branching and execution calibration¶
- capture host, optimiser, lifecycle, controller, RNG, dataloader, task and future state;
- implement Academy-exact restore and common-future replay;
- implement branch adoption or validated replay;
- pass the Academy determinism gate;
- add repeated calibrated-stochastic branches;
- measure ranking, decision and tail disagreement against Academy;
- and keep Field profiles disabled until the declared gate passes.
Phase E — Urabrask QA¶
- implement
TestPlan; - implement runtime semantic and gradient conformance;
- implement numerical, determinism and regression checks;
- implement multi-horizon measurements;
- produce signed
QualityReport; - establish Academy versus Field QA;
- and implement escalation from uncertain Field evidence to Academy-exact retest.
Phase F — Augustin adjudication and controls¶
- implement hard eligibility;
- implement mandatory no-op policy;
- implement provider-blind utility and risk;
- add screen and independent audit rules;
- bind warrants to evidence;
- add random, analytic, retrieval and bounded online controls;
- and produce the QA-cost and adjudication-regret curves.
Phase G — Sarpadia and Momir bootstrap curriculum¶
- migrate legacy stock blueprints into a versioned reference population outside Kasmina;
- store complete pools, reports, decisions and no-op cases;
- implement blinded views, lineage, equivalence and ancestry records;
- collect frozen-state teachers, parents, mutations and failures;
- train structural reconstruction and behavioural prediction;
- train local mutation, recombination and ancestry dropout;
- pass Momir's independent design-prior withdrawal gate;
- retain stock seeds as blinded controls after ancestry withdrawal;
- and consume losers through ranking or utility objectives.
Phase H — Nissa routing and Narset lifecycle integration¶
- publish one Nissa observation directly to Narset and Momir;
- replace blueprint actions with
GrowthIntent; - implement deterministic
GrowthRequestresolution; - train local commissioning and lifecycle behaviour with known-good candidates;
- run anti-collusion and assignment-brief tests;
- enable generated requests;
- and add nursery re-qualification.
Phase I — Emrakul maintenance¶
- implement continued-tenancy QA;
- implement Augustin maintenance decisions;
- calibrate sedation, decay and lysis execution;
- and separate host dependence from intrinsic value.
Phase J — Tamiyo strategic allocation¶
- introduce multiple regions or cells;
- allocate global resources;
- and train or search strategic policies after local behaviour is stable.
Phase K — Oona and scale¶
- complete event projections and audit bundles;
- expose architecture-smell events;
- expand the grammar;
- add image tasks;
- and scale only after synthetic reliability gates pass.