Migration
23. Current-to-Target Migration¶
The target is an evolution of the existing lifecycle architecture, not a wholesale rewrite.
23.1 Tolaria remains the training substrate¶
Legacy ordinary host training remains under Tolaria. Snapshotting, replay and flash cloning extend this authority; they do not replace or fork the training substrate.
Migration introduces explicit execution regimes:
- wrap the current deterministic training path as
ACADEMY_EXACT; - capture complete replay state and pass the bitwise gate on the supported reference profile;
- add repeated stochastic branch execution and calibration records;
- certify bounded
CALIBRATED_STOCHASTICprofiles against Academy; - enable
FIELDprofiles only after decision-aware gates pass; - retain Academy as the CI, causal-label, disputed-case and divergence-debugging path.
The target is not bitwise validation on every future production configuration. The target is one maintained exact reference profile and a measured path from that profile to realistic execution.
23.2 Controller rename and split¶
- The existing local/tactical policy responsibility moves to Narset.
- Tamiyo becomes the strategic controller over regions, budgets, capacity and long horizons.
- Transitional code may use an explicit name such as
LegacyTamiyoController, but the final API does not useTamiyofor the local controller. - Any prior allocator implementation migrates to Tamiyo’s target
StrategicEnvelopeinterface.
23.3 Observation, commissioning and candidate design¶
- Nissa becomes the canonical source of Momir's diagnostic input and publishes the same observation identity directly to Narset and Momir.
- Narset emits
GrowthIntent, not a complete design-bearingGrowthRequest. - Leyline and Kasmina deterministically resolve tensor, grammar and budget constraints.
- Fixed blueprint selection is removed from Narset's production action space.
- Momir becomes the generated candidate designer.
- Legacy Norm, Attention, Convolution and related seeds migrate from Kasmina's internal blueprint library into a versioned Sarpadian reference population and research-control package.
- Those reference seeds serve as temporary Momir ancestry and permanent experimental controls, not the production ontology.
- Compatibility aliases are permitted only during migration and must not reintroduce topology fields into
GrowthIntent.
23.4 Structural and compilation pipeline¶
- Elesh is inserted after design and before compilation.
- Tezzeret becomes the explicit compiler.
- Candidate identity is split into raw design, canonical semantic identity and executable artefact identity.
23.5 QA and adjudication split¶
Any existing combined screening/economy component is decomposed:
- Urabrask owns test planning, runtime checks, branch measurement, regression, uncertainty and
QualityReport. - Augustin owns hard eligibility, no-op anchoring, policy utility, admission, continued tenancy and warrants.
This split is mandatory. Renaming the old judge to Urabrask while leaving admission logic inside it does not satisfy the target design.
23.6 Kasmina and lifecycle¶
Kasmina retains host topology, reversible slots, gradient isolation, maturation, blending, commitment, decay mechanics and state serialization.
The change is constitutional:
- influence-increasing transitions require Augustin warrants;
- post-commit ordinary removal requires maintenance warrants;
- Narset and Emrakul have disjoint authority windows.
23.7 Memory and observability¶
- The static seed catalogue becomes a versioned Sarpadian reference population plus an episodic lineage-aware archive.
- Sarpadia records whether ancestry was supplied, withdrawn, or used only as a blinded control.
- Existing telemetry backends remain Nissa, with direct publication to Momir added as a locked route.
- Existing operator surfaces migrate under Oona.
- Oona gains a newsroom projection showing source observation, assignment, draft, standards, production, QA, judgement, placement, correction and archive.
- Event schemas remain in Leyline and training remains independent of Oona availability.
23.8 Namespec migration¶
Package moves and telemetry names are versioned. Compatibility aliases are time-limited and must be visibly marked as legacy.
The migration should include:
- an ADR locking Namespec 1.0;
- package-owner READMEs;
- import-linter rules;
- telemetry producer-name migration;
- persisted-record schema migration;
- dashboard label migration;
- and removal dates for old aliases.