Skip to content

Architecture

7. System Context and Architectural Planes

The system operates inside nested learning and execution loops:

  • Tolaria continuously runs ordinary host training against the task and data stream;
  • Nissa publishes canonical observations at defined decision points;
  • Narset acts at local developmental decision points;
  • Tamiyo updates strategic envelopes at a slower cadence;
  • Momir, Elesh and Tezzeret construct candidate artefacts;
  • Urabrask defines QA plans and requests candidate trials in Tolaria;
  • Augustin adjudicates the certified reports;
  • Kasmina embodies admitted growth;
  • Emrakul manages committed structure under continued-tenancy warrants;
  • Sarpadia accumulates the complete causal history;
  • and Oona exposes the account without becoming part of the control path.
Plane Subsystems Purpose
Constitutional infrastructure Leyline Defines contracts, grammar profiles, compatibility, policy records and invariants
Training and execution infrastructure Tolaria Trains the live host and executes deterministic ordinary, replay and branch worlds
Historical infrastructure Sarpadia Retains lineages, reference ancestry, outcomes, failures and split-safe datasets
Strategic agency Tamiyo Allocates regional resources, permissions and risk over long horizons
Tactical commissioning Narset Decides whether and where to commission growth and manages pre-commit actions
Observation Nissa Publishes what the host is doing without prescribing what should be built
Synthesis Momir, Elesh, Tezzeret Designs, canonicalises and compiles growth
Assurance and adjudication Urabrask, Augustin Establishes empirical evidence, then judges it independently
Embodiment and maintenance Kasmina, Emrakul Introduces growth safely and removes obsolete committed structure
Witness Oona Exposes an auditable account without steering it

7.1 Logical architecture

flowchart TD
    DATA[Task and data stream] --> TOL[Tolaria: training and execution substrate]
    TOL --> K[Kasmina: host and reversible growth physiology]
    K --> N[Nissa: canonical ablated diagnostic observation]

    N -->|permitted coarse summary| TAM[Tamiyo: strategic controller]
    N -->|same TelemetryEnvelope| NAR[Narset: tactical assignments controller]
    N -->|same TelemetryEnvelope| MOM[Momir: candidate designer]
    TAM -->|StrategicEnvelope| NAR

    NAR -->|GrowthIntent: assignment brief| RES[Deterministic request resolution under Leyline]
    TAM -->|authorised envelope| RES
    K -->|RegionContract| RES
    LEY[Leyline: contracts and GrammarProfile] --> RES
    RES -->|Resolved GrowthRequest| MOM

    SAR[Sarpadia: history, lineage and bootstrap ancestry] -->|optional precedents / BootstrapAncestryContext| MOM
    MOM -->|RawGrowthGraph| EL[Elesh: verify and canonicalise]
    EL -->|CanonicalGrowthSpec| TEZ[Tezzeret: compile]
    TEZ -->|ExecutableGrowthArtifact| URA[Urabrask: QA]

    URA -->|TestPlan| TOL
    TOL -->|BranchResults and runtime evidence| URA
    URA -->|QualityReport| AUG[Augustin: judge]

    TAM -->|strategic limits| AUG
    RES -->|request context| AUG
    AUG -->|AdmissionDecision or NO_OP| NAR
    NAR -->|LifecycleCommand plus warrant| K

    K --> COM[Committed growth]
    COM --> EM[Emrakul: maintenance and lysis]
    EM -->|maintenance QA request| URA
    AUG -->|MaintenanceDecision| EM
    EM -->|sedate / decay / lyse command| K

    N --> SAR
    RES --> SAR
    MOM --> SAR
    EL --> SAR
    TEZ --> SAR
    TOL --> SAR
    URA --> SAR
    AUG --> SAR
    K --> SAR
    EM --> SAR

    N --> O[Oona: witness and operator surface]
    TAM --> O
    NAR --> O
    MOM --> O
    EL --> O
    TEZ --> O
    TOL --> O
    URA --> O
    AUG --> O
    K --> O
    SAR --> O
    EM --> O

The two Nissa arrows to Narset and Momir are independent publications of the same observation identity. Narset is not a telemetry proxy. The request resolver is deterministic contract assembly, not a fifteenth agent: it applies Leyline compatibility rules to Narset's intent, Tamiyo's envelope, and Kasmina's region declaration.

7.2 Ordinary training loop

Tolaria owns the operational loop that trains the host:

text load or create TrainingRunSpec ↓ materialise data stream and device topology ↓ execute Kasmina host forward pass ↓ compute task loss ↓ perform host optimiser step ↓ advance schedulers and logical time ↓ Nissa captures and publishes a canonical observation ↓ invoke permitted strategic/tactical decision points ↓ checkpoint, snapshot or continue

Tamiyo and Narset govern developmental actions around this loop. They do not perform SGD. Kasmina supplies the model and its growth mechanics. Tolaria performs the execution.

7.3 Observation and commissioning loop

```text Nissa observes Snapshot S at decision point T ↓ Nissa publishes TelemetryEnvelope O ├──→ Narset: should work be commissioned, where, and under what class? └──→ Momir: what structure would address this observed state?

Tamiyo supplies StrategicEnvelope E Narset emits GrowthIntent I referencing O and E Kasmina supplies RegionContract R Leyline supplies compatible GrammarProfile G ↓ pure request resolution: (O, E, I, R, G) → GrowthRequest Q ↓ Momir receives O and Q separately ```

Momir rejects a call where the observation, snapshot, region, or compatibility identifiers do not reconcile. A GrowthIntent never contains a copy of telemetry.

7.4 Candidate assurance and adjudication loop

text Tezzeret artefact ↓ Urabrask builds a blinded TestPlan ↓ Tolaria restores a common Snapshot and executes: candidate branches + controls + mandatory no-op ↓ Urabrask verifies runtime conformance and certifies measurements ↓ QualityReport ↓ Augustin applies eligibility, then the tail-risk veto, then utility policy ↓ ADMIT one / REJECT / DEFER / REQUIRE_RETEST / NO_OP

7.5 Control hierarchy

```text Tamiyo └── establishes StrategicEnvelope └── Narset chooses local actions ├── WAIT ├── COMMISSION_GROWTH → GrowthIntent ├── BEGIN_MATURATION ├── REQUEST_QA ├── REQUEST_ADJUDICATION ├── BEGIN_BLEND ├── HOLD ├── ABORT └── COMMIT

After COMMIT: Narset relinquishes ordinary ownership └── Emrakul manages physical maintenance ├── REQUEST_REVIEW ├── HOLD ├── SEDATE ├── DECAY └── LYSE ```

Urabrask and Augustin are independent of this command hierarchy. Urabrask certifies evidence. Augustin issues admission and maintenance warrants. Neither performs Kasmina state transitions. Tolaria is beneath the hierarchy as neutral execution infrastructure.

Emrakul's post-commit actions execute Augustin MaintenanceDecision verdicts; the vocabularies map one-to-one: RETAINHOLD (no physical transition), RETESTREQUEST_REVIEW (schedule maintenance QA), and SEDATE/DECAY/LYSE execute directly under the maintenance warrant. Verdicts are Augustin records; actions are Emrakul executions of them.

7.6 The newsroom authority model

The newsroom analogy provides an explanatory overlay, not a second architecture:

text Nissa reports source material Tamiyo funds the desk Narset commissions the assignment Momir authors the candidate Elesh applies standards and canonical form Tezzeret produces the executable edition Urabrask fact-checks and proof-tests it in Tolaria Augustin publishes, returns, defers, or spikes it Kasmina integrates it into the live edition Emrakul corrects, withdraws, or retires it later Sarpadia preserves the complete archive Oona presents the account Leyline supplies the editorial constitution

The analogy is useful because it makes an authority leak audible. “The assignments editor rewrote the source notes before the writer saw them” is the same defect as Narset mediating Nissa's evidence. Appendix E provides the full mapping and review prompts.

8. Subsystem Map

Subsystem Primary responsibility Key outputs Explicitly does not own
Leyline Typed contracts, schema versions, grammar profiles, request resolution, compatibility, policy record formats and ordering invariants Versioned records, validators, warrants, rule vocabularies and resolved constraints Case-specific policy, diagnosis or execution
Tolaria Host training and deterministic execution: data, optimisers, devices, precision, distributed scheduling, snapshots, replay, branches and rollback TrainingRunState, Snapshot, BranchResult, execution manifests Utility weights, candidate preference or verdicts
Sarpadia Append-only history, lineage, bootstrap reference population, outcomes, retrieval and split-safe datasets GrowthRecord, BootstrapAncestryContext, retrieval sets, lineage graphs and blinded views Live-host mutation or self-approval
Tamiyo Strategic budgets, regional priorities, exploration quotas, cooldowns and long-horizon risk StrategicEnvelope Local action timing or candidate choice
Narset Tactical commissioning and pre-commit lifecycle actions GrowthIntent, LifecycleCommand, review requests and escalations Diagnostic captions, ancestor choice, topology hints, unallocated budget or structural generation
Nissa Ablated host telemetry, direct evidence publication and diagnostic provenance TelemetryEnvelope should_grow, structural recommendations, reward or verdicts
Momir Raw candidate topology, parameters, mutation and recombination RawGrowthGraph Intervention timing, structural approval, testing or admission
Elesh Static verification, canonicalisation, semantic hashing and structural legality CanonicalGrowthSpec, structural reports Runtime QA, task utility or lifecycle policy
Tezzeret Lowering, kernel selection, fusion, memory planning and compilation ExecutableGrowthArtifact, compilation manifest Semantic topology changes, QA or admission
Urabrask Dynamic QA, runtime conformance, regression, branch measurement and evidence certification TestPlan, QualityReport Admission, no-op policy or lifecycle commands
Augustin Provider-blind admission and continued-tenancy adjudication AdmissionDecision, MaintenanceDecision, warrants Test execution, compilation or host mutation
Kasmina Host model, insertion-region declarations, slots, gradient routing, maturation, blending and physical lifecycle RegionContract, embodied state and lifecycle events Stock blueprint ontology or whether a growth deserves admission
Emrakul Safe post-commit sedation, decay, consolidation and lysis Maintenance requests and warranted lifecycle commands Candidate construction or judgement
Oona Event projections, flight recorder, TUI/dashboard adapters, audit bundles and alerts Operator views and audit records Training control or source-of-truth schemas

The deterministic request resolver is a Leyline application service operating over immutable GrowthIntent, StrategicEnvelope, RegionContract, and GrammarProfile records. It performs no diagnosis and has no learned policy.


10. End-to-End Operating Flow

10.1 Ordinary host training in Tolaria

Tolaria materialises the TrainingRunSpec, constructs the data and device environment, and advances the Kasmina host through ordinary optimisation.

It owns:

  • batch acquisition and deterministic ordering;
  • host forward and backward execution;
  • optimiser and scheduler stepping;
  • precision and device policy;
  • distributed execution;
  • checkpointing;
  • logical time;
  • callback and decision-point scheduling;
  • and measured execution spend.

Tamiyo is not “the trainer” in this sense. It is a strategic controller around a host that Tolaria trains.

10.2 Strategic allocation

At a slow cadence, Tamiyo consumes coarse host health, regional histories, current capacity, recent interventions and strategic objectives. It emits a StrategicEnvelope specifying:

  • which regions may grow;
  • how much parameter and compute capacity each may consume;
  • maximum concurrent growth;
  • acceptable risk and churn;
  • exploration allowances;
  • permitted grammar profiles;
  • and cooldowns or embargoes.

Tamiyo may reserve capacity for anticipated future pressure, but it cannot select an immediate candidate, choose an ancestor, or issue a blend command.

10.3 Local observation and direct publication

Nissa observes the host at each local decision point. Germination telemetry is taken from the ablated path, so the condition describes the deficit without assistance from an existing growth.

Nissa constructs one canonical TelemetryEnvelope bound to the current Tolaria snapshot and publishes that same observation identity independently:

text Nissa ├── TelemetryEnvelope O ──→ Narset ├── TelemetryEnvelope O ──→ Momir ├── permitted summary ───→ Tamiyo ├── append-only record ───→ Sarpadia └── event projection ───→ Oona

Nissa emits measurements and provenance, not conclusions. Narset is not permitted to produce a second “designer telemetry” object.

10.4 Tactical commission

Narset receives Nissa's telemetry, the active StrategicEnvelope, Kasmina local lifecycle state, prior Augustin decisions, and compact operational history. It chooses among legal local actions:

text WAIT COMMISSION_GROWTH CONTINUE_MATURATION REQUEST_QA REQUEST_ADJUDICATION BEGIN_BLEND CONTINUE_BLEND HOLD ABORT COMMIT ESCALATE_TO_TAMIYO

A COMMISSION_GROWTH action creates a narrow GrowthIntent. Narset chooses whether intervention is warranted, which permitted insertion region owns the assignment, and what operational class applies. It does not select a seed family, ancestor, rank, width, operator, or diagnosis for Momir.

10.5 Deterministic request resolution

A pure resolver operating under Leyline combines:

text GrowthIntent + StrategicEnvelope + Kasmina RegionContract + active GrammarProfile = GrowthRequest

The resolver:

  • validates the intent against Tamiyo's active authority;
  • obtains the region's tensor contract from Kasmina;
  • selects only a grammar profile already permitted by the envelope and supported by Elesh and Tezzeret;
  • derives exact bounded budgets from the requested resource class;
  • binds observation, host state, snapshot, region and schema identities;
  • canonicalises representation and field ordering;
  • and fails closed on incompatibility.

It does not inspect task utility, choose a candidate family, or infer a structural solution.

10.6 Candidate design

Momir receives independent inputs with preserved provenance:

text TelemetryEnvelope from Nissa GrowthRequest from deterministic resolution under Leyline ProposalBatchRequest from orchestration BootstrapAncestryContext | null from Sarpadia during the bootstrap curriculum ordinary retrieval context from Sarpadia where enabled

Momir may design:

  • faithful reconstructions during structural literacy;
  • bounded mutations of successful or failed ancestors;
  • recombinations of compatible lineages;
  • fresh candidates;
  • deterministic proposals;
  • or stochastic best-of-\(K\) pools.

Research controls such as stock Norm, Attention, Convolution, random, analytic, retrieval, online-optimised or human-designed candidates enter through the same downstream contracts and remain blinded during QA and judgement.

Momir never reads Narset hidden state or a Narset-authored diagnostic summary. Its output must remain valid when bootstrap ancestry is absent.

10.7 Structural conformance

Elesh performs:

  • graph grammar validation;
  • tensor shape inference and alignment;
  • insertion-contract checks;
  • cycle and reachability checks;
  • gradient-flow analysis;
  • zero-influence verification;
  • static parameter and memory accounting;
  • forbidden-operation detection;
  • dead-node elimination;
  • semantics-preserving graph simplification;
  • deterministic ordering;
  • canonical parameter layout;
  • equivalence-class detection;
  • and semantic hashing.

A failed candidate receives a structured rejection report and is stored in Sarpadia. A successful candidate becomes a CanonicalGrowthSpec.

Elesh may simplify only where semantic equivalence is established. Predicted usefulness is not structural equivalence.

10.8 Compilation

Tezzeret lowers each canonical specification for the target Tolaria runtime.

It may perform:

  • kernel selection;
  • operation fusion;
  • layout planning;
  • memory planning;
  • device-specific lowering;
  • eager or compiled realisation;
  • and asynchronous build where supported.

It emits an ExecutableGrowthArtifact plus a compilation manifest linking it to the canonical semantic hash.

Compilation success does not imply runtime validity or admission.

10.9 QA planning and execution

Urabrask creates a blinded TestPlan covering:

  • canonical-to-artefact runtime equivalence;
  • finite outputs and gradients;
  • gradient-routing conformance;
  • zero-influence behaviour;
  • determinism;
  • budget adherence;
  • immediate task effect;
  • short- and longer-horizon trajectory effect;
  • integration shock;
  • regression tests;
  • and mandatory no-op comparison.

Tolaria restores the complete snapshot and executes the requested branches over identical future data.

Urabrask certifies the returned evidence into a QualityReport. It may mark hard defects, incomplete evidence and uncertainty. It does not issue a verdict.

10.10 Independent adjudication

Augustin receives a blinded QualityReport, the resolved request, the active StrategicEnvelope, and the applicable policy version.

Adjudication is lexicographic (ADR-0004): three ordered stages, no later stage reopening an earlier one.

Stage 1 — hard eligibility:

  • structural and semantic identity valid;
  • compilation conformance valid;
  • runtime and gradient checks passed;
  • required determinism passed;
  • evidence complete enough for the assurance class;
  • and measured spend inside declared budget.

Stage 2 — tail-risk veto: per candidate, Augustin estimates the tail of the intervention-outcome distribution from certified evidence and vetoes any candidate whose estimate breaches the threshold. The veto is not tradeable against measured benefit (INV-45); the assurance class owns the operating point, priced against current snapshot distance. No-op never faces the veto.

Stage 3 — utility competition: Augustin computes policy utility, expected-risk and uncertainty charges for the survivors relative to no-op. It may:

text ADMIT NO_OP REJECT DEFER REQUIRE_RETEST

A growth is not entitled to publication merely because Narset commissioned it.

10.11 Germination and maturation

On admission, Narset issues a warranted command to Kasmina.

Kasmina:

  • verifies the Augustin warrant;
  • verifies the candidate semantic hash and artefact identity;
  • instantiates the growth at zero influence;
  • configures one-shot or nursery mode;
  • isolates forward influence during maturation;
  • applies the declared trainability mask;
  • and records lifecycle state.

Tolaria executes any maturation steps. Maturation spend is charged to the growth.

10.12 Integration and commitment

Before blending, a nursery-matured growth is re-qualified against the current host state. If host drift has invalidated it, Narset aborts or requests fresh QA rather than integrating stale growth.

Kasmina applies the authorised blend schedule. Narset may hold, continue, reverse, abort or commit within its pre-commit authority and the Augustin warrant.

At commitment, ordinary ownership transfers from Narset to Emrakul.

10.13 Continued-tenancy review and maintenance

Emrakul schedules periodic or event-driven maintenance review.

Urabrask requests counterfactual evidence in Tolaria, including separate no-growth or re-adaptation branches where required. Augustin issues a MaintenanceDecision:

text RETAIN RETEST SEDATE DECAY LYSE

Emrakul executes the authorised safe physical transition through Kasmina. It cannot change the verdict or generate a replacement.

10.14 Memory and witness

Sarpadia stores:

  • direct Nissa observation identity;
  • Tamiyo envelope;
  • Narset intent;
  • resolved request;
  • proposal-batch metadata;
  • optional bootstrap ancestry;
  • raw candidates and lineages;
  • Elesh reports;
  • Tezzeret manifests;
  • Tolaria snapshots and branches;
  • Urabrask plans and reports;
  • Augustin decisions;
  • lifecycle traces;
  • spend and rent;
  • determinism manifests;
  • and terminal outcomes.

Oona presents the same causal chain to operators. It may show the newsroom view—source, assignment, draft, standards, production, fact check, publication decision, placement, correction and archive—but it remains read-only.