Super Φ.Vessel is a governed cognitive operating layer that sits between a human operator, AI model capabilities, tools/bridges, persistent state, and executable effects.
The architecture is intentionally built around separation of concerns rather than assuming that a capable model should also be the authority boundary.
Human Operator
│
▼
Intent / Work Context
│
▼
Governance + Authorization
│
▼
Routing / Cognitive Roles
│
▼
Model / Tool Capability
│
▼
Evidence + Verification
│
▼
Effect Boundary
│
▼
Receipts / Ledger / Persistent State
- CAPABILITY ≠ AUTHORITY
- GOAL ≠ PERMISSION
- INFERENCE ≠ OBSERVATION
- PROPOSAL ≠ EXECUTION
- retrieval ≠ prompt admission
- readable page ≠ usable service
- usable service ≠ established session
- established session ≠ proven interaction
- evidence transport does not transport authority
The standalone runtime has evolved around subsystem families such as:
- Conversation / Work context — bounded user/model work state
- Crane Fly / Brain Registry — governed routing and executor selection
- Dreamer / Translator / Builder / Ledger — role-separated cognitive chambers
- Reality Gate — deterministic and policy review before promotion
- Promotion / Evidence Fabric — immutable claim/evidence custody
- Memory — bounded Vessel-scoped context with explicit admission rules
- Browser / Research lanes — observation separated from interactive authority
- PhiOS bridge — proposal/context handoff separated from execution leases
- System Map / Inspector — read-only runtime truth surfaces
- Labs / instruments — bounded experimental compute surfaces
- Chamber Fitness / Challenger Bench — evidence for human model-routing review, not automatic promotion
Documentation should classify subsystems using explicit states such as:
- LIVE
- CONDITIONAL
- EXPERIMENTAL
- LOCKED
- LEGACY
- UNWIRED
The presence of a subsystem in source or documentation must never be interpreted as evidence that it is active or authorized in the current runtime.
Until the runtime is modularized, the current standalone HTML artifact is the source of truth for shipped behavior. It belongs under runtime/.
This document is architectural orientation, not an executable specification. Protocol-specific invariants belong under protocols/.
An extracted Sparse Frontier Routing (SFR) candidate now explores a bounded escalation layer between cheap/local decision work and specialist/frontier investigation.
Its purpose is not to choose the “best big model” for every task. It asks first whether deeper reasoning is warranted at all.
Current extracted status: EXPERIMENTAL / UNWIRED.
Key boundaries:
- local/deterministic work remains preferred when adequate;
- frontier escalation receives only a bounded evidence-linked region;
- GA108 specialists remain persistent dormant identities and only a small subset may be recommended;
- SFR recommendations do not activate a model or grant authority;
- learned NBG signals remain log-only in v0.1;
- routing usefulness is recorded in shadow mode and cannot change live thresholds yet;
- Reality Gate / existing authorization remains the action boundary.
See packages/sparse-frontier-v0.1/README.md.
An extracted P2 reference now defines the application-level synchronization seam above the qualified P1 local store.
The boundary is intentionally narrow:
- the local PV-DLAM service decides memory admission;
- peer identity/signing and network transport remain replaceable dependencies;
- Infinite Porch is the intended transport/identity donor, not the memory source of truth;
- only signed scoped MEMORY_ADMIT and MEMORY_TOMBSTONE objects are accepted;
- SQLite/WAL files, vectors, context packets, opaque routing state, and action authority are not synchronized as canonical objects;
- remote claims receive collision-safe local replica identities;
- concurrent incompatible claims remain visible;
- same-peer source rewrites quarantine rather than overwrite;
- tombstone frontiers suppress stale replay;
- revocation epochs invalidate queued old-authority envelopes;
- signed receipts are evidence, never authority.
The extracted package lives at packages/dlam-p2-v0.1/. It is not yet live-wired to Infinite Porch or the canonical runtime.
P3 begins with an extracted observability layer over the qualified P1/P2 substrate.
The observatory records immutable finalized outcomes anchored to exact P1-C route receipts and optional Sparse Frontier decisions. It creates scorecards across the complete GA108 roster while preserving dormant identities with zero evidence.
The observatory may summarize:
- per-Genius task-class outcomes;
- exact model usage;
- quality/success history;
- confidence calibration;
- evidence satisfaction;
- corrections;
- latency/tokens/cost;
- SFR escalation usefulness;
- governance violations and critical misses.
It deliberately does not store raw prompt/memory text and does not activate learned routing.
P3 shadow rankings carry:
activation_status=SHADOW_ONLY_DORMANTranking_is_live=falsemay_change_live_route=falseauthority_granted=false
Live dispatch remains P1-C static policy until a later P4 qualification explicitly activates a bounded learner.
P4 begins with an extracted shadow-only learner. P4-A does not replace P1-C.
The routing order remains:
- current authority must be valid;
- GA108 candidates are sparsely shortlisted;
- P1-C hard model eligibility is evaluated;
- only hard-eligible exact Genius/model routes reach the learned scorer;
- exact-route support floors and non-tradable breach blocks apply;
- the learned candidate may emit a shadow recommendation;
- the live router remains P1-C static.
P4-A uses a fixed 24-feature pre-route schema and task-class-specific batch ridge parameters with a UCB uncertainty term. Features contain no raw memory, prompt text, hidden model state, policy secret, or post-outcome information.
Changed exact model artifacts begin with zero exact-route support, so a model swap cannot silently inherit the prior artifact's learned eligibility.
P4-A learned snapshots are immutable and pin source P3-C evidence plus an explicit rollback parent. There is intentionally no activation method.