sica-fondt/docs/plans/C3-inference-cycle.md
Claude 20e99f434c
docs/plans: remaining component specs — Identity, Cognition, Border, Storage, Periphery, Cross-cutting
Wave 1 remainder + Wave 2 design-first specs, completing one build plan per sub-organ:
- B1 tarot emulator, B2 SOUL/Big-3, B3 Celtic Cross
- C2 4+4 metacog, C3 inference cycle, C4 Brain
- D1 Ada border, D2 the medium, D3 mini-rag
- E1 3 GnuCOBOL invariant stores, E2 VARIANT stores, E3 RAG family
- F1 MoRAG, F2 SAE monitor, F3 subagents
- G1 stress-loop contract, G2 governance, G3 defense model

Each: contract + stubs for independent build; unfitted constants marked C1.

https://claude.ai/code/session_015hmgREHNsxYCuim33yUF2c
2026-06-13 20:08:19 +00:00

2.5 KiB

C3 — Inference cycle orchestration (the 23-step pipeline)

1. Component

The forward-only sequencer that runs one turn: input → enrich (Drive-Box + MoRAG across Ada) → init → the 4+4 passes interleaved with Celtic-Cross draws → final cognition → mini-rag → route → synthesize → coherence check.

2. Status / certainty

Exists in Ada (organs/ada_medium/ada_medium.{ads,adb}, 344 L, SPARK) with a protected orchestrator enforcing legal phase progression — but defunct-flagged ("must be deleted/replaced"). Reuse-vs-rebuild is the key decision here. Structure C4.

3. Language & location

Open. The sequencer could stay Ada/SPARK (it's gate-like, forward-only) OR move to the harness side with Ada only policing crossings (D1). Decide in build step 1.

4. Does / does-not

  • Does: order the 23 steps; enforce forward-only progression (illegal jumps = error); call each organ at its step; carry the coherence/drift check at the end.
  • Does-not: be any organ — it sequences C2/B3/A1/F1/D3/D1, it doesn't implement them.

5. Interface contract

  • run_inference_cycle(session_key, input) -> result (today's Ada_Medium.Run_Inference_Cycle).
  • Step map (canonical): 1 input · 2 enrich (A1 secretes + F1 injects, D1 polices) · 3 init · 4-17 wMC/eMC + draw CC (C2/B3) · 10/14/18 llmCog · 19 mini-rag (D3) · 20-21 sendAda + route (D1) · 22 synthesize · 23 contrast intent vs final cognition (drift).

6. Dependencies & stubs

Every organ it calls — stub: each returns canned output (the C1 spec already lists a pass-through stub). Standalone-testable as pure sequencing over stubs.

7. Invariants / laws

  • L1 (C4): forward-only phase progression; illegal jumps yield an error state.
  • L2 (C5): tool routing is Ada's job (step 21), not the LLM's.
  • L3 (C4): step 23 is a drift detector, never a reward signal (per C1/L4).

8. Build steps

  1. Decide reuse-vs-rebuild of ada_medium (audit its SPARK proofs vs the "defunct" flag) and where the sequencer lives. 2. Lock the step map + L1 as tests. 3. Wire real organs as their specs land, replacing stubs. 4. Add Energy-gating (skip later C2 passes at low E).

9. Tests

Phase-progression test (legal path passes, illegal jump errors); routing-is-Ada test; drift-check fires on mismatch. (Existing mafiabot_core/tests/cycle_tests.adb if Ada path is reused.)

10. Open items

  • Reuse-vs-rebuild + home of the sequencer (the central open call here).
  • Level1 placement in Hermes → Level1 → Ada (C1, shared).