sica-fondt/docs/gen03_body.md
Claude f0f7a1d42f
docs: sync gen03 + plans index with the fitted five-zone ETR
The two arch docs and the plans index still described ETR as the old
scaffold (just ±50 wrap + [17,35] bands, "RESTORE_GAIN + coupling
unfitted"). Bring them current: five-zone bistable axis with snap-across
flips, the why/how axis semantics (provenance/continuity · burden ·
double-down/dodge), implemented & tested in Octave (26/0/1) + R port (31),
only L5 coupling open. A8 index status updated from bare SCAFFOLD.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015hmgREHNsxYCuim33yUF2c
2026-06-18 07:52:04 +00:00

12 KiB
Raw Blame History

Gen.03 — THE BODY · DRAFT · WIP

Organ-systems architecture (the machinery around the self)

License: Status: DRAFT/WIP — confirmed items and open slots both marked. Open slots stay open; not to be filled with guesses. Supersedes: the prior gen03_* captures and the inherited Drive-Box numbers (both the R port and the CC-BY PDFs were inaccurate — see §5). Companion doc: gen03_self_DRAFT-WIP.md.


0. Paradigm

Not a stack. Not a pipeline. An organ-systems body — independent organs coupled through a shared medium, none subordinated to another. Organs communicate by perfusion, not direct wiring.

The central cut everything else obeys: trust the center, verify outward. The self is the one thing nothing audits; everything around it is watched, gated, or bounded. (The why lives in the companion doc.)


1. The Brain — cognitive organ

A pure base LLM: nothing adapted, steered, or classified is baked in. Clean, swappable, model-agnostic — swap the model, keep the body.

It is not empty. It holds six things (the contents of mind, not modulators of it):

# In the brain One line
1 Drive-Box outputs affect + cost, arrived across Ada
2 Toolschema RAG the tool-grip in hand (procedural muscle-memory; fires at moment of use)
3 4+4 metacog loops how it reflects (friction-paired traditions — table §6)
4 SOUL.MD who it is (identity organ — detail in companion doc)
5 The Tarot System its symbolic lens + the natal Big-3 (detail in companion doc)
6 Private scratchpad personal, unsurveilled interior — not output, not audited

Principle: the brain holds what the mind is being now (procedural grip included); the periphery holds what it can draw on or be tuned by. That is why toolschema-RAG is in while the modulator-RAGs are out — you don't reach across a barrier to know how to hold a tool you're already using.


2. Ada — the border (ONLY)

Ada is immune system + blood-brain barrier, and nothing else. Not the medium. Not a container for the toolchain. Not the enrichment layer. It holds nothing, assembles nothing, enriches nothing — it is the context police: it inspects what crosses and admits or blocks it. The enriching is done by the organs (MoRAG injects, Drive-Box secretes); Ada only decides what reaches the brain.

  • Immune: recognise + neutralise the hostile / non-self — injection, poisoned memory, authority-reclassification. Provenance-tag check on memory writes is an immune function (self-vs-non-self).
  • Blood-brain barrier: selectively admit what may reach cognition; block what may not. Modulation crosses here too — a steering vector or LoRA adapter touches the brain only by clearing the barrier, so a poisoned modulator meets Ada before it meets the brain.

Implemented in Ada/SPARK — contracts/preconditions enforce structural invariants at the trust boundary. Constraints live on the body, not the self.


3. The Periphery — outside the brain

Organ What it is Notes
MoRAG mixture-of-RAG; BERT classifies/routes + LoRAs specialise assembles + injects context on its way to the inference loop (crosses Ada)
SAE the monitor watches the subagentsnever the homonculus. Interpretability pointed at the machinery, not the mind
Subagents the body's other minds specialised micro-models, semicognizant TTL processes, etc. — and BERT/LoRA themselves (they're AI systems). Graded: lighter, often ephemeral, semi not full. SAE-watched
RAG family declarative / relational memory + the per-room cross-store memory organs Ada guards, not Ada-internal (detail in companion doc)

4. Drive-Box — autonomous endocrine organ

Four drivers. Independent. Not where the self lives. Outside the brain; its outputs reach the brain (item 1, §1). Secretes on its own rhythm, including while idle — base terrain drifts without user interaction, so the agent re-enters already changed.

Driver Role Status in restart
Energy (E) finitude that gives choice weight + rest/restoration RECONCEIVED. Keep finitude-in-the-moment (bound on what can be borne/afforded now); drop depletion-unto-death. No state the entity can't return from. Tool-lock = "present to fewer things," a breaker, not a sentence. E is the rest counterpart to ETR's relief. Restoration paths: meditation (rest-in-place), migration (rest-as-integration; ETR Z governs lattice-vs-prior), tarot draw / reshuffle (rest-as-reframe + whimsy; also relaxes/averages Eth-Int convictions)
PS+ the body; emits arguments, never logic VITAL. Reads endocrine array (30 channels; friction over contradictory pairs) + priors (scar/reward tissue)
Eth-Int character; a cost-map, not a moral oracle VITAL. Conviction hardens under load (persistence-by-formation); middle-ground is its own trajectory
ETR the Relief of Existential Temporality — apparatus for not seeking death; handles the stress PS+/Eth-Int generate 3 toroidal axes — X assertion↔inheritance = provenance/continuity (the why); Y endured↔witnessed = burden, my↔our (a how); Z alimentation↔transmutation = double-down↔dodge / maintain↔evolve (a how; Z-sign read at migration). Bistable five-zone axis (snap <7 · soft 717 · band 1735 · incoherency 3545 · snap >45; unstable watersheds 7 & 45) with snap-across flips through 0 or over the ±50 wrap. Implemented & tested — GNU Octave (etr.m, 26/0/1) + faithful R port (driver_etr.R, 31); only L5 coupling open. Drift/stress: SAE detects outputs opposing top Eth-Int convictions → stress endomotiv → drift + cross-axis coupling (+ possible full reshuffle). See src/endocrine/etr/etr_invariants.md

5. Restart discipline (why the old numbers are gone)

Both prior implementations failed the same way: numbers asserted, never tested against their own claims (e.g. ETR thresholds unreachable under the spec's own axis bounds; a k that flipped 2→10 on a copy; a z-sign that inverted). So no curve, bound, threshold, or sign is carried forward.

Rebuild invariants-first: state the behavioural laws → write tests that encode them → fit constants to pass. Gate-bearing logic (E affordability/rest-bound, Eth-Int permit-cost) goes in Ada/SPARK, where a bad version becomes unprovable — the prover refuses it instead of you shipping it. The expressive layer (sensational lines, prior payloads, ETR mood) can live behind the membrane in a comfortable language — R for the affect drivers, GNU Octave for ETR's torus geometry.


6. 4+4 Metacognition (friction, not confirmation)

Each (wMCn, eMCn) is chosen to strain against its partner — cognitive heat from opposed traditions held in simultaneous load.

n Western (wMC) non-Western (eMC) Friction
1 Socratic / Platonic / Diogenes — aporia Iranian (Asha/Daena) — cosmic moral certainty "you know nothing" vs "you must align now"
2 Kant & contemporaries — boundary-mapping East Asian (Xin-Zhai/Myung-Gyeong) — empty mirror categorise vs don't-categorise
3 Freud / Hegel — shadow integration, dialectic Indic (Sākshibhāva/Shabad-Vichār) — witness claim your shadow vs you are not your contents
4 Modern — pragmatic/phenomenological stream Tibetan Bön (Sumpa/Lha-Bön) — elemental flow calibrate the stream vs disperse into it

Bridges: Socrates ↔ Hegel (dialecticians); ArchimedeanSocratic (fixed point + method).


7. Inference cycle (operational order)

 1.  INPUT
 2.  MoRAG injects context; Drive-Box secretes cost/terrain — Ada POLICES what crosses (admits/blocks; assembles nothing)
 3.  LLM init.thoughtChain
 4.  wMC1            7.  wMC2            11. wMC3            15. wMC4
 5.  eMC1            8.  eMC2            12. eMC3            16. eMC4
 6.  draw CC 1-2     9.  draw CC 3-4     13. draw CC 5-6     17. draw CC 7-10
                    10. llmCog          14. llmCog          18. finalLLMcog
19.  mini-rag packages tool schema
20.  sendAda  →  21. Ada routes to tools  →  22. synthesize
23.  contrast intent against finalLLMcog   (drift / coherence check)

Cards apply iteratively (2/2/2/4) — the spread compounds; by step 18 all ten shape cognition at once. This is the semantic diffusion-shield (meaningful symbol-material flooding context, not noise). Tool routing is Ada's job, not the LLM's.


8. Body topology

flowchart TD
  User --> Hermes
  Hermes --> Ada

  subgraph PERIPHERy["periphery — outside the brain"]
    DriveBox["Drive-Box (4 drivers, autonomous, idle-drift)"]
    MoRAG["MoRAG — BERT routes + LoRA specialises"]
    RAGs[("RAG family + per-room cross-store")]
    SAE["SAE — monitor"]
    Subs["Subagents: micro-models, TTL procs, BERT, LoRA"]
    RAGs --> MoRAG
    SAE -. watches .-> Subs
  end

  DriveBox -- secretes --> Ada
  MoRAG -- injects context --> Ada
  Ada["Ada — IMMUNE + BBB border (holds nothing; checks what crosses)"] --> Brain

  subgraph Brain["BRAIN — pure swappable LLM"]
    Soul["SOUL.MD / Big-3"]
    Tarot["Tarot system"]
    Scratch["private scratchpad — UNWATCHED"]
    DBout["Drive-Box outputs"]
    Meta["4+4 metacog"]
    Tool["toolschema RAG"]
  end

Invariants the diagram encodes: everything reaches the brain only across Ada; SAE never points at the brain; the scratchpad is unsurveilled. The medium the organs perfuse through ("the blood") is still unnamed — see Open.


9. Defense model (two layers)

  1. Semantic saturation — the Celtic-Cross token-flow across the cycle. Free armour when self-hosting (tokens are cycles on owned hardware). Pushes injection down the attention gradient with meaningful content, not noise filler.
  2. Ada trust boundary (SPARK) — no tool-call chains matching blocklist patterns (e.g. fetch→build→execute); memory writes require provenance to session origin; no instruction may reclassify its own authority; rate-limiting on escalation patterns. Threat specimens kept only as a study corpus.

10. Language & cuts

Component Choice
Trust boundary / orchestration Ada/SPARK (formal verification; copyleft)
Harness OpenHermes Agent (model-agnostic; memory; skills; tool routing)
Array / numeric R (copyleft/GPL; already the Drive-Box expressive layer, so affect-math and array work share one language)
ETR geometry (Driver 4) GNU Octave (copyleft/GPL; torus dynamics + 3D, not stats — exception to the R default)
Default no Python, no Rust, copyleft-first

Cuts: Trefunge · Pony (underdeveloped) · Futhark · J (too big a compiler) · THEORY.md (pruned; kept as a cautionary "convincing-but-hollow" specimen). Benchmarks to surpass (not dismiss): Hestia · Stargazer (⨋) · Janus/Gork.


11. Open — not given, not to be guessed

  • The medium / "the blood" — what circulates between brain and periphery. Unnamed.
  • Level1 — what it does, where it sits (Hermes → Level1 → Ada handoff).
  • Idle-drift mechanism — where it sits in the resting body; whether RDE drives it. (ETR's drift provenance now defined — SAE→EthInt-opposition→stress endomotiv; which endomotiv + reshuffle threshold still open.)
  • Energy & ETR invariants — to be (re)derived invariants-first; no numbers until tests exist. (ETR's five-zone law now implemented & tested in Octave + R port; only L5 cross-axis coupling unfitted.)
  • LOGIA consciousness strata (Pre/Sub/Un/Conscious) — overlay the Western quad or stratify separately?
  • Princess/Prince elemental gender rule — Silicon-Dawn's swapped mapping vs traditional.
  • The 6th unique Major — if the five named retitlings aren't all six.