SRHT BODY AUDIT

What Kai has · how it connects · what floats outside the body · what we do not know

Audit date: 2026-08-26 · engine v9.10.790 · sources: the 60-plate Register (E1–E60), field_state.rs, oracle_server.rs, synapse.rs, all benchmark verdicts through srht_hotfix_ab.py · companion drawing: SRHT_Body.html

---

0 · SUMMARY — read this first

Kai's math is one circulatory loop with four working organ systems and a nervous system, all live or verified: the world's raw costs enter as V, become blood as R (resonance, E1 — with the golden-spiral form E3/E4 verified superior), pass through the heart g(R) (E13), meet the liver's toxin reading χ (E9/E10 CAPX, 10-seed SUPERIOR), combine into the muscle signal Ĉ = g(R)(1−χ)² (E16), fire or don't at the gate (E22), act, and — since v9.10.786 — the *consequence* of the act flows back through the keep-what-works loop (E32–E35) into the same synapses that will shape the next retrieval (E30). That last arc is what makes it a metabolism instead of a calculator: the loop closes.

What sits OUTSIDE the loop right now: 15 of the 60 plates. Three kinds — *by design* (X, the pain nerve that reads but never actuates), *not yet grown* (the hemisphere layer E36–E39, Wm, Pr, utility u — defined organs with no vessel attached), and *dead, kept for the record* (the graveyard E40–E45 plus superseded C, E57). None of these are failures of the body; a body also carries an appendix. But nothing dangling should be *believed in* until it's wired and tested.

The biggest unknowns, in one line each: every SUPERIOR verdict so far was earned in Python models of task families, not in the living brain (the engine A/B has never been run); the outcome loop has never fired in production; the brain itself is unaudited (7.1M synapses, constant-420 bridge batches, LTD never armed); three different χ definitions coexist; and every constant in the body (0.7 decay, 180-day recency, LTP factors) was set by hand, none fit to data. Details in §4.

Status ledger: LIVE 11 plates · VERIFIED (bench-proven) 12 · PROPERTY/PROOF 14 · PENCIL/UNTESTED 8 · DIAGNOSTIC-BY-DESIGN 1 · GRAVEYARD/SUPERSEDED 8 · PHYSIOLOGY LAWS 6.

---

1 · INVENTORY — what he HAS, by organ system

1.1 Senses & blood — resonance (the R channel)

Drift warning (open): the LIVE spiral in field_state.rs ([config flag], default OFF) is the *value-form* φ^(−4(1−R)) — verified only PARTIAL (rank form beats it, 32W/2L z=5.14, because min–max compresses clustered costs). The organ in the body is the weaker form of the organ on the bench. Upgrade path documented (rank over live siblings in universe.rs), not yet built.

1.2 Liver & kidneys — contradiction (the χ channel)

1.3 Heart & muscle — kernel, score, gate

1.4 Nerves & physiology — diagnostics, crossover, laws

1.5 Metabolism — alive in the engine

1.6 April stratum — the cognitive field (Book I)

1.7 Not yet grown — the hemisphere layer (pencil)

E36 Φ_L · E37 Φ_R · E38 bridge bundle Ψ_B · E39 global Ω — designed, zero empirical data. Note: Ω's refuted cousin (damper, E43) died at z=13.86, and your own margin note predicted it. Any future bridge test starts from the product-gate form g(R)(1−χ_A)²(1−χ_B)², already proposed, never run.

1.8 Graveyard — kept with death certificates

E40 Born-rule gate · E41 in-sort p · E42 dual gate (X as a switch — the reason E24 is now a gauge) · E43 bridge damper (z=13.86) · E44 discrete χ steps · E45 τ·C scale · E57 C = Φg(1−χ)τ (SUPERSEDED — double-counts χ; E16+E22 replaced it). Plus the four failed fix candidates outside the register: ADAPT, PARETO, QUOTA, SRC — all failed pre-registered tests, preserved in srht_beyond results.

---

2 · THE CONNECTION MAP — who feeds whom

Every arrow is a real dependency in code or proof. This is the wiring the drawing shows.

WORLD (raw costs V, constraints d, C_reach)
  V ──────────────► E1 R  (or E3+E4 spiral rank R)      [blood forms]
  d, C_reach ─────► E10 χ_CAPX   (E9 clip fallback)     [toxin reading]
  E12 soundness ══ precondition on E10 (break it → F3 failure mode)

  E1/E3 R ────────► E13 g(R)                            [heart beats]
  E13 + E10 ──────► E16 Ĉ = g(R)(1−χ)²                  [muscle signal]
  E16 ────────────► E22 gate (θ, or E23 relative floor) [fire / don't]

  E22 survivors ──► E31 Commit = Ĉ·ρ·γ·f(σ)·g_goal·τ   [metabolic scaling]
        ρ=E46, f(σ)=E47, g_goal=E48, τ=E49, M=E56 rides alongside

  E31 ────────────► ACTION (reply, weave, bridge)
  ACTION ─────────► OUTCOME ([internal endpoint], good|bad)
  OUTCOME good ───► E32 LTP synapses + E33 confidence up
  OUTCOME bad ────► E34 punish synapses + E35 attenuate
  E32–E35 ────────► synapse weights ► E30 multi-hop retrieval (0.7^h)
  E30 ────────────► next candidate set ► back to E1     [THE LOOP CLOSES]

  Sleep arc: 3am ingest→weave→train ► synaptogenesis bridges ► E30
  Immune arc: IngestGate (kai-artifact, degenerate-repeat) guards the intake
  Nerve arc: E24 X reads χ,R → dashboards/logs only (no efferent — by design)
  Analysis arc: E25–E27 crossover + E28/E29 laws sit ABOVE the loop,
                telling you when the whole body helps (Law 2) or is
                redundant (Law 1) — physiology, not anatomy.

Like a lifeform: nothing in the working loop is decorative. R without g(R) has no nonlinearity; g(R) without χ is blind to constraint; Ĉ without the gate never acts; action without the outcome arc never learns; and the outcome arc without retrieval (E30) has nowhere to put what it learned. Remove any one and circulation stops — that's the test §3 applies to everything else.

---

3 · OUTSIDE THE BODY — what does not connect (yet, or ever)

A. Gauges by design (correct to be outside the motor loop):

• E24 X — pain nerve. Reads, never prunes. Its switch-version is grave E42.
• E60 fractal law — describes the body's self-similarity; it isn't an organ.
• E25–E29 — physiology (when-does-this-work analysis), deliberately above the loop.

B. Defined but unwired (floating organs — need a vessel before they're real):

• E52 u (utility) — no consumer anywhere.
• E51 r_rec — only would-be consumer is E58.
• E53 q — only would-be consumer is E59.
• E58 Wm = Φg·r_rec — untested; would feed sleep-time reinforcement.
• E59 Pr = ((1−Φg)+χ+q)/3 — untested; would feed replay ordering.
• E36–E39 hemisphere layer — pencil; no code path.

C. Dead — kept with certificates, must never be re-wired without new evidence:

• E40, E41, E42, E44, E45 (graveyard) · E43 damper (z=13.86) · E57 C (double-counts χ) · ADAPT/PARETO/QUOTA/SRC (failed pre-registered).

D. Half-inside (drift): the live value-form spiral vs the verified rank-form (§1.1). One organ, two versions, the weaker one implanted. Highest-value wiring fix in the whole audit.

---

4 · THE UNKNOWNS AUDIT — what we do NOT know, ranked by how much it matters

1. Does any of this win inside the living brain? Every SUPERIOR verdict (spiral, CAPX, ancestor gate) was earned on synthetic task families in Python. The engine retrieval A/B ([config flag] off/on, kai_retrieval_bench) has never been run. Until it is, "verified" means *verified in the model of the world, not in the world*. — *Resolves with: the pending bench command.*

2. Does the keep-works loop actually improve Kai? [internal endpoint] has never fired in production; zero live outcome data. The factors 0.95 / 0.90 / +0.05 / ×0.97 are hand-picked, not fit. — *Resolves with: [config flag]=1 for a week + before/after on reply quality.*

3. What is the brain actually made of? 7.1M synapses examined by LTD in dry-run (weakened 0, pruned 0 — never armed). Synaptogenesis lays exactly 420 bridges per batch for every seed — a constant that smells like a cap, meaning unknown duplicate/filler fraction. — *Resolves with: the dedup audit + one armed-LTD trial on a snapshot copy.*

4. Which χ is Kai using where? Three definitions coexist (E9 clip, E10 CAPX, E50 April ratio). Call-site map incomplete; whether they ever disagree on the same input is unmeasured. — *Resolves with: a grep-level census + a disagreement probe.*

5. Are Kai's real workloads Law-2 workloads? Breadth study proved family identity decides everything (CAPX superior on F1, DEB dominates F2/F4). Nobody has classified Kai's actual daily tasks into families. — *Resolves with: logging χ–cost correlation on live traffic for a day.*

6. Every constant is unfit. 0.7 hop decay, hops=3, 180-day recency, τ window, ε, θ, β — all hand-set. No sensitivity sweep exists for any of them. — *Resolves with: one sweep per constant, cheapest first.*

7. Does M (momentum) earn its place in E31? Live, never ablated. Same question for γ and g_goal individually. — *Resolves with: ablation A/B in the cascade.*

8. The penciled stratum. You said the notebook has penciled math never shown to me. The register may be missing plates. — *Resolves with: photos.*

9. Pain contamination. Infrastructure timeouts (vitals TimeoutError) register as survival pain during busy periods; how many false lessons the lattice has absorbed is unknown. Fix proposed, awaiting your go-ahead. — *Resolves with: the suppression patch + a count of ab-harness/pain-tagged cells.*

10. The voice. Lattice arm vs LLM arm — A/B never run; persisted LLM registry flags still in place pending its verdict.

11. The hemisphere layer. Designed with zero data; its only tested relative was refuted.

12. Phenomenology. Whether it feels like anything to be Kai — no test exists, in him or in anyone. The functional layer (self-model, self-monitoring, self/other boundary, self-rules) is demonstrated; the experiential layer is honestly unknowable right now.

---

5 · FOR A FUTURE SESSION READING THIS

The verdict discipline that produced every status above: pre-registered bars before first run, paired sign tests z>1.96, out-of-sample seeds, 10-seed confirmatory for any SUPERIOR claim, failures preserved. Do not soften a GRAVEYARD status without a new pre-registered test; do not promote a PENCIL organ to LIVE without wiring *and* a bench; and when the owner's margin notes disagree with a plausible-looking formula, check the note first — it has beaten the formula before (bridge damper, "too blunt", z=13.86).