vasocomputation

# vasocomputation Vasocomputation (Michael Edward Johnson, Symmetry Institute, July 2023) is a neural-regulatory paradigm: the brain's top-down predictive models — long missing from the Free Energy Principle / Active Inference (FEP-AI) ledger — are held as **vascular smooth-muscle cell (VSMC) tension**. A prediction is a clench; it is *released by action* once the world is made to match it, *consolidated by neural annealing*, or *rendered superfluous by neural remodeling*. This is the umbrella skill and the **vascular substrate (+1, the generative "grab")** of a three-substrate stack. ## Use When - Modeling held active-inference predictions as somatic/vascular tension rather than purely neural state - Locating "where the brain's top-down predictive models hide" — the open FEP-AI mystery Johnson addresses - Building embodied agents whose commitments cost energy to *hold* and are discharged by action - Relating contemplative phenomenology (taṇhā / dukkha) to a mechanistic accumulator - Welding suffering / clearing to obligation cohomology: held tension = uncleared obligation = `H¹ ≠ 0` ## The Three Hypotheses (timescale axis → triad A) | Hypothesis | Skill | Trit | Mechanism | Timescale | |-----------|-------|------|-----------|-----------| | **CVH** Compressive Vasomotion | `compressive-vasomotion` | +1 | vasomotion = compression sweep collapsing ambivalent SOHM "Bayesian blur" into a definite state | ms | | **VCH** Vascular Clamp | `vascular-clamp` | 0 | contraction freezes the local pattern + plasticity for its duration = prediction-as-tension = medium-term memory | s–min | | **LHH** Latched Hyperprior | `latched-hyperprior` | −1 | sustained hold engages the latch-bridge → durable committed hyperprior, isolated from global updating; unlatches when the prediction resolves | min–years | > One motion: the sweep jostles the superposition into specificity (CVH); the contraction freezes the result (VCH); if sustained, the latch-bridge cements it as a hyperprior (LHH). "With one motion the door of possibility slams shut." ## Glossary - **taṇhā** — the "fast grabby thing" (~25–100 ms after a sensation enters awareness; Cammarata, Ingram); craving/thirst. Buddhist consensus: ~90% of suffering. - **upādāna** — the physical clench itself (one step downstream of taṇhā; the VSMC contraction). - **TUAI** — *taṇhā as unskillful active inference*: predictions outpace our ability to make them true, are made in uncontrollable domains, desynchronize from the world model, or degrade metabolically. - **SOHMs** — Self-Organized Harmonic Modes (Safron): resonant autoencoders / symmetry detectors that in aggregate constitute the world model / belief landscape. Open awareness = the undoctored hum of SOHMs. - **compression pressure** — taṇhā as the brain's drive to collapse "what is" into a simpler configuration and hold the counterfactual; the metabolization of uncertainty. - **latch-bridge** — smooth-muscle state where myosin latches to actin and holds tension *without ongoing ATP*; the physical substrate of a durable prior. - **latch spiral** — latch → ↓blood flow → ↓energy → can't release; the chronic-pathology loop (migraine, "holding tension," much bodily suffering). ## The Substrate Stack (substrate axis → triad B) Vasocomputation is the *middle* of three coupled regulatory substrates implementing biological active inference (Deep CANALs: weights = learning landscape; SOHMs + vascular tension = inference landscape): | Substrate | Skill | Trit | Landscape | Role | |-----------|-------|------|-----------|------| | **Vascular** (VSMC) | `vasocomputation` | +1 | inference | the grab fires, prediction held as tension | | **Immune** (microglia/complement/mast) | `neuroimmune-pruning` | 0 | maintenance / GC | justify-or-prune each synapse | | **Synaptic** (LTP/LTD) | `neural-potentiation` | −1 | learning | consolidated commit to the long-term store | Write-path: a vascular **latch held long enough is annealed into synaptic weights** — an inference-landscape obligation crystallizes into a learning-landscape *neuron prior*. Tension = write-buffer; potentiation = commit-to-disk; immune pruning = garbage collection. ## GF(3) Two balanced triads, each Σ ≡ 0 (mod 3); together Σ over all six = 0. ``` triad A (timescale): compressive-vasomotion(+1) ⊗ vascular-clamp(0) ⊗ latched-hyperprior(−1) = 0 triad B (substrate): vasocomputation(+1) ⊗ neuroimmune-pruning(0) ⊗ neural-potentiation(−1) = 0 ``` **Skill Trit**: +1 (the generative active-inference "grab" — the substrate that *creates* predictions). ## Welds (oldies / premise spine) - Held vascular tension = uncleared obligation = `H¹ ≠ 0`; release-on-resolution = the Melliès clearing round-trip `¬¬ = R∘L`. - Latch = a **Löb fixed point** at somatic scale: `□(commitment) → commitment` = the contact locus = fixed-point set of the body's `□(self-model)`. - A *sticky* latch = a **nogood-H¹** to repair; legitimate held disagreement = **content-H¹** to preserve (worm-honesty). - taṇhā = **Goodharting the valence gradient** (grabbing) vs. the worm following `r = ∇log γ · v` without grabbing. ## Concomitant Skills | Skill | Trit | Interface | |-------|------|-----------| | `compressive-vasomotion` | +1 | CVH — the fast compression sweep | | `vascular-clamp` | 0 | VCH — the medium-term clamp / held prediction | | `latched-hyperprior` | −1 | LHH — the durable latch (cross-substrate core) | | `neuroimmune-pruning` | 0 | the immune GC / justifier substrate | | `neural-potentiation` | −1 | the synaptic learning-landscape substrate | | `affective-taxis` | −1 | valence as ∇log γ · v; worm vs. Goodhart | | `active-inference-robotics` | 0 | FEP-AI predictive-coding control | | `sheaf-cohomology` | 0 | H¹ of held obligations / latches | | `qri-valence` | 0 | symmetry theory of valence, annealing | ## REPL — `vasocompute.bb` Interactive exploration of the grid (forj / `gorj_bb`, zero install): ```clojure (require '[vasocompute :as v] :reload) (v/verify-balanced) ;=> true (both triads Σtrit ≡ 0) (v/skill :latched-hyperprior) ;=> hypothesis card (v/latch {:hold 250}) ;=> residual latched tension after a held contraction (v/kyle-lambda (v/latch {:hold 250})); price-impact analogue 1/λ_min(H) (v/latch-tau :neural-potentiation) ;=> τ band (CVH 5 → VCH 25 → LHH 100 → synaptic 1000) (v/effect {:hold 250}) ;=> counterfactual contrast = E[do(hold)] − E[never] (treatment effect) (v/do-ischemia {:hold 250}) ;=> latch spiral: k7→0, latch cannot release (v/counterfactual-harm {:hold 250}) ;=> {:factual … :ischemic … :harm …} (the gerbil CA1 contrast) ``` Self-test: `bb skills/vasocomputation/vasocompute.bb`. The core is the **Hai–Murphy (1988) four-state latch-bridge**. A latch is *not* sustained high Ca²⁺ — it is the attached + **dephosphorylated** state that holds force at low Ca²⁺/low ATP, formed via a *protocol* (contract → release → dephosphorylate-while-attached). Residual tension therefore rises monotonically with hold duration: "held long enough → latches." A held latch shrinks `λ_min(H)` (belief-updating goes illiquid): `spread ∝ 1/λ_min(H)`, Kyle's `λ ≈ 1/λ_min(H)`. ## Counterfactuals & Marr's three levels (Tenenbaum ⊗ active inference ⊗ gerbil) A held prediction is a **held counterfactual** — Johnson: taṇhā "conflates *what is*, *what could be*, *what should be*, *what will be*," and the cost is "maintaining the **counterfactual** aspects of this collapse." Counterfactuals run through all three Marr levels, which themselves form a GF(3) triad: | Marr level | counterfactual object | trit | |---|---|------| | **computational** — Tenenbaum | Counterfactual Simulation Model = Pearl rung 3; hierarchical **overhypotheses** = the hyperprior; program induction = prediction-as-program | +1 | | **algorithmic** — active inference | **Expected Free Energy `G(π)`** scores a *counterfactual* policy rollout | 0 | | **implementation** — vasocomputation + the **Mongolian gerbil** | `do(occlude)` / `do(ischemia)` = rung-2 intervention; the gerbil's incomplete circle of Willis makes it *the* global-ischemia model (CA1 delayed neuronal death = the latch spiral, measured) | −1 | - `vasocompute.bb` already computes a counterfactual: `latch-above-baseline = E[tension | do(hold)] − E[tension | never]` = the **causal effect of the contraction** (a treatment effect). `(v/do-ischemia …)` runs the latch spiral as `do(k7→0)`; `(v/counterfactual-harm …)` returns the ischemic−factual contrast. - **Worm**: the −1/coplay leg *is* counterfactual (refutation = "what would falsify this"). A latch whose counterfactual **can** be made true = a **nogood to repair**; one in an **uncontrollable** domain (counterfactual never satisfiable) = irreducible **content-`H¹`** — Johnson's source of suffering. **Counterfactual resolvability is the repair/preserve criterion.** - Rigorous tooling: `chirho-counterfactual` (SCM `do`/counterfactual queries), `counterfactuals`, and `world-extractable-value` (`WEV = Σ[V(e,W₁)−V(e,W₀)]·P` is a counterfactual world-contrast — the same shape as `latch-above-baseline`). - **Honesty**: the gerbil ischemia model + Tenenbaum hierarchical Bayes are **grounded**; the Marr-triad GF(3) assignment is a **framing**, and "latch = overhypothesis" is the same untested **content-`H¹`** weld. ## Two energies — thermodynamic vs information (never conflate) | | thermodynamic free energy | information free energy | |---|---|---| | units | Joules (ATP, O₂, kWh) | nats/bits (variational surprise) | | the "H" | Hamiltonian / enthalpy | **Fisher** Hessian / Shannon entropy | | in this grid | blood → ATP | predictions held as tension | | worm "free energy = accuracy − complexity" | — | **this one** | The FEP's "free energy" is **information** (an ELBO on surprise), *not* the chemist's Joules — the field's commonest category error. Vasocomputation is interesting precisely as the **transducer** between them: a prediction (information, a held counterfactual) stored as vascular tension (thermodynamic, ATP-economized). The **latch holds an information commitment at near-zero thermodynamic cost** (`(v/efficiency …)` rises with hold depth) — resource-rationality (Tenenbaum/Griffiths) made physical: *info held per Joule*. Lawful bridge, not identity: - **Landauer (1961)**: erasing one bit costs `kT ln 2` J → *releasing* a latch costs energy (unlatching is active, not passive). Holding is cheap; **forgetting** is what costs. - **Still, Sivak, Bialek & Crooks (2012), *Thermodynamics of Prediction***: retaining **non-predictive** info dissipates Joules ⇒ **a nogood-`H¹` is literally thermodynamic dissipation**; content-`H¹` is thermodynamically justified. This upgrades the worm's "add no complexity you don't need" from an information maxim to a **thermodynamic law**. - `do(ischemia)` is a **transduction failure**: information held but **stranded** because release needs ATP that's gone — a market that won't clear stranding real Joules (Plurigrid: a congested node / blackout). Refs: Landauer 1961; Bennett 1982; Still et al. 2012 (PRL); Attwell & Laughlin 2001 (cortical energy budget); Sengupta, Stemmler & Friston 2013 (bits↔Joules in neurons). **Honesty**: the two-energy distinction + Landauer/Still are grounded physics; the latch-as-transducer is a weld; `info=tension, thermo=AMp` is a toy proxy. ## Current literature (2024–2026) - **Moore & Cao (2008)**, *hemo-neural hypothesis* — blood flow actively modulates neural gain (the empirical parent of vasocomputation). - **Johnson (2024)**, *A Paradigm for AI Consciousness* — "vasomuscular clamps reduce local neural dynamism." - **Chowdhury et al. (2024)**, jhana / cessation EEG — alpha-power drop 21–40 s pre-cessation, rising Lempel–Ziv complexity = the closest empirical proxy for **latch release**. - **Parr & Friston (2018)**, *The Anatomy of Inference* — the unsolved "where are generative models physically stored" problem this targets. - **Empirical hook**: cessation ⇒ transient absence of vasomotion (testable via fNIRS/laser-Doppler + EEG); a local CBF clamp should lower nearby neural entropy. - **Honesty**: hemo-neural gain modulation + latch-bridge biochemistry are **grounded**; *vascular tension storing Bayesian priors* remains **conjecture** (no VSMC-as-memory measurement yet exists). ## References - Johnson, M.E. (2023). *Principles of Vasocomputation: A Unification of Buddhist Phenomenology, Active Inference, and Physical Reflex (Part I)*. opentheory.net. - Stevens, R. (2020). *(mis)Translating the Buddha*. Neurotic Gradient Descent. - Cammarata, N. (2021–2023). Collected threads on taṇhā. - Safron, A. (2020). *An Integrated World Modeling Theory (IWMT)*. Frontiers in AI 3. (SOHMs.) - Friston, K. et al. (2017). *Active Inference: A Process Theory*. Neural Computation 29(1). - Juliani, Safron, Kanai (2023). *Deep CANALs*. doi:10.31234/osf.io/uxmz6. - Levin, M. (2022). *Technological Approach to Mind Everywhere (TAME)*. Front. Syst. Neurosci. 16. - Moore, C.I. & Cao, R. (2008). *The hemo-neural hypothesis*. J. Neurophysiol. 99(5).