RHSF Detailed Design Basis: Mead-Mahowald Retinomorphic Analog VLSI Foundation
RHSF Detailed Design Basis: Mead-Mahowald Retinomorphic Analog VLSI Foundation
Design-basis document for the Retinomorphic Hamiltonian Self-Field (RHSF) framework that grounds a hardware substrate in Mead-style analog VLSI, combining subthreshold MOS transconductors, non-pixel retinomorphic receptor lattices, clock-free accumulation, and duty-cycled Gm-C Hamiltonian relaxation to implement low-power, latched attractor computation. Intended to support schematic design, SPICE validation, and device-level implementation planning (not a claim of completed silicon).
Consensus round triggered on 2 dimensions
Resolved: 2 - Still contested: 0
The document is structurally coherent in its signal-flow architecture (receptor → evidence → accumulation → latch → Gm-C relaxation). Design constraints (bounded N_max, bias-gating during accumulation, tile-local sparse projection) are stated early and respected throughout. However, a central definition drift exists around the term 'Hamiltonian.' Section 11 defines Hamiltonian mechanics as the conservative canonical system dq/dt = ∂H/∂p, dp/dt = −∂H/∂q preserving a scalar H. Section 10 implements a driven, damped, mean-field-coupled, nonlinear ODE with explicit dissipation (−G_diss Vp,i) and latched forcing current (I_u,i^latched). No modified Hamiltonian, port-Hamiltonian decomposition, gradient-Hamiltonian hybrid, or Lyapunov function is defined that would connect the conservative reference to the implemented dynamics. The document acknowledges the mismatch (Section 11: 'RHSF is damped and driven, not purely conservative') but this acknowledgment does not resolve the drift — it merely names it. Because the term 'Hamiltonian' appears in the framework title (RHSF), the core computational primitive description ('Hamiltonian relaxation'), and the central functional claim ('relaxes into an attractor'), the missing definition of what mathematical object qualifies as 'Hamiltonian' in the implemented circuit is a central inconsistency under the rubric's definition-drift rule. The strongest opposing concern from the 5/5 assessment — that terminology is used consistently — is not persuasive here because 'consistent use' of a qualifier whose meaning has shifted from its explicit definition does not constitute internal consistency. Secondary issues include: (i) the mean-field term V̄q is used without definition (normalization, coupling topology, boundary conditions); (ii) state variables Vq, Vp are not mapped to canonical q, p with scaling units; (iii) Gm coefficients are not constrained by sign conditions needed to guarantee relaxation behavior. These secondary issues would independently reduce the score, but the central definition drift caps the score at 2 under the rubric rule. A consensus round resolved an earlier panel split before this score was finalized.
Many local equations are standard and dimensionally plausible as stated (weak-inversion exponential law in Sec. 3; U_t = k_B T/q; differential-pair tanh and its small-signal Gm in Sec. 4; accumulator ODE and leaky-memory form with τ = RC in Sec. 9). However, the strongest opposing concern (raised by peers arguing for ≤3) is decisive for this score: the central RHSF dynamics in Sec. 10 are asserted without a reproducible derivation from either (i) an explicit circuit schematic/netlist whose KCL/KVL reduces to the stated ODEs, or (ii) an explicit scalar Hamiltonian/energy/Lyapunov function plus dissipation/forcing that would justify the ‘relaxation into an attractor’ behavior. In addition, the mean-field term G_mf(λ_i) V̄_q uses V̄_q without definition (average over which nodes? normalized by N_a or N_max? boundary conditions?), which affects correctness of coupling scaling and whether an energy function could exist.
Consequence chain (load-bearing): the architecture’s core functional claim—short duty-cycled inference via convergence/attractor selection after latching—depends on Sec. 10 having the intended stable attractors and on the latch/injection faithfully producing the intended forcing currents. If Sec. 10’s form is not physically realizable by the intended transconductor network or does not possess robust attractors under mismatch/noise, then the ‘Hamiltonian relaxation’ computation mode is not mathematically supported by the present document (though it could still be validated empirically in later SPICE work). This is why the mathematical_validity score is capped at 3 under the unverified-central-derivation rule.
Addressing the strongest opposing point for a higher score (4–5): the argument that “these are merely design primitives, so missing proofs don’t count” does not remove the load-bearing gap, because even for a design-basis document, presenting a specific coupled ODE as the RHSF ‘Hamiltonian cell’ and using it to justify an attractor-computation concept requires at least a derivation path (circuit-to-ODE or energy-function-to-ODE) or explicit stability conditions. Without that, the central mathematical claim remains unverified. A consensus round resolved an earlier panel split before this score was finalized.
Using the empirical/methods rubric. The work is reasonably testable because it lays out a staged validation ladder with concrete acceptance criteria: subthreshold differential pair tanh transfer and Gm-bias scaling, accumulator leakage/retention, sample-hold droop and charge injection, single-cell attractor polarity under latched bias, and 2-10 cell collective convergence. These are experimentally falsifiable with standard SPICE and later bench characterization. The main limitation preventing a 5 is that most criteria are qualitative or tolerance-relative rather than fully quantitative system-level benchmarks; e.g., 'does not alter attractor decision' and 'agrees with input class' need sharper numeric thresholds, noise margins, and power budgets.
For a working design document, the structure is strong: lineage, objective, device physics, circuit primitive, receptor architecture, accumulation, core cell, validation ladder, risks, and work packages. A scientifically literate reader can follow the intended architecture and development path. The main clarity weaknesses are conceptual rather than organizational: 'Hamiltonian' is used as an inspiration label for a system that is explicitly damped, driven, and attractor-seeking; several placeholders remain high-level; and some symbols/variables are introduced only at a design-equation level without an accompanying physical readout definition or explicit mapping to classification outputs.
The individual primitives are well-established Mead-style analog VLSI (subthreshold transconductors, tanh differential pair, resistive retina network, Gm-C integration). The novelty lies in the synthesis: a duty-cycled, bias-gated Gm-C 'Hamiltonian relaxation' core fed by a generalized non-pixel retinomorphic receptor lattice with clock-free accumulation and latched forcing. The generalization of the retina beyond optics to arbitrary physical domains and the accumulation/commit/relax duty-cycle framing is a reasonable and somewhat original architectural combination, but the core computational mechanism is not sharply differentiated from existing analog Hopfield/attractor and neuromorphic relaxation circuits. It is an interesting recombination with modest new elements rather than a fundamentally new mechanism.
For a v0.5 design-basis document targeting SPICE validation, the submission is remarkably complete. It covers: prior-art lineage with specific attribution to Mead/Ismail and Mead-Mahowald; subthreshold MOS physics with correct governing equations; the tanh transconductor model with small-signal and saturation limits; the non-pixel receptor lattice abstraction with domain-specific examples; directional evidence channels; the P-matrix excitation structure with bounded cell count; clock-free accumulation dynamics including leaky integration; the asynchronous commit functional; the Gm-C cell state equations; design rules; a risk/falsification table; and a six-level SPICE commitment ladder with explicit acceptance criteria. Work packages with deliverables and exit conditions are provided. The SPICE netlist skeleton is minimal but intentionally so and labeled as such.
The one notable gap is λᵢ in the Gm-C cell equations (Gₘ₂(λᵢ) and Gₘf(λᵢ)) — it appears without definition, leaving the mean-field coupling mechanism underspecified at the circuit level. This is a secondary detail rather than a core argument gap, since the document's purpose is to establish design basis rather than full circuit specification. The floating-gate/Fowler-Nordheim adaptation path is mentioned in design rules but not developed in the circuit-block mapping. The commit detector SPICE implementation uses a behavioral voltage source that is acknowledged as a placeholder. These are appropriate gaps for a v0.5 design basis but prevent a score of 5.
As a framework document evaluated in FRAMEWORK-MODE, the evidence roadmap is partially developed. Strengths: the six-level SPICE commitment ladder (Section 14) provides a clear, sequenced falsification path with explicit acceptance criteria at each level, which is genuinely useful structure. The risk/falsification table (Section 17) identifies seven specific failure modes with required tests — this is a meaningful evidence roadmap for hardware validation. Work packages provide decomposable, testable deliverables.
Weaknesses: The framework's testable predictions are primarily process-level (SPICE curves, attractor convergence, retention time) rather than system-level performance claims. There are no quantitative targets for system-level metrics — e.g., what classification accuracy, energy-per-inference improvement, or latency advantage over conventional approaches would confirm the architecture's value. The motivation for why latched Hamiltonian relaxation specifically offers advantages over simpler Gm-C attractor circuits is asserted but not connected to measurable phenomena. The Boltzmann generator analogy (Section 12) provides conceptual justification for tiled architecture but no quantitative scaling prediction. The framework does not identify existing experimental observations or anomalies that motivated its development — it is purely a design prescription. This limits evidence_strength to 3: the hardware falsification roadmap is clear, but the system-level performance claims that would distinguish RHSF from prior art are not quantitatively bounded.
The RHSF Detailed Design Basis is a carefully scoped engineering planning document for a Mead-Mahowald-derived analog VLSI attractor-computation substrate. The panel's fixed scores reflect a framework that is well-organized and intellectually coherent at the block level (clarity: 4/5, falsifiability: 4/5, completeness: 4/5) while carrying meaningful technical debt in its central dynamical model (internal_consistency: 2/5, mathematical_validity: 3/5, novelty: 3/5, evidence_strength: 3/5). The following narrative synthesizes the specialist findings and identifies the specific locations where work is needed.
The most significant and pan-panel concern is concentrated in Section 10's RHSF Gm-C cell equations. Three of the four math specialists and both science specialists converge on the same structural gap: the momentum-node ODE (C_int dVp,i/dt = Gm2(λi)Vq,i − Gnl Vq,i³ + Gmf(λi)V̄q − Gdiss Vp,i + Iu,i^latched) and its companion (C_int dVq,i/dt = Gm1 Vp,i) are stated as the RHSF Hamiltonian cell without derivation from either a circuit netlist or an explicit scalar Hamiltonian, port-Hamiltonian decomposition, or Lyapunov function. This is flagged as a HIGH-severity mathematical risk flag by two independent math specialists. The panel's central_consistency score of 2/5 reflects the finding that the term 'Hamiltonian relaxation' — which appears in the framework title, the executive summary key design thesis, and throughout Sections 10–11 — acquires a different meaning between its explicit conservative-mechanics definition in Section 11 (dq/dt = ∂H/∂p, dp/dt = −∂H/∂q) and the driven, dissipative, nonlinear system actually implemented in Section 10. The document explicitly acknowledges this gap ('RHSF is damped and driven, not purely conservative'), but acknowledgment alone does not constitute a mathematical bridge. The missing object is a single consistent mathematical artifact — whether a port-Hamiltonian decomposition, a Hamiltonian-plus-Rayleigh-dissipation split, or a named Lyapunov function for the attractor-convergence claim — that would justify retaining the 'Hamiltonian' descriptor for the implemented dynamics. One math specialist (claude-opus-4-8) assigns internal_consistency 3 on the grounds that the explicit disclaimer in Section 11 partially cures the drift; the majority (gpt-5.2, DeepSeek-V4-Pro, claude-opus-4-8 in the consensus round) hold at 2 because the drift is load-bearing given the framework title and central thesis. The panel resolved this at 2/5. Additionally, the mean-field coupling variable V̄q used in the Section 10 ODE is undefined: no normalization, averaging graph, boundary condition, or scaling with active cell count Na is provided. This is flagged by all three math specialists and directly affects whether any energy-like function could exist for the coupled system.
A secondary cluster of mathematical risk flags (all rated MEDIUM by the math specialists) concerns dimensional completeness. The quantities sk,d(t), P(i,k,d), ui(t), Gin,i, and the weights wi in the commit functional Eacc(t) = Σi wi|Vacc,i| are not given explicit unit conventions; the coupling constant Gnl must carry units of A/V³ to make the cubic term −Gnl Vq³ dimensionally consistent as a current, but this is never stated. The parameter λi appears in Gm2(λi) and Gmf(λi) in Section 10 without definition, leaving the mean-field coupling mechanism underspecified at the circuit level — a concern independently raised by both sources specialists. The small-signal transconductance linearization Iout ≈ Gm(V+ − V−) is correctly flagged as approximate in Section 4 but later Gm-C dynamics in Section 10 use linear conductance terms without propagating the operating-range restriction or connecting it to the nonlinear tanh saturation. Finally, the SPICE skeleton in Section 15 contains a self-referential behavioral hold source (BHOLD HOLD 0 V = if(V(TRIG)>0.5, V(ACC), V(HOLD))) flagged as LOW severity but noted by all specialists as relying on simulator-specific implicit state semantics rather than constituting a mathematically explicit sample-and-hold model. The document already labels it a placeholder, which is the appropriate handling.
Where the document is strong, it is genuinely strong. The weak-inversion MOS current relation (Section 3), the thermal voltage definition Ut = kBT/q ≈ 25.8 mV, the differential-pair tanh characteristic Iout = Ib tanh[κ(V+−V−)/(2Ut)] and its small-signal limit Gm ≈ κIb/(2Ut) (Section 4), and the accumulator/leaky-integrator ODEs with τ = Rleak Cacc (Section 9) are all dimensionally sound and correctly stated standard results from Mead-style analog VLSI practice. The six-level SPICE commitment ladder (Section 14) with explicit per-level acceptance criteria and the seven-item risk/falsification table (Section 17) represent genuine engineering discipline that is rare in early-stage analog design proposals. The signal-flow architecture (physical field → receptor evidence sk(t) → projection ui(t) → accumulation → latch → RHSF relaxation) is consistently maintained across Sections 2, 8–10, 13, and 14. The bounded-fabric constraint Na(t) ≤ Nmax is introduced in Section 8 and respected consistently throughout design rules and work packages. The generalization of the Mead-Mahowald photoreceptor to a domain-agnostic receptor transducer (Section 6, with mapping table) is a well-executed conceptual extension. On evidence_strength, the panel scores 3/5 under framework-mode evaluation: the hardware falsification roadmap is credible and modular, but no quantitative system-level performance targets (energy-per-inference, classification accuracy, convergence time relative to competing architectures) are specified that would distinguish a successful RHSF implementation from prior analog attractor circuits. Citation hygiene issues affect References [1], [2], and [4], which appear in the document with truncated identifiers; the underlying works (Mead & Ismail 1989, Mead Adaptive Retina 1989, Schebek et al. 2026) are plausible and well-known, but the identifiers as listed could not be independently confirmed — these should be corrected to full bibliographic form before the document is used as a formal design-basis record. No specialist described any reference as fabricated; the issue is identifier incompleteness only.
This work departs from mainstream consensus physics in the following ways. These are not penalties - they are informational flags that highlight where the author proposes alternative interpretations of physical phenomena. The scores above evaluate rigor, not orthodoxy.
- ◈The framework does not depart from mainstream physics or engineering consensus in any notable factual sense. It operates entirely within established subthreshold MOS physics, analog VLSI design practice, and Hamiltonian mechanics as a theoretical reference. The main departure is terminological rather than physical: using 'Hamiltonian relaxation' to describe a driven, dissipative, nonlinear Gm-C system is a conceptual extension of the Hamiltonian framework beyond its standard conservative-mechanics definition, but this is an internal consistency and naming issue rather than a claim that contradicts observational data or established physical law.
This review was generated by AI for research and educational purposes. It is not a substitute for formal peer review. All analyses are advisory; publication decisions are based on numerical score thresholds.
Key Equations (3)
Subthreshold MOS drain current in weak-inversion; basis for low-current continuous-time analog computation.
Mead-style differential-pair transconductor: saturating tanh I–V characteristic and small-signal transconductance expression used to set time constants and coupling gains.
RHSF momentum-state dynamic equation: latched forcing current drives a second-order Gm-C element with nonlinearity, mean-field coupling, and dissipation.
Other Equations (3)
Clock-free, leaky charge-domain accumulator for projected signed evidence; forms the asynchronous memory prior to commit.
Companion coordinate-state equation forming the q–p second-order Gm-C dynamical pair that implements Hamiltonian-like relaxation.
Asynchronous commit detector functional that triggers latch when weighted accumulated evidence exceeds a threshold.
Testable Predictions (4)
A Mead-style subthreshold differential transconductor macrocell will exhibit I_out versus ΔV matching a tanh characteristic and an extracted small-signal G_m that follows G_m ≈ κ I_b / (2 U_t).
Falsifiable if: Measured I_out(ΔV) and extracted G_m from SPICE or silicon deviate outside the design tolerance such that the tanh shape or G_m vs bias relation cannot be fit within specified error bounds.
Clock-free differential accumulators implemented with C_acc and pseudo-resistor leakage will retain projected evidence over the target accumulation time T_acc (i.e., leakage and kT/C noise remain within the design window so evidence is conserved until commit).
Falsifiable if: Measured leakage, retention, or noise cause the accumulator voltage to decay or vary such that committed evidence is lost or corrupted before commit (retention below specified threshold over T_acc).
A single RHSF Gm-C cell driven by a latched positive or negative injection will produce the expected attractor polarity; small tiles (2–10 coupled cells) on a mean-field bus will converge collectively and yield readout confidence matching input-sign classification criteria.
Falsifiable if: SPICE or device-level tests show the single cell does not settle to the expected polarity, or coupled tile readout does not converge reliably to input class within specified time/noise margins or confidence thresholds.
The proposed duty-cycled architecture (low-current accumulation, bias-gated RHSF inference) will provide an energy advantage over an always-on core, measurable as lower energy per commit when reporting E_acc, E_trigger, E_latch, and E_inf separately.
Falsifiable if: Measurements of accumulation and inference-phase currents and energies show no net energy advantage (energy per commit not lower than an always-on baseline) or duty-cycle overheads negate expected savings.
Tags & Keywords
Keywords: analog VLSI, subthreshold MOS, Mead-Mahowald retina, retinomorphic receptor lattice, Gm-C circuits, transconductance amplifier, Hamiltonian-like dynamics, clock-free accumulation, SPICE validation, duty-cycled inference
You Might Also Find Interesting
Semantically similar frameworks and papers on TOE-Share