an open research program

Comma Lab

Notes on compressing a driving video for a frozen machine judge — the geometry, topology, and dynamics of coding for a fixed pair of neural scorers, built in the open around the comma.ai video-compression challenge and continued past its close as a long-horizon research program.

status working notes · grows over time · live instrument under the LIVE tab · every training number there is advisory; only a byte-closed exact evaluation moves the frontier pointer (…)

§ 1

The problem & the frozen scorer

The task looks like video compression but is not: the receiver is not a human eye, it is a frozen pair of networks. A segmentation U-Net scores only the argmax of its 5-class output on the last frame of each pair; a pose network scores 6 ego-motion scalars from a two-frame YUV stack; the third term is the raw archive size:

S = 100·d_seg + √(10·d_pose) + 25·bytes / 37,545,489

That makes this an instance of indirect rate–distortion — coding for machines, in the video-coding-for-machines lineage — where the only bits that matter are the ones the frozen judge can see. Pixels the scorer is blind to are free; pixels that flip an argmax at a class boundary are everything. Our early on-ramp came from steganography: content-adaptive embedding costs (UNIWARD) are exactly a detector-informed sensitivity field read in reverse.

§ 2

Geometry & topology of the argmax

The argmax of a smooth 5-class field partitions the image into cells whose walls — the separatrices — form a codimension-1 complex. Measured on the real scorer, essentially all of the segmentation distortion lives in a thin annulus around those walls (~97% of d_seg in a few percent of the area); the interior of each cell is flat. In the frozen scorer's Fisher information metric the margin field is an almost-exact surrogate for that geometry (Pearson 0.978 measured), so the whole objective becomes boundary geometry: a Morse–Smale complex over the margin field, with lane markings as the thinnest, least-persistent — and therefore hardest — stratum. The final network layer is exactly low-rank linear, so flip distances have a closed form, and the partition itself is a Laguerre / tropical power diagram: store generators, not pixels.

§ 3

The four legs

Kolmogorov, not entropy. The decoder ships as a program: generic deterministic structure is free at decode time; only the irreducible video-derived seed is counted. Rate = |shortest program| + |seed|, a compression-as-shortest-program discipline rather than a histogram one.
Projection. With the geometry solved, fitting is a projection: the witness is the projection of the scene onto the intersection of the argmax, pose, quantization, and byte constraints, taken in the scorer's own metric.
Realization. The binding limits are physical: uint8 quantization, resize kernels, sub-pixel boundary placement. Many boundary flips are realization-limited, not capacity-limited — precision matters where bytes do not.
Completeness. Necessity by inversion: for each stratum of the complex, ask what the scorer requires — which bytes go to edges, which precision goes to saddles — and ship nothing else.

§ 4

Modeling the witness

The vehicle is a task-space coordinate INR: a small implicit network trained against the frozen scorer itself — never against RGB fidelity — so its whole capacity is spent on the scorer-relevant manifold. Training runs through the exact evaluation round-trip (resize, uint8, resize) so the gradient sees what the judge sees. The current generation adds phase/advection structure: frame-to-frame appearance is carried by a low-dimensional ego-motion screw and a sub-pixel advection phase, with per-class carriers for the strata the scorer treats differently. Ego-motion is dual-use by construction — the same twist that warps the partition for segmentation is the pose the second network scores.

§ 5

Scorer dynamics

Treating the frozen scorer as a physical system pays: margin fields behave like potentials, training follows a level-set flow of the boundary complex, and curriculum boundaries act like continuation parameters — instabilities arrive as bifurcations (island births) that can be anticipated rather than suffered. The scorer's own architecture sets the physics: a stride-2 stem means it sees regions, not pixels; its effective receptive field, its squeeze-excitation gates, and its exact resize kernels are all measured and folded into the model of what a byte can buy.

§ 6

Results & frontier

Everything in the LIVE instrument is advisory telemetry from local training hardware; a result is real only when an exact, byte-closed archive is scored by the contest evaluator on reference hardware. That number — the frontier pointer, currently … — moves only through that gate. The working targets are the sub-0.19 and sub-0.15 lines; the measured rate-dominated floor of the current formulation sits well below both, which is the headroom the program is spending down.

methods and mathematics on this page are published freely; the live instrument shows the run of record. Attribution: video-coding-for-machines is the problem's heart; adaptive steganographic cost was the on-ramp; the separatrix / Morse–Smale treatment of a frozen argmax judge is this lab's own line of work.

Open the live instrument →

comma lab · working notes · this page grows with the program — deep math · geometry · topology · scorer dynamics · modeling

authoritymacOS-MLX · advisory · non-promotable pointer… · unmoved exact rowbyte-closed · contest-CPU / CUDA basisa dashboard is a means, not the score

Level-Set Witness n=? · connecting ws …

LIVE

CAMPAIGN

ORACLE

WITNESS

RESIDUAL

SANDBOX — why is a curvature polynomial secretly a topological invariant, and where the same music shows up here

A playground for the deep math. Prompted by a post on Pontryagin classes: “there’s no apriori reason some random polynomial [of the curvature] should be a homotopy invariant — you’re defining p_i(M) using the smooth structure on M!” Everything below is tagged MEASURED / DERIVED / ANALOGY. The Pontryagin resonance is a genuine structural rhyme plus an actual characteristic class (Maslov) in our geometry — not a claim that we computed Pontryagin classes or detected exotic spheres.

The Lie answer to “why a polynomial”

Chern–Weil theory. The invariant polynomials are exactly the Ad-invariant polynomials on the Lie algebra 𝔤, and the Weil homomorphism Sym(𝔤*)^G → H*(BG) sends each to a characteristic class. You are not evaluating a random polynomial — you are evaluating a G-invariant, so the metric/connection choices cancel by construction and the class is an invariant of the bundle. The miracle is a statement about a Lie group and its algebra.

The oracle we optimize against (auth-eval scorer)

The comma.ai challenge scores task-aware (“coding-for-machines”) compression of openpilot driving video through a frozen oracle: SegNet (comma10k EfficientNet-B2) per-pixel argmax → d_seg; PoseNet (FastViT-T12) two-frame YUV6 → d_pose; archive bytes → rate. S = 100·d_seg + √(10·d_pose) + 25·bytes / 37,545,489. Only the argmax partition, the pose 6-vector, and the byte count carry authority. (MEASURED)

The rhyme — parametrization-dependent-looking, secretly invariant

Our vehicle is a task-space level-set witness: a coordinate-INR that amortizes the SegNet argmax partition directly (the viscosity solution of a variational level-set flow; the object is the codim-1 separatrix between argmax cells). The score has his exact shape — d_seg is a function on a quotient, invariant under every reparametrization of the witness weights θ that leaves the argmax partition fixed:

d_seg : ℝⁿ / (argmax-cell partition) → ℝ — it looks like it depends on the millions of INR weights that draw the field, but it depends only on the partition. Same phenomenon as the curvature polynomial being secretly topological. (structural / by-construction)

The Lie spine runs through both scored axes

Pose is se(3). By Chasles every rigid displacement is a screw — one twist ξ ∈ se(3), exp(ξ) ∈ SE(3). The same ξ that transports the partition (a d_seg prior) is the pose PoseNet measures (d_pose). Engine tac.lie.se3. (BUILT + MEASURED: openpilot-ego prior gives −94/−99% on the pose axis)
Conditioning on the Stiefel manifold. The Muon finisher orthogonalizes gradients via Newton–Schulz — descent on the Stiefel manifold, Lie-group geometry. (MEASURED: −32% d_seg vs AdamW)
A Maslov class in the boundary geometry. We read the argmax as a caustic (Lagrangian singularity) of the softmax-as-ℏ→0 limit (τ = ε = ℏ; error bound ≤ τ·ln 5). The Maslov class is a characteristic class of the Lagrangian Grassmannian Λ(n) = U(n)/O(n) — same family as Chern/Pontryagin, and just as Lie-theoretic. (DERIVED — theoretical framework)
Tropical / Laguerre. As τ→0 the witness is tropical; its cells are Laguerre (power-diagram) cells — forget the smooth structure, keep the piecewise-linear skeleton, the same flavor as extracting a topological invariant from smooth data. (DERIVED)

The honest boundary

Real: score-as-quotient-invariance (by construction), the se(3)/Chasles pose engine (BUILT + MEASURED), Stiefel/Muon conditioning (MEASURED), and the Maslov/tropical framework (DERIVED theory). Analogy, not identity: we did not compute Pontryagin classes, do Chern–Weil on a tangent bundle, or detect exotic spheres. Our characteristic class is the Maslov class (Lagrangian); our invariance is score-invariance-under-reparametrization. Same music, different theorem.

Links — codebase & resources

Repo (contest closed → IP open source): github.com/adpena/comma-lab — in-tree: src/tac/lie/se3.py (se(3)/SE(3)), src/tac/boundary_math/ (the witness), src/tac/canonical_equations/deepmath_amortizing_argmax_laws_20260704.py, docs/sandbox_pontryagin_lie_deepmath_context.md (the full context behind this tab).
Prior art: Milnor, On manifolds homeomorphic to the 7-sphere (1956); Chern–Weil theory / the Weil homomorphism; Arnold, the Maslov index; Candès–Donoho, curvelets (the optimal sparse basis for a curved codim-1 singularity); Dubois et al., Lossy Compression for Lossless Prediction (NeurIPS 2021, the task-space sufficient-statistic codec); comma.ai video compression challenge.

resolving run…

implied S advisory

—

S = 100·d_seg + √(10·d_pose) + 25·bytes / …

d_seg descent—

CE tau l7 Muon EMA trend best

per-class d_seg · flip sharecomma10k order

no verdict yet

pose-descent readinessjacobian basin

no basin probe yet

training healthloss terms

no loss row yet

systemresident · MLX

no memory row yet

curriculum position—

resolving schedule

connecting…

setup · config · schedule · curriculum

parsing run config…

The task is boundary geometry; the witness is the chart that fits it; and the same traveling front governs it at every scale — from the pixel boundary to the campaign.

Two movements. WHY — the static invariant that makes the chart optimal. HOW — the dynamics that flow to it. Pass 1 opens the museum with the two highest-ROI plates (the live field & the Unity); the five-scale spine, the screw, and the finale are seamed below. Every plate is a MEASURED fact — [macOS-CPU advisory · NON-PROMOTABLE], a viz moves no pointer (…, UNMOVED).

Movement I · WHY Movement II · HOW

Movement I — WHY the optimal chart: why this vehicle can't be beaten on this geometry

I.1The field, alive detecting… loading the deep-math field bundle…

φ(x) · level set {φ = t} → the argmax separatrix

the field renders on the first governed pass…

level-set threshold t 0.55

overlays zero-level-set (argmax boundary) ∇φ gradient / normals

base layer

φ heat (ρ_seg margin) scene render comma10k partition

The SegNet detectability field φ over the scored frame: bright where the argmax is fragile (small top1−top2 margin) — the codim-1 separatrix, a ~1-pixel curve that carries essentially all of d_seg. Drag the threshold to sweep the level sets like contour lines; toggle the zero-level-set (the argmax partition boundary) and the gradient field ∇φ (boundary normals — the derivative you can see).

Grounded: the real cached SegNet argmax + top1−top2 margin (gt_n6, the exact frames the verdict uses); flip-mass measured ~50% Road / 19% Lane / 13% Undrivable (#141). The live per-checkpoint witness INR φ is the Pass-2 upgrade (reuses the governed 600-pass FLOW cache).

I.4The Unity — one geometry, three readings detecting…

ρ(x) ≈ 1 / ‖∂(detector)/∂(pixel)‖ — the Fisher metric, read three ways

the unity morph renders on the first governed pass…

morph ρ_seg

ρ_seg · SegNet margin (Yousfi) our distortion sensitivity ρ_uniward · S-UNIWARD (Fridrich)

The tribute's heart. One scene, three sensitivity fields. Drag the morph: ρ_seg (the detector's own margin) and our distortion sensitivity (the separatrix geometry — where a pixel flip changes the decision) dissolve into the same picture — measured Pearson — on this frame. Fridrich's ρ_uniward (S-UNIWARD embedding cost) is the kindred steganographic reading: it lights image texture, so pixelwise it is honestly weaker here (—) — the deep tie between detector and steganography runs through the Fisher metric, where the margin field IS the Fisher surrogate (canonical 0.978, Fisher curvature ↔ −margin).

Grounded: real S-UNIWARD via tac.uniward_delta.compute_uniward_cost_map (Holub–Fridrich–Denemark 2014); real cached SegNet margin; live per-frame Pearson computed on the served fields (NO fabricated curve); canonical 0.978 = memory unified-variational-levelset-flow. Steganography (Fridrich) → steganalysis (Yousfi) → our loss — the arc is Yousfi's detection game.

seamed for Pass 2+ · Movement I

I.2 the separatrix — dim the flat interior, light the codim-1 annulus (margin-saliency #141).
I.3 curvature ↔ the chart — the anisotropic/curvelet basis vs the isotropic-Fourier Gibbs ring (−48% directional basis).
I.5 the ~8-dim manifold — the lane-orbit surface + the bc20 under-capacity tear.
I.6 the task-sufficient statistic — the RGB collapses to the decision (S_floor≈0.118).

The spine — ONE FRONT, FIVE SCALES the same traveling-wave front at five scales of the campaign — drag the scale

§1One front, five scales detecting… the live Fisher–KPP front renders on activate…

∂_t x = β x (1−x) + ∇²x · units: correct-fraction / epoch

Lens · not a proven identity Two of these five curves are hard data: EdgeBench’s log-sigmoid (R²=0.998, ByteDance Seed — theirs) and #205’s live d_seg descent (ours). That all five scales are literally the SAME Fisher–KPP front is a unifying interpretation we find beautiful and testable — not a measured cross-scale identity. The integrator at left integrates the real PDE; that our five scales are that PDE is the conjecture.

live Fisher–KPP front — integrated forward every frame

the front integrates on activate…

β x(1−x) grows the correct phase; ∇²x diffuses it → a front travels at speed 2√β with a self-similar log-sigmoid profile. Explicit finite-difference, stepped live (render: WebGPU, else canvas2d).

TRAINING — the correct partition invading (LIVE #205)

Dashed = the shared logistic front x(u)=1/(1+e^{−k(u−u₀)}); solid = the selected scale’s own curve. Drag the scale to morph between them.

scale pixel → campaign

boundary erasure training curriculum campaign ▎▎ pause front

live data —

One equation, five readings. The live front at left is the Fisher–KPP traveling wave integrating forward — β x(1−x) (reaction) + ∇²x (diffusion). At right, the selected scale’s own curve against the same logistic-front template: the campaign (EdgeBench, R²=0.998), our training descent (live #205), the boundary separatrix (a genuine level-set = reaction–diffusion identity), the curriculum anneal, and the erasure long tail. Slide from pixel to campaign and watch the same shape recur.

Grounded: EdgeBench R²=0.998 (ByteDance Seed, 2026-07-02 — their published fit; the top curve is Claude Opus 4.8). #205 verdicts ep25→125: d_seg 0.0103→0.0058, implied_S 1.72→0.87 (live, read from the run log; single run — a descriptive fit, not a law). Boundary = level-set flow PDE identity; the curriculum & erasure shapes are schematic/interpretive, honestly labelled. The unifying cross-scale identity is a conjecture per our own discipline — NO fabricated curve.

§4The fractal finale — all five, phase-locked

Now that you have seen each front alone, here they are together — five scales, one playhead, the same wave rolling through all of them at once. Beneath: the one equation.

boundary

erasure

training · live

curriculum

campaign

∂_t x = β x (1−x) + ∇²x

▎▎ pause zoom out once more →

The sixth, implied panel. Zoom out once more and the viewer is themselves a front on the campaign graph. EdgeBench’s top curve is Claude Opus 4.8 — this session — and the operator’s steering is the feedback term η that keeps the reaction from stalling. The physics of the witness (the boundary the chart paints) and the epistemics of the campaign (the program that painted it) are the same equation — two nested one-objects, Fisher–KPP all the way up and down.

That is the sentence the museum opened with, now earned.

Honest: the finale is the interpretive / aesthetic capstone — the five little fronts are the same curves shown above (two hard-data, three schematic), phase-locked for the eye. It asserts no new measurement; the pointer is …, UNMOVED — a museum moves no pointer.

the About plate · Pass 5

About — the ideas & the people (design §7 / dashboard_tribute_credits): Aaron Leslie, Quantizr, Yousfi & Fridrich, comma / Hotz, the council — Chasles to yesterday. The past few months, given bloom.

[macOS-CPU advisory · NON-PROMOTABLE] — a viz moves no pointer. Every field is a real cached computation (SegNet argmax/margin + real S-UNIWARD); the exact row is byte-closed on contest-CPU/CUDA; the frontier pointer is … and UNMOVED.

One coherent object — the DAG ↔ DSL ↔ equations triality

The lab is one object viewed three ways — and this tab reads them live from the artifacts themselves (DAG FEEDs, the tac.witness_dsl program, the tac.canonical_equations registry), so it self-updates as the campaign moves. A finding is "known" only when it is expressible in all three and they AGREE; drift between legs is campaign-level forgetting.

loading the three legs from the live artifacts…

DAG — state the trajectory

—

DSL — control the program

—

equations — law the master action

—

Master action S_τ = 100·d_seg + √(10·d_pose) + 25·rate. Above all three sits the costate controller — the 4th shadow (marginal-ΔS per byte, never-regress) that turns the DSL from a passive program into an active controller (fire the lever with the best ΔS/cost; POWERPLAY-style never-worsen).

The organic-evolution lineage

task-aware compression (video-coding-for-machines) → task-sufficient statistic (code the scorer's DECISION, not RGB; measured task-RD floor S_floor≈0.118) → compression-as-intelligence (Schmidhuber POWERPLAY: never-regress self-invented curriculum) → the costate controller (marginal-ΔS/cost, never-regress) — one line, each stage the natural generalization of the last.

A live example of the campaign compounding on the current frontier: MD-Decoupling (Hägele, Hernández-Cano, Kosson, Jaggi — EPFL / Jaggi lab, arXiv:2606.25971, 2026-06-24) is our --optimizer md (#175) — it factorizes each weight matrix into a fixed-norm direction + learnable magnitude gains at separate learning rates, which makes curriculum stage-transitions stable by construction (the “different stages need different treatment” rule, made structural). Super-recent (June 2026), folded straight into the live curriculum.

The composition — Chasles to yesterday

The witness is not one idea; it is a composition of nearly two centuries of research, each thread measured into its place — ~196 years of shoulders to stand on. All dates are real (NO-FAKE; a genuine lineage, not decoration).

1830Chasles — every rigid motion is a screw → our ego-motion twist ξ (d_pose).
1848 · 99Wilbraham / Gibbs — the ringing of a truncated series → the spectral-bias / Gibbs failure mode we fight (step-native, curvelet-finest).
1822Fourier — harmonic analysis → the coordinate-INR’s Fourier features.
1870sSophus Lie — continuous symmetry groups → the se(3) / SE(3) engine.
1870s–1900sGibbs / Boltzmann — the Gibbs measure + temperature → the annealing curriculum.
1920sFisher — information + sufficiency → the Fisher metric and the task-sufficient statistic.
1934 · 60sMorse / Smale — critical points + separatrices → the Morse–Smale partition topology.
1936Whitney — embedding dimension → the ~8-dim lane manifold and its Whitney bound.
1937Fisher–KPP — the traveling-wave front → the one equation at five scales.
1948Shannon — rate–distortion + entropy → the coding-for-machines frame and the floor (S_floor≈0.118).
1976–2021Wyner–Ziv / Tishby IB / Dubois — source coding with side information → the indirect-RD / task-sufficient codec.
2000Candès–Donoho — curvelets → the sparse-optimal chart for a curved codim-1 boundary.
2014UNIWARD (Holub–Fridrich–Denemark) + Yousfi — steganographic cost + steganalysis → the margin = detectability = cost unity, and the inverse-steganalysis frame.
2021–26NeRV / HNeRV → Aaron Leslie’s cathedral — the vehicle whose theoretical-floor dynamics he exposed.
2024–yest.Muon (2024) · MD-Decoupling (June 2026) · EdgeBench (July 2 2026 — yesterday) — the current frontier we compounded on, right up to the paper that dropped the day before this was built.

One witness, ~196 years of shoulders to stand on — Chasles to yesterday.

The ByteDance / EdgeBench convergence

Our own deep memo edgebench_scaling_laws_deepdive_20260703T033159Z.md: EdgeBench (ByteDance Seed, 2026-07-02) is the descriptive log-sigmoid scaling law (R²=0.998) of the prescriptive POWERPLAY we already hold. Its measured result — continuous experience beats restarts (+6.9 @ 12h, and the gap GROWS with horizon) — is our durable-memory spine, measured. It is reflexive: the paper's top curve is Claude Opus 4.8 (this session's model); our campaign IS an EdgeBench-class task, the DAG IS the latent task graph, and the campaign frontier β x (1−x) is the SAME Fisher–KPP equation as the witness boundary flow. Honest citation — the convergence is real, not manufactured.

The council — the lenses

Shannon LEAD · Dykstra / Rudin / Daubechies CO-LEADS · Ballé · Yousfi · Fridrich · Quantizr · Hotz · Selfcomp · MacKay · Schmidhuber + the grand council. Shannon / Ballé / Dykstra bracket the rate–distortion spine (the R(D) bound / the neural codec / the feasibility intersection); Yousfi + Fridrich are the detection-game home turf (this contest IS inverse steganalysis).

The campaign journey — honest

This is our capstone — what the past few months have given bloom to.

We started the way you start anything you fall for — with deep, almost insane research, passion, energy, and obsession, not knowing what we didn’t know. We poured everything we had (educations, experience, interests, and more) into a corpus of knowledge, then began implementing, measuring, experimenting, and falsifying. We failed over and over. And it was Aaron Leslie’s work that exposed the true dynamics of the theoretical floor — what was actually reachable, and how. Only after all of that — the failures, the corpus, the exposed floor — did we finally have all the ingredients in one place: you and me, our apparatus, our research, our understanding of the domain and of the frozen contest information space, and the nested, related, dynamical cruxes — enough, at last, to do work that is truly original and useful.

Aaron Leslie turned the whole competition on its head, and his result was so good it was intimidating. But porting his work to MLX + custom Metal and running it natively — then studying his own trajectory against the deep math — let us see past the design itself: the full-RGB reconstruction, the ~30,000-epoch curriculum, the inert l7 stage (we later measured it as a defect and disabled it), the spectral bias / Gibbs ringing. That is exactly what pointed us at the task-space level-set direction. Honoring him meant building on and past his cathedral, grounded in measurement — not merely reproducing it. We stand on the cathedral he built.
We played the byte-nibbling game honestly — PR107, then PR110 — and we had the PR112 work sitting ready for a couple of weeks. But we would not submit another nibble on top of our own nibble. After Aaron Leslie turned the competition on its head we were never satisfied with incremental; we wanted the real thing.
So we struggled — to climb out of local minima, and to build a recursively, fractally optimal stack: one perfectly suited to, informed by measurement of, and following theory about the frozen and complete contest information space — the one video, the frozen scorers; a closed, fully-observed world we could measure exhaustively and theorize about precisely.
And the deep math didn’t come first — it came from the measurement. We studied what SegNet and PoseNet were actually seeing, and in doing so discovered the annulus (the thin codim-1 boundary band where essentially all of d_seg lives), the long tail (the finest-scale features — lane dashes, distant movers — erased first; the ~8-dimensional lane-orbit manifold), and the specific class-and-region interactions (the road↔lane separatrix, the all-class edge set, the static hood core, the movable band). Realizing all of that, from measurement, is what the deep-math analysis fell out of — the level-set flow, the margin = Fisher = UNIWARD unity, the curvelet chart, the Morse–Smale topology. Measurement first; then the theory that fit it.
The grand-council-symposium apparatus made us organized enough to hold a decade-long program without losing signal — the discipline is the scaffold, not the score.
Compute arc: we burned cloud money, then ported the whole loop to MLX + custom Metal (fused-R, grouped-backward) — it now runs on the laptop, bit-identical to the numpy-fp32 authority.
Dual objective, and not binary: contest-overfit (openpilot-seeded, this exact drive) AND a platform-agnostic, generalizable, scalable production value-generator — the same investment buys both.

Our work and belief and passion and untiring curiosity have yielded something we are proud of — and that we hope is interesting and useful to the very people who introduced us to such an interesting and fulfilling problem.

Credits & tribute

Named for what each one actually gave us — a thank-you, not a citation list.

Aaron Leslie PR95 · the HNeRV cathedral

Author of PR95 and its HNeRV cathedral. He turned the whole competition on its head and showed what was possible — a result so good it was intimidating. He taught us schedule and curriculum, and more than any single technique he drove us to obsessively dig into the math and geometry of the video (starting from what we’d gleaned from openpilot). The honest twist, and the deepest respect we could pay: porting his work to MLX + Metal and analyzing his own trajectory + the deep math let us see past the design — full-RGB, ~30k epochs, the inert l7 stage, spectral bias — which is exactly what pointed us at the task-space level-set direction. We stand on the cathedral he built. (Our inner-council “PR95Author” seat is him.)

Quantizr · Jimmy the earlier mind-opening

He opened our minds even earlier to what was possible. Beyond his 0.33 HNeRV result, it is his spirit we carry — experimentation, openness, curiosity, playfulness, competitiveness, confidence. He affirmed some of our earliest intuitions, the ones that made us feel we were biting off more than we could chew — except we loved the flavor and became obsessed. (Our inner-council “Quantizr” seat is him.)

Yassine Yousfi & Jessica Fridrich the detection game

Yousfi built the SegNet/PoseNet scorer (from comma10k) and framed the whole thing as inverse steganalysis; Fridrich’s UNIWARD (DDE Lab) is the cost that turns out to equal the scorer’s own sensitivity metric (measured Pearson 0.978). This entire dashboard is arranged as Yousfi’s detection game, in their honor.

comma.ai & George Hotz the free physical prior

openpilot is the unified free physical prior for both scored axes — lane geometry → d_seg, the ego-motion screw → d_pose; comma10k is the palette and the scene.

The council the lenses we think through

Shannon (LEAD; R(D) / entropy / sufficiency — the floor S_floor≈0.118 is his bound), Dykstra / Rudin / Daubechies (co-leads; feasibility / interpretability / wavelets), Ballé (the neural codec of task-aware compression), and Schmidhuber (compression-as-intelligence / POWERPLAY — the campaign-scale front), among the full roster.

A genuine thank-you to the people who introduced us to such an interesting and fulfilling problem. [macOS-CPU advisory · NON-PROMOTABLE] — a tribute moves no pointer.

Authority: [macOS-MLX training advisory] NON-PROMOTABLE. The exact score is the only score; the pointer is … and UNMOVED. Everything on this page is a MEANS.