Energy First Architecture — from mechanism to the fabric

Act I — the mechanisms, proven at nano scale

Each idea, isolated and driveable

Before integrating, every EFA mechanism was built as a self-contained on-device demo and verified — not argued. The sparse-positive latent holds a Hebbian memory and is a plannable world model; a dendritic gate lets one model learn several physics-worlds in sequence without forgetting; and a policy can be trained entirely inside the learned world model — Dreamer's "learn in a dream" — at a fraction of the real experience.

The nano program also mapped the honest limits — that amortizing a planner into a fast reactive policy is a structural dead-end at fixed capacity, and that the one thing which reliably improves with scale is world-model multi-step fidelity. That single finding set the whole direction.

Act II — where EFA sits

It doesn't pick a frontier bet — it unifies them

Each EFA mechanism maps to a distinct state-of-the-art line. No single comparable does all of them in one continual, on-device model. That unification — not a benchmark win at toy scale — is the claim.

Act III — EFA on the fabric

It trains and acts on the real cross-fabric stack

The nano proof pointed at scale, so EFA moved onto Ferric — one pure-Rust compute stack that runs the same code on a datacenter GPU, a laptop, and a browser tab. There the EFA world model trains with autograd + Adam batched on the GPU, and acts by planning — the full perceive → plan → act loop on the fabric.

On the maze, pure energy-descent is provably trapped at the wall (the goal's latent looks close through it — a local minimum), so acting needs a learned value of cost-to-go around the gap. A goal-conditioned value, trained with Monte-Carlo returns + Hindsight Experience Replay + a MuZero-style search, routes around it — and improved steadily as each principled lever was added, until a change of algorithm broke the plateau:

value-guided reach: 11% → 13% → 19% → 23% → 39% (+HER · +gap-seed · +MC returns · +lookahead · +Fitted-Value-Iteration & target network)

And 39% is not the wall either. Holding the excellent world model fixed at 0.951 and scaling only the value network's width, cross-wall reach climbs monotonically — still rising at 2× the base net. On the easy grid, scaling saturated because the world model had no headroom; on the hard maze, scale is the lever:

value-net width → reach: 16k → 23% · 33k → 27% · 67k → 39% · 133k → 45% (params → cross-wall reach)

Then the hardest test — and the one that made us walk back a too-tidy claim. A two-wall S-corridor (a ~2× longer detour). At a fixed training budget, small models fail outright (a real capacity floor) and reach rose with size, which first read as a clean super-linear "scale is the lever" curve:

fixed-budget joint size → reach: 26k → 0% · 54k → 1% · 93k → 3% · 250k → 19%

A harder look corrected it. Doubling the training budget took the same 250k model to 43% — so 19% was under-trained, not a size floor. And at that budget reach is non-monotonic in size (43 → 26 → 51%): the 480k model had the best world model (0.977) yet the worst reach, so the noisy bottleneck is the value-learning conversion, not fidelity.

2× training → reach: 250k → 43% · 480k → 26% · 785k → 51% (best config ≈2.7× the first pass)

A direct test settled the dip: fix the 480k model and just give its value net more FVI compute — reach climbs 26% → 37% → 43% (22k/44k/66k iters, still rising), recovering to the 250k model's 43%. So the "dip" was an under-converged value net, not a size penalty.

480k value-RL compute → reach: 22k → 26% · 44k → 37% · 66k → 43% (FVI iters → reach, world model fixed)

Honest read: below ~250k you fail the S-corridor (capacity floor); above it, absolute reach is value-RL-convergence-bound — bigger value nets just need more FVI iterations. So "scale is the lever" holds only in the qualified form — capacity floor, then compute — not a clean bigger-is-better law. The clean monotone result stays the single-wall value-net sweep above; the hard task's noise was convergence, and it resolves with compute.

The other lever: test-time search, not training

EFA is a training-free system by construction — it acts by planning at test time. So there's a second lever entirely: spend more search, not more training. We swapped our greedy depth-3 lookahead for MPPI — population trajectory search in the latent, scored by the value net (the recipe Meta's V-JEPA 2-AC uses on real robots) — on the same world model and value net that gave 39%:

single-wall reach (same value net): energy-descent 2% · greedy depth-3 39% · MPPI K=16 12% → K=64 23% → K=256 69%

MPPI nearly doubles reach — 39% → 69% — with zero additional training, purely by searching harder, and reach rises monotonically with the search budget K. That's the test-time-compute substitution law (the principle behind o1-style reasoning), here demonstrated for control: search substitutes for value-function error.

Priced honestly: at low K, MPPI loses to greedy — our depth-3 is exhaustive (all 4³ sequences), while MPPI under-samples the 4⁶ horizon; it overtakes only at K≈256, where its longer horizon and value bootstrap pay off. So MPPI is a compute knob, not free. Together with the convergence result, this reframes the ceiling: hard-task reach was both value-RL-convergence-bound and test-time-search-bound — more search on a fixed value net buys as much as more value training.

And the memory rule, upgraded gradient-free

EFA's Hebbian memory Z is written at inference by outer products — the primitive corner of the modern test-time-memory design space. Two gradient-free upgrades from that literature — a delta-rule write (store the reconstruction error, not the raw pattern) and a modern-Hopfield softmax read — were characterized on EFA's exact latent. Recall cosine as memory load grows past the dimension d:

Honest reading: the delta rule dominates within capacity (M ≤ d — near-perfect where the additive write already interferes), but beyond M > d no fixed d×d write survives — that's a rank limit, not a write-rule flaw. The Hopfield read is the only thing perfect past capacity, at the cost of keeping every pattern. For EFA this is exactly right: Z does episodic recall (few patterns ⇒ M ≪ d), the regime where both upgrades give near-perfect recall and additive quietly loses ~15%.

The whole stack, one model, on the fabric

Every run above was fully-observed, so the memory sat idle. The final test makes it carry weight: a partially-observed maze where the goal is shown only at t=0, then hidden. The agent writes the goal into its Hebbian Z at t=0, then recalls it each step and plans toward the recalled goal with MPPI — exercising every mechanism together (sparse latent → Z → world model → value → MPPI):

The goal lives nowhere in what the agent sees after t=0, yet it still reaches it — the Hebbian Z is load-bearing. Blind fails (0–5%); with memory, reach matches the fully-observed maze exactly (39 / 69%), so recall is lossless and the MPPI acting win carries straight through the memory. The whole EFA stack now runs as one model on the real cross-fabric stack — the nano fusion, at fabric scale.

Many memories, and the upgrades compound

One goal makes recall trivial. The real test of the memory rule is a landmark tour: K goals shown at t=0, all hidden, visited in order — so Z holds K associations and recall quality gates success. This is where the acting and memory upgrades meet. Fraction of the tour completed:

Two training-free upgrades, compounding. The acting upgrade (MPPI) lifts completion 3–19% → 37–59% — reach precision was the real bottleneck. And once a good planner removes it, the memory upgrade surfaces: delta/Hopfield beat the additive write by +9 points at K=16 (46 vs 37), a gap that widens as the memory fills — exactly the interference the capacity study predicted, now paying off as task success.

Honest bounds: at short tours recall is clean, so the memory rule doesn't matter (additive even edges it, 59 vs 58 — noise); the upgrade earns its place only as the memory fills. And the first MPPI run read 0% everywhere — a harness bug (a planning seed that resampled the same actions in place), caught and fixed before drawing any conclusion.

How good is it — measured against the exact optimal

No fabricated peer numbers. The honest benchmark: EFA's planner against the exact optimal (shortest paths computed by breadth-first search — the true 100% ceiling, stronger than any learned method) and a naive euclidean floor. On reach and path-length optimality:

EFA closes to 87% of the exact optimal and is still climbing with search budget — and it beats the naive floor by solving the detour goals naive can never reach (euclidean-greedy is optimal on easy goals but trapped by the wall; EFA's whole edge is the planning-required ones). The one open gap was path efficiency: EFA reached but meandered — ~2.5× the shortest path.

That gap had a clean fix. Scoring each plan by its terminal value rewards ending up near the goal, not getting there fast — and discards the fact that a plan passed through the goal early. Reward reaching high value early instead (max_h [ V(x_h) − c·h ]) and both axes improve:

Not a tradeoff — both improve. Crediting a plan for reaching the goal at any point in its rollout (not just the end) lifts reach 87 → 96%; a small step-cost tightens paths from 2.48× → 1.93×. EFA now reaches 96% of the exact optimal. The last honest gap: still not path-optimal (1.93× vs 1.00×) — closing that fully needs a distance-shaped value, not the reachability-shaped one the planner learns.

The honest ledger

What's proven, and where the edge is

Every nano finding replicated at scale on the fabric: fidelity is the lever; a heavy variance regularizer over-spreads the latent and hurts planning; energy-descent traps at obstacles; value-guided planning is the right actor for hard tasks. And the one soft number got the honest treatment — diagnosed, then corrected under test:

The science question is answered. EFA works — it trains, plans, acts, and scales on the real cross-fabric stack, every mechanism validated. And on the one hard open number, the honest process — over-diagnose, then correct under test — nearly doubled it. The path onward is engineering, not a redesign.

Internal research record · Charlot Lab · Institute for Physical AI @ Bailey Military Institute. Full design + measurements: v2/docs/SOTA-PAI-MODEL-DESIGN.md. On-fabric code: fourteen EFA examples in ferric/crates/ferric-tensor/examples (efa_world_model · efa_scaling · efa_maze · efa_maze_scale · efa_maze_hard · efa_maze_hard_big · efa_maze_conv · efa_maze_mppi · efa_memory · efa_pomdp · efa_tour · efa_ensemble · efa_benchmark · efa_pathopt) + a sqrt autograd op contributed to Ferric. Nano spikes are driveable, on-device, local-only. Every figure on this page is a measured result, not an estimate.