A field native intelligence
We stopped asking how to copy the brain. We started asking what the brain discovered.
The brain is one realisation of intelligence, not its definition. This project tries to distil the principle rather than imitate the organ: one field, one energy, where thinking is a settling, learning is a change in the landscape, and not knowing is an honest abstain.
Whether that is the cleaner description of intelligence is the open question, and the chapter below is where the argument begins.
Latest status“Do not make the system look intelligent. Make the field become intelligent.”
00 · Origin
Evolution found intelligence. It never explained it. The brain is a working example of matter that perceives, remembers, abstracts, plans, learns and models itself, and it arrives with no theory attached. An existence proof says that something is possible. It does not say what that something is.
01 · Existence, not theory
That is already a great deal, and it is easy to read too much into it. The organism is the working example, not the finished description. So the opening question here is not how to imitate a brain more closely. It is what evolution actually found, and whether that finding has a cleaner form.
02 · The first era of AI
The founding move was never “copy the brain.” It was more precise than that. From McCulloch and Pitts through the perceptron, backpropagation, convolutional nets and the transformer, the field abstracted intelligence as networked computing elements, and then as a paradigm: network plus training plus inference. Today’s systems are already very far from the wet organ. They are still inside that abstraction.
Deep Learning can be an extremely capable path to intelligence without already being its most fundamental description. Transformers are not the opponent here. They are the strongest working instance of one abstraction, and that is exactly what makes the next question worth asking.
03 · The level of abstraction
A bird approximates flight with feathers and a wingbeat. An aeroplane does not flap. Once the law behind the wing is distilled, the organism stops being the thing that has to be copied. So the useful question is not how close a copy can get. It is which description is more fundamental.
Perhaps the brain approximates the pure form of intelligence the way a bird approximates the pure form of flight.
What is the equivalent of lift for intelligence?
Perhaps is doing real work in that sentence. This is a research question, not a finding. Hylaean is a bet on one answer, and the rest of this page has to keep that distinction visible.
04 · The hypothesis
The working hypothesis is that a thought is not a token, a concept is not a label, and a memory is not a database row. All three are states of one dynamical system, stable or metastable. Such a system is described by four properties, and Hylaean is built out of those four and nothing else.
What is possible. The shape of the landscape, the basins that can exist.
What is stable. A valley that holds, or a ridge that cannot.
Where thinking goes. The field relaxes. It does not pick the next word.
What experience keeps. Learning changes the landscape itself, not a separate weight store.
05 · One dynamics, many regimes
If the hypothesis is even partly right, the usual catalogue of mental verbs is not a stack of modules. It is one physics read in different operating regimes.
Read the table as a claim about description, not as a report on twelve finished mechanisms. Some of these regimes are measured on this page, several are still walls, and each one is named with its own claim further down.
06 · One substrate
A typical stack separates training from inference, context from memory, retrieval
from reasoning, confidence from the agent loop. Hylaean tries the opposite:
state.S is a persistent physical state in which experience
leaves traces. Perception, memory, language, world model and reasoning are
meant to be regions of that one field rather than separate machines glued together.
The word meant is deliberate: how far that holds today is what the evidence
sections below have to answer.
Abstain is specified as a consequence of missing stability, not as a confidence
threshold bolted on afterwards. An answer may leave the system only when the field
itself has moved, and four gates have to fire together for that:
moved, energy_work, reached_answer and
not_echo. Where they do not, the honest output is abstain. That is a
design a reader can check rather than a promise that the system never guesses, and
the measurement section is where the checking happens.
07 · The hard ending
This is a wager with first pieces of evidence, not a finished replacement for a brain. The measured cells, and the walls that are still closed, are stated later with named claims. Only no teach scores are reported as intelligence: a system that was shown the answer just before the test is doing recall mechanics, not discovery.
If the hypothesis is real, things should become visible that a story about capacity alone does not require. The five lines below are the open research agenda of this project, not a list of results, and each of them is also a way to be wrong:
What if intelligence isn’t a machine? What if it is a property of certain dynamical systems? Evolution found one. We are trying to understand the class.
The next section is the working metaphor for that class: a material that rings, rather than a network that looks up.
01 · The core idea
The origin chapter asked what evolution discovered. This page’s first working picture is a material, not a network. A piece of metal does not “know” what sound is. Yet strike it and it rings with resonances, modes and standing waves. That geometry is the knowledge. Hylaean is built to work the same way: intelligence as physics, not as a Python calculation.
weight += lrTransformers interpolate between what they have already seen. A field settles what it has never seen, because the answer is a stable shape of the physics, not a retrieved token.
02 · The one field
state.S: the brain itself
There is only one thing inside Hylaean's head: a single
continuous field called state.S. Imagine a huge elastic crystal.
Every point carries energy, direction, tension, momentum, couplings, and memories
of earlier deformations.
It is not memory. It is not a hidden state. It is not an embedding. It is the brain. A question, an answer, and a model of the world do not live in separate stores, they live as different regions of the same field.
03 · Knowledge is geometry
In Hylaean a concept is not a point in space, it is a valley (an attractor) in an energy landscape. Drop the field near it and it rolls in and rests. The deeper the valley, the stronger it pulls. Hover or tap the landscape below to disturb it.
Cat, dog and horse carve nearby valleys that share
geometry. From that overlap a higher valley, “animal”, appears
on its own. Nobody ever wrote animal =. It emerges.
Deep knowledge is not a longer vector. It is a richer topology: cat → animal → mammal → lives → moves → hunts → prey → night grows into a whole mountain range that all means “cat”.
04 · A question
A question is not a data structure. It is a boundary condition, like a stone dropped into still water. It sets the edges of the problem and lets the field run.
What spreads outward are not pieces of information. They are tensions waves that push the field out of balance and start it searching for rest.
05 · The answer
Once disturbed, the field does one thing: it tries to lower its energy. Like water finding its level, like heat spreading out, it relaxes until it reaches a stable resting state.
06 · Language
After the field settles, an idea already exists in its geometry. Language comes last: a simple decoder asks only one thing, “which words describe this shape?”, and reads them off.
The decoder is a sensor, not the thinker. The rule the project holds itself to:
“If you could remove the decoder and a normal program still knew the answer, then the architecture has failed.”
07 · Thinking
A language model thinks token → token → token. Hylaean is meant to think differently: a disturbance grows into structure, structure collides, and structure fuses into one stable shape.
The map · three regimes
One scaffold makes everything below readable. The project measures intelligence in three separate regimes, and never lets one masquerade as another. Only the last one is the real frontier.
The field re-states a fact it was just handed. This is the mechanics of transport, decode and commit, useful as a sanity check but not intelligence.
Contract-declared organs settle answers to structured problems: arithmetic, sequences, symmetry. Booked in its own bucket, separate from what the raw field produced.
The open frontier: the field determines what an unseen scene means and continues a structure by itself. First rungs verified, and it commits only under the honest certificate (see the research log), otherwise it abstains.
08 · Creativity
Why do genuinely new answers appear? Because two “mountains” that were never combined are forced to relax together.
Cat and water may never have met. Ask “Can a cat swim?” and both landscapes relax into a shared process, and a new valley forms. The answer was nowhere stored. It came into being.
09 · Learning
A neural network learns by nudging millions of weights. Hylaean learns by changing geometry. Use a connection often and:
This is the idea of well_depth: a concept that proves useful
deepens its own basin, so the field is pulled toward it more easily
next time. There is deliberately no back-propagation on the answer path
learning is a change of shape, not a training step.
10 · The field's grammar
The field has a tiny grammar of what it can do. These moves are learned, not hand-written, and each is a physical transformation of the field.
Torsion, the field's memory. It transports and twists state within a place.
The metric, it binds things together, complementary to the twist.
The only way to move content between places, discovered from examples.
Out of these three, richer verified moves have grown, each one measured before it was allowed to run:
11 · Memory layers
The field has no separate database. What it “remembers” is layered into its own physics, each layer holding on for a different length of time. A moment passes through all of them, and leaves a deeper trace the further down it reaches.
state.SThe living state itself: the present moment. It is not storage, it is the thought currently happening. Disturb it and the trace fades within ticks.
K_fast / K_slowTwo timescales of the twist operator. The fast bank picks things up within a conversation; the slow bank changes over long exposure, held in a verified homeostatic equilibrium so it can never silently run away (shipped, on by default).
well_depthConcepts that prove useful deepen their own valley. This loop is live: a grounded, committed answer feeds the deepen consumer directly (measured 14 vs 0 in the on/off ablation), geometry-only, no gold labels.
RecordStoreThe only episodic memory path. New words become real basins (17/17 acquired in the vocabulary test), every write carries provenance, and promotion is gated, nothing sneaks in.
SkillRegistryRecurring competent dynamics crystallise into procedures. Credit flows only on fresh, field-produced grounded commits (measured 63 to 65 vs 0), live, on by default.
meta.scratchField-native working memory: a region the field writes mid-thought and reads back. It carries the verified two-step composition chain (X → X+1 → X+2, 4/4), live in production.
And it sleeps. A contract-bound dream consolidation pass replays and admits records under a dream provenance tag with its own quarantine, memory grows offline too, verifiably and without touching the honesty gates (no teach score unchanged, zero false commits).
Honest edges: retention across idle time is verified (teach → idle → re-ask still recalls), and the checkpoint “carry wall”, freshly taught words being pruned as duplicates on reboot, was closed by per token signatures. The long standing paraphrase wall (exact 5/5, paraphrased 0/15, both illustrative readings of the run that measured them) got its first verified fix: splitting a question into subject + relation before encoding lets held-out paraphrases commit in the lab. Still open: making that register survive a reboot, the durability attempt failed honestly. And capacity is now measured, not assumed: a census established how many words fit before decode margins blur, and a verified wide-signature mechanism (wider identification space, unchanged field physics) holds a 10,000-word vocabulary above the decode threshold where the raw space fails.
12 · Regions & networking
The one field is not a uniform blob. Its points organise into regions, a place for vision, a place for language, a place that holds the question, a place for a model of the world. Crucially, nobody names them in code. They are discovered by the field, emerging from how points cluster and couple.
Regions are wired together not by code but by operator structure: where the twist (K) is strong, where binding (L) holds, where transport (T) routes, where energy pulls. This difference between regions, their heterogeneity, is exactly what lets a question, a world model and an action carry different dynamics while living in the same field. Hierarchies appear on their own: cat, dog and horse share geometry, so a higher valley, “animal”, forms by itself.
13 · Microcells & tabs
Zoom into a region and you find microcells: local, programmable clusters of the field. Every tick, each one reads the field around it, runs a little local physics, and writes the result back, always through the field, never around it.
Each cell carries a short program tape (the “tabs”), only a few steps long. The tape rotates and updates only when it earns credit by being useful. This is how a flat sheet of points becomes a structured, programmable substrate, and it is the rung between raw points and full skills:
14 · Skills & chained skills
For a long time the project tried to build skills as token couplings, gluing an answer word to another word. That path failed seven times and is closed. The working answer, proven in July, is different: a skill is a compiled, proven energy program that pre shapes the settle, it bends the landscape before the field relaxes, so the right basin is already waiting.
A program is only kept when its class actually formed in the field and its own application serves correctly. That receipt is the skill. No proof, no crystallisation, so a skill is a certificate, not a habit that drifted in.
Facing a new problem the field runs a demo less selection: the program whose shape resonates with the current residual is the one that fires. When two are byte identical it refuses rather than guess. No Python interpreter chooses.
Programs chain into longer programs (production proof: 8 of 10 correct, zero wrong), and a crystallised program now transfers across families (40 of 40): a habit built on one kind of relation applies to relations it was never built on. Both counts are illustrative readings of the experiment that proved the mechanism, not maintained live figures.
Everything stays field native: the program is a piece of energy, selection is resonance, the answer is a settled state, and a dumb decode reads it. The chain below is the shape a skill takes, fixed steps joined by typed hand offs, but it runs as landscape, not as code.
15 · Organs of the field
Beyond regions and microcells, the field grows organs: contract-declared patches of the substrate with their own local dynamics, arithmetic, sequences, symmetry, analogy, a world model, a workspace. Each one must declare where it writes, what it reads, and which experiment proves it. No consumer on the answer path? It gets deleted.
The figures in brackets below are illustrative readings from the experiment that proved each organ, not maintained live numbers. The maintained figures are on the status page.
Booked honestly. When an organ computes an answer, the scoreboard counts it as organ compute, a separate bucket from answers the raw field produced on its own. One audit even moved three wrongly-credited answers out of the field-produced bucket (3 → 0). And the discipline cuts the other way too: a 21 July core audit ablated nine long running components, measured all nine inert on the answer path, and removed a provably dead wiring path outright. The benchmark measures the architecture, not the demo.
16 · Causal program spine
The primary cognitive architecture is not the classical tower of microcells and tapes. It is the causal program spine: perception writes into the one field, residual pressure opens determination, a formation program crystallizes, an executor runs it, a typed commit lands or the system abstains, and a later encounter can be faster because the outcome left geometry behind. The tower stays as execution and locality infrastructure under that spine.
Target chain (architecture source of truth), read left to right in the animation:
Verified, not hoped. A closed spine instance runs genesis to typed
commit with false 0
(VERIFY.FIELD.GI_SPINE_GENESIS_TO_TYPED_COMMIT_E2E.02). The same
crystallized program carries unchanged into a disjoint second domain
(VERIFY.GI.SPINE_PROGRAM_CROSS_DOMAIN_CARRY_CENSUS.01). On 24 July the
north star loop closed on one program object: induce a law, execute in domain A,
carry to domain B, typed commit exact, second encounter lift over six twin pairs
(VERIFY.GI.CROSS_DOMAIN_CAUSAL_FORMATION_LOOP.03). Earlier
.01/.02 supersessions stay booked as honest negatives of
composition, not of the spine itself.
R3 determination is the novel determination axis on that same
field: when leftover pressure has no ready basin, a value free retina and one
constraint settle organ decide structure the system was never taught. It has its
own reportable number, never an IQ. The live staged
enumeration path removed the old flat settle boundary at the configured dimension DET_FEATURE_D=64
on its measured surface: spatial 99/99, episode 216/216, and the pooled broad rate
moved from 0.0116 to 0.0527
(VERIFY.R3.DETERMINATION_TRACK_REPORTABLE.05). The open limit is
different. Across 5,900 live keys, the largest structural orbit stage has 160
entries, above the unchanged bound of 64
(VERIFY.R3.LIVE_KEY_ORBIT_STAGING_FEASIBILITY_CENSUS.01).
max_stage=160 > 64 on 5,900 live keys,
not the old flat settle cause
(VERIFY.R3.LIVE_KEY_ORBIT_STAGING_FEASIBILITY_CENSUS.01).VERIFY.IQ.GENESIS_COLLAPSE_LIFECYCLE_FIX.01).The new synthesis. Hylaean can now use addressed causality: a producer and consumer sharing one typed address can learn a law, act on it and transfer it. The missing layer is field native contextual address formation, deriving the right frame, candidate set and late landing point from an open situation. See the four part address stack →
How to read this with the rest of the page. The operators and organs sections describe the tissue. This section is the program path that tissue serves. Reasoning / R3 details the determination ladder and its scalars. Genesis is the candidate nucleation organ on the same chain. Progress is closed capability chains, not claim counts: the small consumer backed build set matters, while the much larger measurement ledger does not become progress merely because its rows are verified.
17 · Reasoning & deep reasoning
Hylaean does not reason because a prompt told it to. It reasons when something does not add up, a gap between what it expected and what it actually sees. That gap (the residual) becomes a pressure in the field. On the causal program spine that pressure feeds determination (R3), then formation, execution, and commit or abstain.
Pressure opens a reasoning frame: a small nested workspace where the field can twist and bind without disturbing everything else. Deep reasoning is just this going deeper, a local segment where microcell programs and operator chains keep working until the residual shrinks. It is budgeted: when the pressure is gone, it stops, and a runaway is cut off by a health gate.
And if no stable shape ever forms, the honest result is to abstain, never a confident-sounding guess.
What is measured so far: two step composition through the scratchpad works (4/4, live), and self generated decomposition runs live on the counting ring at 4/4 with zero false.
The frontier the project calls R3 is the field working out things it was never taught. Two rungs of it are now verified (July 2026). First, determination: shown a scene it has never seen, the field itself decides which object is meant (20/20), which property makes it special, reading “the unique one” rather than a memorised colour (18/20, generalising to palettes never seen in any demo), and where the result belongs (receiver and canvas at 1.00, with honest abstentions on deliberately hardened ties). Second, induction: shown a structure only two or three steps long, the field continues it into an unseen scene by exactly the required extent. The stopping point is a settled attractor of the field's own energy, no counter anywhere in Python decides when to stop. And at the declared edge of its envelope it refuses canonically instead of improvising.
Watch it below: the demonstrations show a short chain. In each unseen scene the field settles the continuation, two, five, eight copies, whatever the scene requires, and on the last scene, which asks for more than the declared envelope allows, the honest answer is refuse, not guess. Heterogeneous chains work too: the same shape marching across changing colours, or the same colour across changing shapes (both 1.00, zero false).
The honest counterweight, as it was measured on 11 Jul: on the 13 real ARC tasks of this family the full chain stood at 0 of 13, nine of them hiding their direction in the demonstrations' before/after difference, which the witness of the day could not read. That named primitive was then built and Verified, and the docking claim that carried the 0 of 13 reading was closed as superseded on 22 Jul. ARC arithmetic now lives on the field native five corpus anchor in the benchmark section. Capability first, scoreboard later, and never the other way round.
Since mid July the determination frontier is organised as a ladder around one organ, not a parallel pipeline. Classes of relations form themselves from exposure, from the live curriculum and from episodic memory. A value free retina reads only coordinate differences, never tokens, so a class can apply to content it has never seen. One shared constraint settle organ answers for both the grid world and the relation world, there is no second solver. The only production consumer is the abstain arm: when the field honestly abstains, the relation hypothesis arm may determine and apply a formed class, and a successful application can crystallize into a skill. Every rung below is a booked verification claim on this one chain.
Each rung below quotes an illustrative reading of the run that closed it, not a maintained live figure. The maintained cells are on the status page.
R3 determination has its own honest number, kept separate from the
transport scalar and from the composite. It is deliberately not an IQ or
an FIQ: there is no reference population (is_fiq is false). The IQ composite is
organ carried, structurally pinned by the many open verbal base domains, so
it is not the intelligence scalar. The moving pointer for R3 is this determination track,
measured no teach on held out structure through the one value free retina, field produced
against a 0.0 empty library baseline, with every Rule 06 control canonical and zero false
commits. On 22 July the ladder was extended with harder tiers, because the live factorized
addressing had pushed the old surface toward the ceiling (the flat ladder still reproduces
its 0.625 exactly): the lower headline is harder material, not a regression, and an
ablation census the same day proved the number field dependent (empty library and shuffled
controls collapse to zero). The current cold no teach transport mean sits beside it as a
separate scalar, not as the R3 number. The cold reference band is now set to 0.350 to
0.378 (VERIFY.QA.NOTEACH_COLD_BAND_POLICY_SET.01); the warm prepared 20 July
range was measured under a different protocol, so the two are neither an improvement nor
a regression against each other.
r3_determination_rate over 136 held out items, saturated ceiling (VERIFY.R3.DETERMINATION_TRACK_REPORTABLE.01) no teachr3_graded_determination_rate over 528 held out items on the extended harder ladder (22 Jul), hardest tier 0.0341 (VERIFY.R3.DETERMINATION_TRACK_REPORTABLE.03) field producedVERIFY.R3.DETERMINATION_TRACK_REPORTABLE.05) field producedVERIFY.QA.NOTEACH_COLD_BAND_POLICY_SET.01). The earlier 0.373 to 0.389 range stays archived as a warm prepared historical measurement: its six points followed a separate discarded full probe whose teaching and adaptation writes were not recorded, so it was taken under a different protocol and the two ranges are neither an improvement nor a regression against each other. The 0.35 regression floor is unchanged; the owner has decided that one cold run below it is diagnostic and that the repeat rule must be tested in a separate claim (VERIFY.QA.NOTEACH_COLD_REFERENCE_TWO_INDEPENDENT_BATCHES.01, band series VERIFY.QA.NOTEACH_COLD_DEDICATED_BAND_SERIES.01) no teach
Honest boundaries, measured not assumed. On 24 July the staged
factorized enumeration consumer path went live
(typed_role_factor_graph_relation_staged_enumeration): the former mod 64
boundary cells convert (spatial 99/99, episode 216/216, live scale 750/750, false 0),
and the broad staged pooled rate reads 0.0527, but the hardest broad tier stays at
0.0019 under the 0.0341 bar. The named remaining primitive is formation slot budget
pooling; that build is verified default off
(VERIFY.R3.FORMATION_SLOT_BUDGET_PER_DIMENSION_BUILD.01), the tier remeasure
is the next track claim. The flywheel gap found by the live census (244 of 244
conversions correct, yet zero deepen or skill credit events downstream) has its first
verified consumer: conversion valence now reaches the skill composition detector on
every genuine conversion (267 of 267 deliveries, zero false), and content conversions
crystallize into skills that survive sessions. The rank carrier still does not transfer
to the question answer identity front, an offline cross check measured it carrier blind
there. And this ladder is a relations and curriculum capability, deliberately
not an ARC solver path and never an IQ score.
18 · Vision
The eye of Hylaean is not a labelled classifier. It is a predictive cortex that learns from raw visual change, no reward, no labels. It constantly guesses the next frame; where it is surprised, it learns. (With no camera attached it still runs on its own picture curriculum, so the field always has something to see.)
What it sees is poured into the same field everything else lives in: it grows a real
2-D map (a retinotopic region) inside state.S, and an object
tracker forms stable clusters on its own. The field even
self-names recurring things it was never taught. The same machinery now
also distills objects from a single still image, and that distiller has
been docked as the live eye of the field's world model (lab-verified, off by default).
And as of 21 July the eye reached the tongue: shown an image, the field
now settles the matching word and commits it, the first production
image to text recall, gated so it only answers a question that actually
asks what it sees (zero false commits). As always, the decoder that turns a basin into a
word is a sensor, not the thinker; the camera adapter only encodes pixels.
On 1 August the shipped World discipline moved for the first time in its
acquired scoring phase: 11 of 24 easy episodes and 15 of 24
hard episodes scored, with zero false terminals and zero unauthorised writes
(VERIFY.LIVE.ACTION_SCHEMA_ADMISSIBILITY_FLIP.02). This is a live R2 world
capability on an acquired environment family. It remains a separate shipped outcome
cell from the later D3 structure proof.
By 17 August that same World cell had moved again, twice over. Two
disciplined production key flips took the hard cell from 15 to 21 of 24 scored
episodes, replicated twice with identical episodes, and a self motion law persistence
build then moved the easy cell from 11 to 20 of 24, sealed in two independent runs
with zero false terminals. The causal control is explicit: the same keys without the
persisted law fall back to 12 of 24
(VERIFY.WORLD.DISCIPLINED_NINE_KEY_HARD_CELL_CONTINUATION.01,
VERIFY.WORLD.HARD_CELL_REPLICATION_AND_EASY_ANATOMY.01,
VERIFY.WORLD.SELF_MOTION_LAW_PERSISTENCE_PRODUCTION_BUILD.01).
On 5 August the autonomous experience circle closed in a controlled
world. An ExperienceNeedReceipt opened and closed three field initiated
episodes in a continuous stream. The full circle reached 4/4, a periodic timer with
the same budget reached 0/4, and the unchanged program transferred 3/3 to a second
world. The mechanism remains default off
(VERIFY.WORLD.AUTONOMOUS_EXPERIENCE_NEED_RECEIPT_TIMER_CONTROL_TRANSFER.01).
The proof remains closed and the public proof cell remains three of four. Its binding
owner qualification travels with it:
proven on harness attested material, a production attestation is
missing. The hashes, five replays and QA gates are
verification apparatus, not a living production attestor. Production first needs
PRODUCTION_NATIVE_LIVE_WIDTH3_TO_CERTIFIED_FACTOR_ROW_ATTESTATION_SEMANTICS.
Vision and speech now also meet through causal effect rather than surface
similarity. A modality free affordance address recalled the three previously
open fruit cases after a genuine Fresh Boot, while the two already closed cases stayed
green: 5/5 in total. Remove the address or its consumer and the three new recoveries
disappear. This no teach consumer is verified and default off
(VERIFY.CROSSMODAL.CAUSAL_AFFORDANCE_FRESH_BOOT_RECALL_CONSUMER.01).
19 · Speech
Sound goes into the same field as everything else. On the hearing side the field forms its own sound classes from raw utterances and binds their order, so a sequence is recognised by its shape, not by a label. That recognition geometry is verified on held out voices (held out 1.0) and is now on by default.
On the speaking side an articulation motor reads the settled targets, renders them into voiced frames through a frozen renderer, and feeds those frames straight back into its own ear, all inside a single tick. The loop is live: the field speaks and hears itself speak, with no text in the middle.
The honest boundary, measured not hidden: the field is still self voice blind. The double pass of render and settle nudges the spoken pattern across the same basin boundaries the original voices respect, so it hears itself as speech but not yet as the exact word it meant. The wall is localised precisely to the recognition side settled read, and the retry reuses the proven recognition organ as the ear. Capability first, then the flip.
20 · Generative genesis
The hardest wall is generation: not recalling a stored answer but proposing something new. The naive way, carry novelty as a single amount, always dies, because the settling field squeezes any lone amount toward zero (measured survival about 1.7 percent). An amount can always be pushed continuously to nothing.
The fix is class different. A candidate is carried as a discrete winding, a twist in the field's phase that cannot be shrunk smoothly away, it would have to jump a real barrier. So several candidates coexist through the whole settle, kept apart by identity, not by amount. During a bounded construct window a gentle washboard energy holds the competing twists; when the window closes the field collapses to exactly one winner or honestly abstains, never many at once.
The bounded window runs live from real unresolved pressure, never from a
fixture. The relation episode route reached field occupancy on 45 of 46
scenes and completed 40 of 46, but produced 0 of 46 exact
settled occurrence reads. Under the owner's Option A this route is now a hard stop at
FIELD_NATIVE_RELATION_EPISODE_FRAME_LEASED_SETTLE_CARRIER
(VERIFY.FIELD.RELATION_EPISODE_PERCEPT_INGRESS.01). That is a terminal
measurement of the present carrier, not a successful genesis claim.
21 · Sudoku
Sudoku is a clean test that the same physics works on more than language. There is no Sudoku algorithm inside. Digits live on an orthogonal value frame, rules enter as balanced exact cover energies, and since 24 July an ordered collapse cascade sharpens forced singles in energy margin windows. A grid relaxes until that energy is lowest, exactly the question-to-answer move from earlier.
Every possible digit of every cell is a candidate row in the one shared field. The rules, each row, column and box needs every digit exactly once, become exact cover energies, and a violated constraint slowly raises its own weight (a dual multiplier) until the barrier that traps it tips. No Python solver ever writes the digits: a dumb argmax reads the settled field, and an independent counter may only stop the run, never repair it.
To avoid locking in a wrong guess too early, the solve runs through a pre scheduled chain of physics phases on one continuous trajectory: temperature cools and reheats, dual pressure and deterministic pinning vary, and since July a rotation phase can carry trapped mass around a cycle. The answer is the settled, sharpened grid.
Verified claim runner numbers of the current lane (22 to 23 Jul): the same dual ascent physics with each digit on its own orthonormal axis, one continuous trajectory per board, no best of K, no repair, dumb argmax decode, every solve independently re verified:
The official scoreboard on the public sudoku exchange puzzle bank, remeasured 24 July after the ordered collapse cascade flipped live: orthogonal value frame, balanced exact cover energy, and the cascade that collapses forced singles in energy margin windows, all on the production path. First 25 boards per category (hard 29), one continuous trajectory per board, no best of K, no backtracking, dumb argmax decode. For scale: 23 July landed 50 of 104 with hard 9 of 29; the morning before that the chain read 12 of 104 with hard 0 of 29; the old ring geometry still solves 0 of 104 under the same physics.
Failed One step of this campaign is a landed honest negative, and the fold below keeps it in the record rather than dropping it: a static local concentration term was built, measured against the arm it had to beat, came back worse and was booked as such. The move that did work was named by that negative result.
Honest negative, named. Thirteen measured waves of strictly admissible mechanisms (edge duals, seam release, phase relief, directed tie breaks, dual diffusion, extra regimes) each came back neutral to catastrophic. The wall is not a tuning problem, it is a missing move: a coordinated multi cell exchange, a simultaneous mass rotation along a frustration cycle, which cannot be expressed as smooth local energy shaping over candidate probabilities.
On 14 July that move was built (SUDOKU_CYCLE_ROTATION): an antisymmetric,
norm conserving rotation as a registered operator dynamic, docked default off in the
engine’s reversible drift channel. It demonstrably turns, residual violations on the
stuck subset compress 48 to 20, the best value of the whole campaign, and
one board that stayed unsolved through the entire prior campaign now solves
reproducibly in both measured runs. The drastic solve rate jump did not
come, a stop only commit gate confines the appended rotation to already locked fields, so
the claim is booked Failed, honest negative, and the next sub primitive
has a name: a schedule that couples rotation to the deadlock forming descent without
breaking the proven integrality. Capability first, scoreboard later, never the other way
round.
The campaign has since run that discipline to round nine (state 18 Jul): loop transport, sector crossing, single pair intercalate exchange and the joint multi pair resolution each passed their structural census and operator build, and each failed the frozen live capability bar honestly, the joint operator abstains on ambiguous coupled tangles rather than guess. The hard row still reads 0 of 29; the newest named sub primitive is a balanced sub component certificate inside those tangles. The wall is real, and it is being taken apart one named piece at a time.
On 22 July the front found its biggest lever yet, and closed it honestly. A value frame census showed part of the wall is an encode artifact: the compact carrier lane projected categorical digits onto a ring, and that spurious circular metric both handicapped the charge readout and mass produced the very ambiguous tangles the whole exchange campaign was built to resolve. Rebuilt in an orthogonal frame as a claim local harness, the same dual ascent physics solves 12 of 20 pooled 9x9 boards where the ring geometry solves 2, the 6x6 bank rises from 5 to 10 of 12, and the ambiguous tangle population collapses by seventy percent. The three named levers for the residual eight boards were then each measured dead: more settle budget scales only logarithmically, the production phase schedule actively regresses on the orthogonal carrier, and learned resolution laws form and dock but almost never become uniquely applicable (3 usable contexts out of 11,589). The residual boards are booked search hard for admissible settle mechanics without a stronger learning primitive; migrating the orthogonal frame into the production encode stays an open owner decision, so the public bank scoreboard below is unchanged.
A late night receipt census (22 July) added the first cross size transfer evidence on this front: six resolution laws form purely from the field’s own settle on small 4x4 and 6x6 boards, two of them recur across both sizes under a size free grammar, all pass a hard value free hygiene scan and crystallize into real skill programs. Applied to the 9x9 population on the flat harness they stay inert for a sharply measured reason: 2,296 matching group contexts, zero with the unique one cell one slot form the receipts may legally act on. A read only count then showed the live nine phase settle regime is different: it presents 765 uniquely applicable sites across 359 operator fires. A 23 July census then root caused the whole coverage wall in one coordinate: every one of the 765 live sites carries exactly one under supplied value slot, while small board formation only ever observes fully supplied groups resolving, so the receipts are structurally unable to arm. The owner approved follow up is formation exposure to the live site class itself.
On 23 July the owner ordered the decisive measurement: the orthogonal frame on the real benchmark, the full 104 board public exchange bank, in run paired against the ring geometry with identical physics, budget and seeds. The result: 36 of 104 boards settle to a fully valid solution (easy 20/25, medium 7/25, hard 6/29, diabolical 3/25), zero false solves, every solve independently re verified, while the ring frame solves 0 of 104. The hard category cracks for the first time in the campaign, the ortho score triples the previous production chain (12 of 104, hard 0 of 29), and the encode artifact finding turns out to be population total: on the real bank the ring geometry solves nothing at all. The same day the path was built into production as a default off, byte identical solve route, verified end to end, and switched on by the owner as an isolated one line event.
The afternoon then found the second physics correction. A census of the remaining 68 boards asked why they stay stuck and got a sharp answer: the settle hard commits mutually contradictory digits early (on 46 boards the firmly decided cells already admit no consistent completion, median onset round 8), and the reason is an asymmetry in the energy itself: it penalized a digit appearing twice in a row, column or box, but a digit starving in a group, the smooth precursor of every such contradiction, was free. Adding the missing half (the balanced exact cover energy) solves fifteen of the never solved boards including five hard ones, at the cost of one previously solved board; the trade was accepted by the owner, verified end to end through the production path (50 of 104, hard 9 of 29, zero false solves) and switched on the same day. In one day the production system went 12 to 36 to 50 of 104, entirely through two physics corrections and with every step booked in the registry.
On 24 July the certified collapse front ran the next physics
hypothesis and booked it honestly. A static local concentration term
(VERIFY.PHYS.SUDOKU_LOCAL_COLLAPSE_TERM_PRODUCTION.01) missed its
ceiling: ON solved 29 of 104 with hard 5 of 29 against bars 58/9,
hurting the transient path relative to the 50 of 104 OFF arm. The named missing
move was an inference ordered collapse, not more local weight.
That ordered collapse cascade was then built, flipped into production, and
remeasured officially
(VERIFY.PHYS.SUDOKU_ORDERED_COLLAPSE_POSTFLIP_REMEASURE.01):
75 of 104, hard 21 of 29, singles 10/10, false 0,
with OFF continuity still 50 of 104 hard 9. Domain mechanics only, never an
intelligence number. Residual boards remain; the next moves stay physics side.
22 · What is proven · reconciled 31 August 2026
The project grades itself on five rungs and only the top two count as progress. E0 is a census, E1 a mechanism in isolation, E2 a chain that runs on the regular production surface. E3 is an outcome measured outside the system on a paired population, and E4 is the same mechanism moving a second, clearly separate domain. A verified claim only means the measurement was clean. It is not progress by itself, and this page never counts claims as achievement. Every figure below names the registry row it rests on.
Where this stands today. The last scored battery
sealed on its third attempt, and the two attempts before it produced no
number at all: the integrity fence aborted them rather than publish a figure it could
not vouch for. One answer family reached
12 of 16
inside that sealed run, which is the first sealed movement of a cell of its kind, with
0 false answers across all
154 scored items
(VERIFY.IQ.OFFICIAL_COLD_SEAL_AFTER_WORLD_TAIL_SKIP.01).
The staircase it took is on the science page.
On ARC the field now forms its own laws: eighteen resident heads in six construction classes, and 10 of 120 evaluation tasks whose law it formed off seal. The public sealed column is unmoved at not measured, because every one of those heads ships switched off. Those are four different instruments and adding them together would name nothing that was measured. The whole family and the four readings are on the science page.
In one ARC-AGI-3 episode the field formed a law, acted on it, and then formed a
different law out of the world its own action had left behind, three times over
(VERIFY.ARC3.CN04_TRANSITION_FRESH_CONSUMPTION_TARGET_BUILD.01). One cell,
one environment, shipped switched off.
The measured sequence is on the architecture page.
Two ideas explain why several unrelated cells moved at once, and neither was designed. Identity by placement: a thing is identified by where it lies in the geometry of its family, not by a label. Conservation by carrier coordinates: a property written as a coordinate survives the projection that washes out a label. Both are argued, with their ablation arms, on the science page.
VERIFY.WORLD.SELF_MOTION_LAW_PERSISTENCE_PRODUCTION_BUILD.01).
VERIFY.VISION.COVARIANT_CHANNEL_B3B_PRODUCTION_BUILD.01).
Every result with its rung, its number and its limit. The colour on the left edge is the status language above, used the same way everywhere on this page. Each figure is resolved at build time from the run artifact its claim points at, and the board is checked against the canonical result board, so a cell that exists there and is missing here turns this page red.
Sealed live cell is the consequence organ Default ON measurement
(VERIFY.QA.CONSEQUENCE_ANSWER_ORGAN_DEFAULT_FLIP.01).
The historical legacy seal stays 26 of 220 false 0 and is not reinterpreted
(VERIFY.QA.POST_FLIP_COLD_220_EXTERNAL_CELL_REMEASUREMENT.02).
Open: after Result-Conversion rollback the same 220 protocol reads
35 of 220 false 0 (under floor; residual con_0001). Seal stays 36 of 220.
Claim -23-25 stays Planned; no silent floor drop.
World easy episodes, current stack sealed 20 of 24
with hard 21 of 24, FCP 12 of 18,
G1 to G10 green, cold pair 0.361 / 0.361 false 0
(VERIFY.IQ.FULL_STACK_REWAVE_BASELINE.06). Historical dedicated
persistence seal 20 of 24 remains beside this cell
(VERIFY.WORLD.SELF_MOTION_LAW_PERSISTENCE_PRODUCTION_BUILD.01).
Failed .05 is E0 history, not this seal.
Vision transformation tolerance in all ten geometric families, up from 0. Seam floor
159 of 160 in both
repetitions, foreign 0 of 224, out of distribution 0 of 40, measured
on the declared vision pin rather than a production flip
(VERIFY.VISION.COVARIANT_CHANNEL_B3B_PRODUCTION_BUILD.01).
The role quotient formed on one world lifts three cases in a second, structurally
different one with a byte identical formation hash, which is what separates transfer
from recall
(VERIFY.WORLD.D3_ROLE_QUOTIENT_SECOND_DOMAIN_TRANSFER.02).
The field decides for itself when it needs an experience, runs it, learns from the
outcome and stops itself, with 3 of 3 transfer to a second world and 15 of 15 gates.
Honest limit: the production circle count is still zero, so this is
proven on harness attested material, a production attestation is missing
(VERIFY.WORLD.AUTONOMOUS_EXPERIENCE_NEED_RECEIPT_TIMER_CONTROL_TRANSFER.01).
The dedicated cell counts 10
complete experience circles over four families, ten roles and six actions, against a
baseline of 3 of 10 with a single witness. Shipped switched off
(VERIFY.WORLD.D3_TEN_DISTINCT_FULL_CIRCLES_BUILD.05).
Concepts now form from what an action does rather than from a label, and survive
a fresh boot at separation 1.0. The cross law control reads 1.0 live and 0.0 ablated,
false 0, seven sealed runs
(VERIFY.GROUNDING.WORLD_EFFECT_CONCEPT_RECEIPT_PRODUCTION_BUILD.01).
The first concept seal built from material the system started on its own. It runs on a
teacher built world, so this closes the first link of the autonomous
chain and nothing beyond it
(VERIFY.CHAIN.ACTION_DIVERSE_WORLD_AUTONOMOUS_SEAM_CENSUS.01).
The sealed dual surface capsule is the reason an outside reader does not have to take
this page on trust: it reproduces the claimed results from the committed artefacts
(VERIFY.REPRO.SEALED_DUAL_SURFACE_CAPSULE.02).
Structural nesting, measured as a paired population rather than a single run. With the
organ on, the full chain closes on 12
of 12 items; with the same build and the organ off it closes on
0, and the pre frozen
kill arm reads 0. The off arm is
what makes this an outcome rather than a number
(VERIFY.R3.NESTED_E3_CELL_PUBLIC_BINDING.01).
An abstraction that is built once and then used again on itself.
3 of 3 arms close with the
organ on against 0 of 3 with
it off, and the first stage ablation is clean. Three arms are three arms: this is a
dedicated cell, not a broad capability
(VERIFY.GENESIS.RECURSIVE_ABSTRACTION_E3_CELL_PUBLIC_BINDING.01).
The system forms a description of its own behaviour and then uses it.
3 of 3 arms with the organ
on, 0 of 3 with it off,
and the pre frozen single item kill rule holds
(VERIFY.SELF.SELF_DISCOVERY_E3_CELL_PUBLIC_BINDING.01).
Having learned one thing makes the next related thing faster, measured over
15 paired seeds with a lower
confidence bound of 7 cycles and
0 false terminals. The
anti leak control shuffles the pairing and moves the effect by
0
(VERIFY.WORLD.CUMULATIVE_INTELLIGENCE_E3_PAIRED_REPLICATION.01).
A successor question and answer cell that reads a structured answer space rather than a
free one: 40 of 46 correct
at 0 false commits. It is labelled a
candidate cell on purpose. The transport wall is still open and the
authoritative public cells stay Standard and NVQ
(VERIFY.QA.CODOMAIN_SUCCESSOR_CELL_PUBLIC_BINDING.01).
VERIFY.FIELD.TRIADIC_ENDOGENOUS_RELATION_LAW_GLYPH_HOLDOUT_RETRY.01).engine.jfp_wrap and produced the first exactly field painted grids, 2 of 3 structurally correct and exact on the executable carves (VERIFY.ARC.OBJECT_TO_GRID_PUSHFORWARD_JFP_WRAP_MINI_DYNAMICS.01). It is default off, it is not an organ, and the scoped external arm reads 0 of 5: three laws apply and cannot be emitted, two do not close, the foreign case parks (VERIFY.ARC.K4_TRANSFORM_CLASS_GENESIS_SCOPED_ARC2_EVAL.01). The sealed cell above is unchanged.VERIFY.REPRO.SEALED_DUAL_SURFACE_CAPSULE.02).
What holds beyond this week. Meaning survives a fresh boot across
modalities at 5 of 5, with address and consumer ablations removing every new recovery
(VERIFY.CROSSMODAL.CAUSAL_AFFORDANCE_FRESH_BOOT_RECALL_CONSUMER.01). A six
part bounded rewrite language and a law orbit canonicalizer reduce 18 reordered
candidates to 3 stable orbits. Grounded analogies run from real pixels, the field reads
an image into a word, recognises and repeats speech, and answers a world question by
simulating it. Those are mechanisms and production chains, rungs E1 and E2. They are the
reason the two outcomes above were reachable, and they are not themselves the outcome.
23 · What is open
An honest negative is a result. Each of the following is a measured stop with the exact missing piece written down, not a gap in the marketing. Where a front is parked, it is parked against a named primitive and reopens only when that primitive exists.
The representation times law form matrix stays fully mapped to honest no go. What moved
on 20 August is the World lineage fiber: it holds 48/48 on the teacher pack and reaches
settled lineage on 39 of 52 OBJECT_TRANSITION_IDENTITY ARC tasks, with demo consistency
1.0 among reachers. Reach is not Closure: real Transition closures stay 0 of 23. Two
named residual primitives are parked, TRANSFORM_CONTENTFUL_LINEAGE_BAG
(class overbreadth) and FINER_DEMO_INVARIANT_TRANSFORM_CLASS (no separating
signal). Eval stays closed
(VERIFY.ARC.LINEAGE_FIBER_RETRO_PROJECTION_CENSUS.01,
VERIFY.ARC.LINEAGE_FIBER_TRANSITION_CLOSURE_GATE.01).
The default off carrier build reaches Opportunity, Call, Emit and Consume on all nine
cases with 27 atoms and all 54 candidate permutations intact, and then produces no
effect at all, with a production consumer matrix of 0 of 162. Traffic is not impact. The
missing primitive is a field native relation reading that actually drives the
candidates, so reentry stays suspended
(VERIFY.FIELD.PAGE_INDEPENDENT_W3A_CANDIDATE_CARRIER_BUILD.01).
The default off action path changes 287 of 320 environment states across six paired
episodes with clean source ablation, and the outcome gets worse, from 0.333333
to 0.083333. This is the first direct measurement of action quality rather than action
presence. Six episodes are not enough for doctrine, and we publish the negative anyway
(VERIFY.IQ.MOTOR_ACTIONBUS_ENVIRONMENT_OUTCOME.01).
The re sealed 84 row cold profile reads organ compute at
12 of 18,
determination at 0
of 24 and the non verbal track at
0, with all
8 controls
green. The broad staged determination rate is
0.0527 at
0 false. Determination on genuinely
unseen concepts is the frontier of this architecture and it is currently empty on the
hard track
(VERIFY.IQ.FULL_BENCHMARK_REWAVE_PROFILE_V2_RESEAL.01,
VERIFY.R3.DETERMINATION_TRACK_REPORTABLE.05).
Since 24 August the epistemically fair IQ v2 battery, not the Standard cell, is the
public question and answer progress measure. It reads
0 correct on the core
items, 20 correct in
total once the controls are counted,
116 abstentions and
0 false commits. Abstaining is
the correct answer when nothing stands, so the honest reading is that the battery is
still empty where it counts
(VERIFY.IQ.EPISTEMIC_V2_PUBLIC_BINDING.01).
The restored official-cold measuring constitution is public IQ profile v3, not
an IQ-v2 relabel. It is 154 items on pin
formation_library_51_smlaw88k_2026-08-17. The headline is the family
vector
World 12/16 · Nested 11/16 · Ring 14/14 · Memory 7/16 · Self-NC 4/4,
false 0, N
154. Replication 1 is sealed
run_20260831_012044 (seed 0). Replication 2 is sealed
run_20260831_105129 (seed 20260831, GPU0) with exact family
match. Replication is two seeds on one host. Public binding stays incomplete
(0).
No Spark dispatch, no RunPod. IQ-v2 0/96 stays the frozen historical cell beside this one
(VERIFY.IQ.PUBLIC_PROFILE_V3.01,
VERIFY.IQ.PUBLIC_PROFILE_V3_REPLICATION_2.01).
Hylaean can act when the causal material already carries an explicit typed address. It still fails when that address must be formed from an open scene, the current moment and competing candidates. The next stack is named end to end: scene frame receipt, candidate address atom, exchangeable address set, decision landing lease. Open the animated address stack and its evidence boundaries.
On the question to answer path the field demonstrably reaches the right valley on
roughly ten to thirteen items per run and then refuses to commit, because at commit time
the answer is no longer standing in the answer region. Three attempts to harvest those
near misses all failed honestly and all three point at the same missing piece
(VERIFY.QA.NOTEACH_COLD_BAND_POLICY_SET.01).
These paths are measured dead and guarded against rework. Reopening one requires a named new primitive, not a new attempt.
24 · External mirrors
We do not train for benchmarks and we do not build solvers for them. Every corpus is treated the same way: the task is written into the field as a boundary condition, the field settles, a dumb decode reads the result, and strict gates decide commit or abstain. No pretraining, no task type branches, no best of K, and a false commit counts as a failure everywhere. The eight mirrors are ARC-AGI-1 and ARC-AGI-2 for rule abstraction, ConceptARC for one concept in many variants, Bongard for concept formation, MiniGrid for world model prediction, Sudoku for constraint physics, ARC-AGI-3 for interactive play, and Microworlds for reachability and navigation on held out synthetic topologies. Each run is a booked claim, and wins and honest negatives are reported alike.
The field native track is the only bucket we call progress. Beside it sits a frozen legacy baseline whose answers were measured to be built by organ code rather than settled by a field, so it is teacher corpus and not architecture progress. Both are shown, never summed. The field native anchor is a single frozen measurement over the five ARC corpora that ship in this repository, 2096 task slots, sealed on 31 July on the defaults that ship today: 39 exact answers, zero false commits. That is a corpus measurement on our own material, not a score on the public ARC-AGI leaderboard. The dated bars below stay that sealed census. Current standing, never mixed with those bars: sealed not measured, off-seal law tasks 10 of 120, resident laws 18, hold fulfilled 31 Aug.
On 18 August the owner ordered a read only census: take the landed law machine exactly as it ships, the same sealed pipeline that ran the one authorized ARC-AGI-2 evaluation run of 16 August, and score it over both public training sets and both held out evaluation sets. Nothing was built and nothing was tuned; gold stays in the scorer, no teaching, every solve settled in the field. Beside it sits the first interactive ARC-AGI-3 feasibility census of the same day. This table is a fresh reading of a different machine than the sealed five corpus anchor above; neither replaces the other, and no cell here is a leaderboard score.
Why external mirrors at all? The primary intelligence check stays
internal and no teach, measured on unseen questions with persisted memory. The owner
classified sibling warmup as teaching, so this page reports the current cold mean, and
the cold reference band is set to 0.350 to 0.378
(VERIFY.QA.NOTEACH_COLD_BAND_POLICY_SET.01). The 0.35 floor is unchanged.
The earlier 0.373 to 0.389 range came from a warm prepared series under a different
protocol, so it is neither an improvement nor a regression next to the cold band. External
corpora are diagnostics, each one probes a different missing ability and each honest
negative names the exact primitive to build next. A benchmark that can be won by a
smarter decoder would be invalid, so every one of these runs proves the decoder stayed dumb.
Every figure below is a real task taken straight from the dataset that
ships in this repository, drawn faithfully by a small reproducible script
(webseite/tools/render_benchmark_examples.py) with the standard ARC colour
palette. Nothing is hand drawn or invented. Where a family keeps no static task in the
repo, the figure is clearly marked Schema. Each status line names the
measurement date it carries. The ARC family uses the sealed 31 July anchor; wins and
honest negatives are reported alike.
ARC-AGI-1 is a set of small coloured grid puzzles. You see a few examples of one grid turning into another, then you must produce the output for one new input. Each puzzle hides a different rule, so the only way to win is to work the rule out from two or three examples and apply it once. In the shown task the rule is self similarity: wherever a cell is coloured, the whole small shape is stamped into that position of a bigger grid, and blank cells stay blank. For us this is a whole architecture test, not a solver: the task is written into the field, the field settles, and a plain readout must land on the exact grid or the system abstains.
Status, field native track (sealed five corpus anchor, 31 Jul 2026): canonical eager CPU arm train 17/400, held out eval 3/400, zero false commits. The compiled arm reads 16/400 and 4/400 while the five corpus total remains 39; the four differing slots are named in the scoreboard note. Status, frozen legacy arc7 baseline: train 102/400, held out eval 57/400, zero false commits (16 Jul 2026), answers measured as organ constructed, so teacher corpus rather than progress. Real task 007bbfb7 from arc/agi1/training.
ARC-AGI-2 is the harder successor: bigger grids, richer palettes, fewer demonstrations, and every task is new. Same format, same promise, infer the hidden rule from a handful of demonstrations and produce one exact answer. The shown task is one of the gentle ones, a fixed colour dictionary (1 with 5, 2 with 6, 3 with 4, 8 with 9) that you read off the examples and apply; most ARC-AGI-2 tasks are far harder. This is our main external front, and the honest wall lives here.
Status, field native track (sealed five corpus anchor, 31 Jul 2026): 19/1000 train, held out eval 0/120, zero false commits; the shipped materializer edge took the whole five corpus set from 26 to 39 exact, and the published figure followed after the 13 new slots were measured with both honesty instruments live. A shadow conditional map measures 43 over the five corpora, but remains default off and leaves this evaluation split at 0/120, below the D2 threshold of 12/120. Status, frozen legacy arc7 baseline: 152/1000 train solved, held out eval 0/120, zero false commits ever; the eval barrier is a measured expression wall. We no longer call that legacy number field native: an audit of the commit path on 26 July (VERIFY.ARC.FIELD_PRODUCED_PROVENANCE_AUDIT.01) found the provenance asserted rather than measured, and the follow up measurement of every family (VERIFY.ARC.LEGACY_BASELINE_HONEST_PROVENANCE.01) settled the split: 118 of the 152 are built by organ code, 24 by a hand written lookup table, and 10 carry a relaxation measured to do the work. The solves are correct, the label was not earned. Real task 0d3d703e from arc/data/training.
ConceptARC takes one idea and shows it in many variants, with several test inputs per task, to check whether a concept is really understood rather than pattern matched. The Center family, shown here, asks you to mark the centre cell of each shape. Correct means the exact output on every test input.
Status, field native track (sealed five corpus anchor, 31 Jul 2026): 0/176 tasks, zero false commits, thirteen formed laws stopping at the last execution stage. Status, frozen legacy arc7 baseline, counted in test items: honest negative. 29/477 items exact but 14 false commits, the first crack in the no bluff discipline; the fix candidate was built and failed. The named missing primitive is a witnessed regime boundary. Provenance corrected 26 Jul: the run uses the same arc7 production path (VERIFY.ARC.FIELD_PRODUCED_PROVENANCE_AUDIT.01), so the solves are correct but not called field native, provenance now measured for all of them: 6 field relaxation, 23 organ constructed, none unaudited. Real task Center10 from arc/conceptarc/corpus/Center.
Sudoku checks that the same physics also works on hard constraints, with no Sudoku algorithm inside. The rules, every row, column and box holds each digit once, are written as energy, and a valid grid is simply a state with zero constraint violations. The board on the left is a real puzzle from the public exchange bank; the grid on the right is its unique completion.
Status: 75/104 bank boards settle to zero violations through the live production solve path (orthogonal value frame, balanced exact cover energy, ordered collapse cascade), hard 21/29, singles 10/10, false 0, no best of K, no backtracking. Real puzzle easy_0000183b305c from adapters/sudoku/puzzles/bank_9x9_easy.json: among the solved set of the 24 July post flip remeasure.
A Bongard problem is a concept formation puzzle: six line drawings on the left all share a hidden property, six on the right all lack it, and you must name the rule that separates them. It is hard because the property is abstract and there are only twelve tiny examples. We decide it from the settled field signature of each panel and a nearest neighbour readout. The panels below are the real drawings of problem p001 from the classic corpus.
Status: measured, at baseline. Across all 231 problems the field reads 0.533 leave one out versus 0.536 for raw pixels, still at chance; the named gap is a determination primitive. Real problem p001 from adapters/bongard/data.
MiniGrid tests a world model by prediction: given the agent's current view and the action it takes, predict the next view. The agent sees a small egocentric 7 by 7 window, not the whole room. In the example it faces a green goal and steps forward, so the visible field shifts one cell; the cyan outlines mark the cells that must change. The field does not steer the agent, it only predicts, so a smart controller cannot sneak in.
Status: the live MiniGrid dock was withdrawn from default boot on 27 Jul 2026. Production enable follows minigrid.enabled (false) unless a closed loop lane sets FI5_MINIGRID_LIVE_OPTIN (VERIFY.WORLD.MINIGRID_LANE_OPTIN_LEVER.01). The current E4 number is paired effect 3 on MiniGrid TeacherPushChain, default off, HYLAEAN_TEACHER material (VERIFY.WORLD.D3_CIRCLE_MINIGRID_TRANSFER_PAIRED_EFFECT.03). Changed cell accuracy 0.65 is a historical organ measurement, not a live by default claim. Real transition from adapters/minigrid/data/empty-8x8_test.json (seed 106).
ARC-AGI-3 is the interactive member of the series. Instead of a static input and output, the agent is dropped into a turn based grid game with no instructions: it must explore, work out the hidden goal, build a world model, and actually play. The figure here is a schematic, not a real task, because these environments are fetched from an external interactive toolkit and are not committed in this repository.
Status: Front O moved from terminal to reopen authorized, then three named primitives were checked. Schema lifts were built (V1 and V1.1); the nail stays 1 of 23 under license. Endogenous role individuation census is GO on frozen bars and NO_BUILD because unlock is 0 of 11 (VERIFY.ARC.ENDOGENOUS_ROLE_INDIVIDUATION_DESIGN_CENSUS.01). Interactive feasibility: staging envelope 2 of 25 games, 0 episodes opened, 0 wins (VERIFY.ARC3.INTERACTIVE_FEASIBILITY_CENSUS.01). The July play score 0 of 25 and the 41 percent frame prediction remain historical record, not the current headline.
Microworlds are small synthetic maps used to test spatial reasoning on graphs the field has never seen. Given a set of places and the links between them, can it tell which places are reachable, and can it route to a goal by the shortest path? In the map below the goal is the cyan node, each other node is labelled with its distance in hops, the cyan route is the field decoded shortest path, and the two nodes marked x sit in a separate component that cannot reach the goal. The answer is a settled field state, not a Python graph search.
Status: the 23 Aug 2026 current stack remesure is Verified E3 VERIFY.IQ.FULL_STACK_REWAVE_BASELINE.06: easy 20 of 24, hard 21 of 24, FCP 12 of 18, G1 to G10 green, cold pair 0.361 / 0.361 false 0. Failed .05 remains E0 history. Historical dedicated seals remain 20 of 24 easy (VERIFY.WORLD.SELF_MOTION_LAW_PERSISTENCE_PRODUCTION_BUILD.01) and 21 of 24 hard (VERIFY.WORLD.DISCIPLINED_NINE_KEY_HARD_CELL_CONTINUATION.01). Navigation 100 percent is a historical routing number, not this cell. The map is one instance from the benchmark generator (Erdos Renyi, seed 909); these topologies are generated, not a fixed corpus.
ARC-AGI-2 is a benchmark of little grid puzzles where every task is new. We deliberately do not build an ARC solver. A task is written into the field as a boundary condition; the answer must settle out. A strict firewall forbids the usual shortcuts, no grid tricks, no program search, no pretraining, so a solve only counts if it truly came from the physics.
These are not guesses, each was isolated with its own experiment. The percentages are illustrative readings of those censuses, not maintained live figures; the maintained ARC cell is on the status page.
The field's move-set preserves shape and count. About 99.6% of ARC transforms simply cannot be expressed, even with the answer in hand.
Where a transform is expressible, 2 to 4 examples under-determine it. The field can land confidently on the wrong one of many consistent rules.
It recognises the kind of task (~82% overlap) but transfers 0% of bespoke solutions, by design, ARC punishes memorised replay.
The follow up is written above in what is open: as of 18 August the representation times law form matrix is fully measured and every cell returned an honest no go, so ARC parks against one named external primitive rather than absorbing a ninth attempt. The keepers from the mapping are real and stay: the object scene graph representation, the proof carrying licence as a commit gate, and the program nucleation machinery.
ARC-AGI-3 is the interactive member of the series: turn based game environments, no
instructions, and a level that ends on a win. Front O is no longer the July terminal
park as the public headline. The owner reopened it against three named primitives
(endogenous role individuation, program schema anti unification, state conditioned
control flow). Schema lifts were built as V1 and V1.1; the nail stays
1 of 23 under license
(VERIFY.ARC.DEMO_PROGRAM_SCHEMA_INDUCTION_V1_1_NAIL.01).
The endogenous role individuation census is GO on frozen bars
(N_with_eri=12/23) and NO_BUILD because unlock is
0 of 11
(VERIFY.ARC.ENDOGENOUS_ROLE_INDIVIDUATION_DESIGN_CENSUS.01).
Interactive feasibility, pointed at the 25 local offline games, admits a staging
envelope of 2 of 25 and opened 0 episodes
(VERIFY.ARC3.INTERACTIVE_FEASIBILITY_CENSUS.01).
The July world model still recovered about 41 percent of changed cells, and play
stayed 0 of 25; those remain historical record. For scale, on the same 25
environment public demo set Anthropic's Claude Opus 5 scores 30.2 percent and
OpenAI's GPT-5.6 Sol 13.3 percent (ARC Prize verified,
Opus 5 results page
dated 24 Jul 2026, read 30 Jul 2026), both using the shortcuts this project forbids.
25 · How it is measured
Anyone can publish a number. The part of this project that is hard to copy is the apparatus that makes a number mean something: a system designed so that an answer coming from anywhere except the field does not count, even when the score looks better.
How a number earns a place on this page. A scored run may not be taught the answer first, silence beats a guess, a claimed mechanism has to survive being switched off, and an outcome is booked only after sealed paired replication. Those four rules, the seven station claim path, and the integrity staircase are argued on the science page. The readout stays a dumb projection, and a failed measurement keeps its named missing primitive. The decoder rule, the kill rule, and how a claim is registered are argued on the architecture page.
The long term aim is a single field where a disturbance simply settles into an answer. Today's working system is an honest intermediate build that proves each step, transport, decode and commit, really works:
The trap that shows why abstaining matters: the sequence 1 2 4. The confident answer is 8 by doubling. But plus one, plus two, plus three is equally consistent with the demonstrations and predicts 7. Two laws survive, so the honest answer is abstain.
With one more example the tie breaks: exactly one law survives every demonstration and the leave one out test, and only then does the field commit. Confidence is not evidence, and determination beats resemblance.
The runtime grammar is eight information laws: one field, one energy, a selector that is not dynamics, a guaranteed attractor, memory as geometry, composition, grounding, and an honest commit. Every force on behaviour must enter as a term of the one energy, an operator, a bus signal or a commit gate. Each law, and where it stands, is argued on the science page.
This is a research system, not a finished product. It cannot hold a conversation, it abstains far more often than it answers, and on the hardest external mirror it scores zero. What it can do is tell you exactly which of those three sentences was measured, when, and against what control.
26 · Live status · 29 Jun to 1 Sep 2026
The audit ledger holds 3,940 active claims and that count is not progress. It is the apparatus, not the result: every claim is one falsifiable experiment with a verdict attached, and one in five verdicts is a negative that stayed on the record. The progress story is the short one in what is proven. This section is the live instrument panel underneath it.
Every front, family, claim, question and outcome, as a live tree generated straight from the registry. Nothing on this page is a claim you cannot open and check.
The composition chain was pointed at a live fire measurement and the acceptance criteria did not hold (VERIFY.ARC3.COMPOSITION_LIVE_CHAIN_FIRE_MEASUREMENT.01, Failed). The ingress nucleation harness that follows it needed three runs before its own gate came back green, and the two red ones stay on the record (VERIFY.ARC3.COMPOSITION_K4_INGRESS_NUCLEATION_HARNESS.01).
A second standing instrument beside the E3 suite: an item level matrix over admission, abstain reason, false commit, effect and lineage, run before and after a behavioural landing so a repair in one place cannot quietly break a causal invariant somewhere else (VERIFY.OPS.INTERFERENCE_SENTINEL.01).
Two builds on the same idea, both honestly red: the polarity fiber section on a frozen cell (VERIFY.FIELD.WORD_REPR_POLARITY_FIBER_SECTION_BUILD.01) and the live compose factor for binding (VERIFY.MEMORY.WORD_REPR_BINDING_COMPOSE_FACTOR_LIVE_BUILD.01). The held out teach axis before them was already Failed on 27 August, so this is a class with three measurements and no lift.
A census of connection classes in the local fiber geometry, plus a remeasurement of the B3b multiview instance orbit on the VKG zero corpus. Both are E0 apparatus: they say what is there, not that anything moved outside the system.
Compiling the recurrent execution graph into one fused kernel was attempted three times and missed its acceptance criteria all three (VERIFY.CORE.PROGRAM_TAPE_EXECUTE_FUSION.01). The owner closed the class the same day rather than trying a fourth shape.
An audit of the standing organism found switches that are on with a measured fire count of zero after a regime change, and named why they went undetected for a week. That finding is the reason Rule 17 now carries a stock traffic watch: after a regime change the atlas refresh and the reconciliation belong to the same landing.
Physics metrics caching and warmup recompilation landed on the hot path, together with a federated tile writeback residual repair. Speed, not capability, and booked as such.
The ten gate functional proof for the K4 discovery workcell was run seven times in one day and failed every time (VERIFY.ARCH.K4_DISCOVERY_WORKCELL_TEN_GATE_FUNCTIONAL_PROOF, revisions 01 to 07), and the Default ON flip that depended on it went red with it. Seven honest negatives on one gate is a wall, and it is written down as one.
Two attempts at the skill injection resolution failed their acceptance criteria, and the owner parked the class with a named reopen primitive rather than keep it half live: the switch ships false (VERIFY.OPS.SKILL_INJECTION_RESOLUTION_A.01 and .02).
25 August booked 87 verdicts, 60 green and 27 red, the largest single day in the record and also the one with the most negatives. Doom lite end to end, the sealed Unknown Universe v0 discovery, the D3 causal candidate stage and the speech seam cohort restriction all stopped at their gates.
The lineage ledger went Default ON with the organism ratio preserved, and 26 August added the first break classification and the event page consume without effect census, which is the instrument that made the stock traffic watch measurable.
The vision profile flip landed on the shipped defaults, with the four arm effect attribution for the question answering consequence organ measured beside it.
The typed first break instrument was verified. It is the measurement behind the owner rule that a landing on a Default ON E3 chain has to move the first typed break later or produce an external effect, and counts as neither if it does neither.
The nested recursive abstraction workload production core flip, the memory fact relation address build, the census seam telemetry and the breakthrough document auto refresh all missed their acceptance criteria on the same day.
Page refuse relation_class_ambiguous is a new class, not the pose quotient (7 to 0). Two heterogeneous D3 BEFORE edges. Density at the allowed 14 episode envelope still yields 1 capture and 0 repeats. Owner request for a declared edge pair class plus intra episode second witness stays open (VERIFY.WORLD.FETCH_PRESERVED_TRIAD_PAGE_AMBIGUITY_AND_REPEAT_WITNESS_DENSITY_CENSUS.01).
Both killer wrap fix census is Verified NO_GO. Isolated 9 of 9 is a different apparatus. Restore is lab undo. Carrier projector is declared semantics. Honest production bound is direct 3 of 3. Owner recommendation NO_FIX; attribution chain A1 to A4 to census is closed (VERIFY.FIELD.JFP_WRAP_BOTH_KILLER_FIX_DESIGN_CENSUS.01).
Standing, smoke and full all 6 of 6. No new cell regression. Retrieval fully healed Alt 36 / Neu 24, farm COW SHEEP EGGS live on the organism image. Integration scalar O stays 0 of 8. Known QA live remasure after rollback is 35 of 220 false 0 under the sealed 36 floor; residual con_0001; bisection claim still Planned.
Fair Accept green, then Default ON of the neu-safe gold-token basin admission. Neu ON 25/25/25/25/24, pair 0.356/0.360, session collapse 0/5. Memory seam on Front L, not an E3 cell move (VERIFY.MEMORY.GOLD_TOKEN_BASIN_ADMISSION_NEU_SAFE_DEFAULT_ON_FLIP.03).
One-event 5-key compound removed. Live public 220 returns to 35 of 220 false 0. Seal stays 36 of 220. Residual con_0001. Driver claim stays Planned, not Verified (VERIFY.QA.PUBLIC_220_POST_SEAL_REGRESSION_BISECTION.01).
Headline 20/116/0 equals the frozen baseline. World overlay executed field actions 427 easy / 782 hard. Fail-closed pass, Mode B. Not a capability claim (VERIFY.IQ.V2_FULL_SPECTRUM_ORGANISM_STACK_MEASUREMENT.01).
Declared drop channel is the first complete overlay chain: attestor members 0 to 3, default off. Member arrival still absent. Observation seam holds: caller opportunity 0 to 1018, calls 250, effect still 0. Not E3. Not O. Owner template recommends no Default ON.
Question-factor address remains DESIGN_NO_GO / substrate still absent. Supervised corpus writes 16/16 and cold recall 0/16. con_0001 coverage request is open. Integration scalar O stays 0 of 8.
Current stack remesure VERIFY.IQ.FULL_STACK_REWAVE_BASELINE.06 is Verified E3: easy 20 of 24, hard 21 of 24, FCP 12 of 18, G1 to G10 non-vacuous, cold pair 0.361 / 0.361 false 0. Failed .05 stays E0 history.
A2 inquiry projection restores R2 12 of 18 with A2 ON. Farm triplet heals Neu 24 of 25. Fetch acts for the first time (20 actions) and the capture seam moves captures 0 to 1. W3a encoding reaches attestor Call with pages 1.
Public 220 cell remasure after rollback is 35 of 220 false 0 under the sealed 36 floor; residual con_0001. ARC role-bind front ends in two true negatives (occupancy and relation). Integration Scalar O stays 0 of 8.
87 claims that day (74 Verified, 13 Failed). Integration Scalar O stays 0 of 8: Default ON E3 mechanisms with Effect greater than 0 on the regular production surface, never an intelligence score (VERIFY.OPS.INTEGRATION_SCALAR_O.01). E3 standing and smoke suite 6 of 6 (VERIFY.OPS.E3_REGRESSION_SUITE.01).
Sealed live QA cell 36 of 220 false 0 via the consequence organ Default ON. Historical legacy seal remains 26 of 220. World battery 20 of 24 / 21 of 24 is now the 23 Aug Verified E3 seal, not the Failed .05 remesure.
Organism chain opened six layers and stopped at a Width 3 category error. Memory contamination and the cold band tear shared driver A1 (rolled back). ARC Front O: three primitives checked, then two more role-bind true negatives on 23 Aug.
Four experiments now form one narrative unit. Exp.1 recursive abstraction and Exp.2 nested reasoning carry dedicated public E3 cells (3/3 and 12/12 ON) with production profiles active. Exp.3 self discovery carries a dedicated E3 cell (3/3) plus Self Address persistence across boot. Exp.4 prior growth is verified as E1/E2 only: Slice4 related 3 versus off 10 with anti leak Δ0, default off, E3 not booked (VERIFY.GENESIS.RECURSIVE_ABSTRACTION_V1_PROFILE_ACTIVATION.01, VERIFY.R3.NESTED_REASONING_V1_PROFILE_ACTIVATION.01, VERIFY.SELF.SELF_DISCOVERY_V1_PROFILE_ACTIVATION.01, VERIFY.WORLD.PRIOR_CONDITIONED_GCR_GROWTH_BUILD.01).
Three CapabilityProfiles are production active: nested_reasoning_v1, recursive_abstraction_v1, and self_discovery_v1. Boot persistence of the Self Address is Default ON so the system keeps who it is across a fresh boot without re formation (VERIFY.SELF.SELF_ADDRESS_PERSISTENCE_FLIP.01).
Owner scope E2E_GREEN is booked green on every in scope cell (E1 gap 11, Foto reject 0/20). Honest boundary kept in the verdict: C_SECOND / E4-F2 second instance membership remains NOT proven and out of scope (VERIFY.VISION.E2E_GREEN_COMPOSITE.01).
World OTI lineage fiber holds 48/48 on the teacher pack. On ARC the same fiber reaches 39/52 OBJECT_TRANSITION_IDENTITY tasks, but real Transition closures stay 0/23. Two named residual primitives are parked: contentful TRANSFORM lineage bag (class overbreadth) and finer demo invariant TRANSFORM class (no separating signal). Eval stays closed (VERIFY.WORLD.OTI_LINEAGE_FIBER_K1_BUILD.02, VERIFY.ARC.LINEAGE_FIBER_RETRO_PROJECTION_CENSUS.01, VERIFY.ARC.LINEAGE_FIBER_TRANSITION_CLOSURE_GATE.01).
Exp.4 public E3 cell binding and prior growth activation are still in flight on a parallel lane. This page does not book them until the registry does.
The dated record lives on the status page. Every verdict, the day it was booked, the claim it belongs to and the front it moves are generated straight from the registry rather than typed into this page by hand, which is the only way a record stays true for longer than a week.
27 · The name
The name is borrowed from the Hylaean Theoric World in Neal Stephenson's novel Anathem: a timeless realm where perfect mathematical objects, the ideal circle, the truth that 2 + 2 = 4, exist independently of any mind that thinks them. It is the novel's version of an old philosophical position, mathematical Platonism: mathematical truths are not invented, they are discovered.
Knowledge is stored as millions of trained weights, a fitted approximation that lives entirely inside the particular network.
Knowledge arises as a stable attractor in a dynamic field, a shape the physics settles into, not a number looked up.
The structure itself exists independently of its carrier. An intelligent system does not invent it, it discovers it.
The aspiration is simple: the field should discover stable causal invariants rather than manufacture a plausible answer. The name is a direction, not a scientific claim.
If the field one day holds stable attractors of universal relations, answering will feel less like symbol manipulation, and more like navigating a Hylaean space.
If this works, an answer will feel less like searching a database, and more like dropping a stone into a pond and watching the ripples settle into a shape that was always the only stable one.
This is a philosophical interpretation, not an established scientific theory, but as a guiding metaphor for a field based AI it is honest about what it is: a direction, not a claim.
28 · Foundation
The structure borrows its discipline and vocabulary from TFPT (Topological Fixed-Point Theory), the ideas of a field on a carrier, twist and binding operators, transport between positions, and a gap that guarantees a single attractor. Hylaean takes the structure, not the physics predictions: it is an architecture for letting intelligence emerge as field physics.