Without a distinguishing relation
Three candidates, one key
Akey q
Bkey q
Ckey q
Position, ordering or an extra settle pass cannot create information the field never observed. A tie is the honest answer, and the system abstains.
01 · The thesis
One substrate, one landscape, one settle path.
The homepage now opens with the origin question: what did evolution actually discover? This page is the architecture that follows if that question is taken seriously. Almost every design decision in Hylaean follows from four commitments. They are restrictions, not features. Each one removes a place where a system could look intelligent without being intelligent.
-
01
One substrate
Everything cognitive lives in one continuous state called
state.S. The question, the candidate answer, the world model and every specialised region are slices of the same material. There is no second store, no side memory and no planner holding state of its own. -
02
One landscape
Everything that pushes the state enters one composed energy landscape called
master_energy. A new behaviour has to arrive as an energy term, an operator, a bus signal or a commit gate. Anything else would be a second physics. -
03
One settle path
One production descent, the joint fixpoint pass, moves the state each tick. Bounded candidate programs may compete, but they compete inside that same landscape and the survivor is a settled state, never a choice made by the host.
-
04
One dumb readout
Readout projects the settled state and formats it. It may report abstention. It may not repair, vote, or resolve an ambiguity. If a cleverer decoder could recover the answer without the field, the result does not count.
Why the restrictions are the point. The hardest question to ask of any cognitive system is simple to state and usually impossible to answer: did the system do the work, or did its scaffolding do it? With one substrate and one descent path there is exactly one place a result can come from, so the question becomes an experiment. Switch the mechanism off and the capability has to disappear with it. Every proof on this page is built that way, and several of them came back negative.
Abstention is physics, not politeness.
When the field cannot separate its candidates, no commit forms. That is not a safety policy bolted on afterwards, it is what an unresolved landscape does. The clearest case is also the oldest one on this page.
With a real third relation
Observed structure gives each a role
Akey q₁
Bkey q₂
Ckey q₃
Now the candidates differ because the episode contains a third relation. The distinction comes from what was observed, not from a label supplied by the host.
These are architectural commitments whose implementation and effects require separate checks.
03 · The one loop
Intelligence is the whole circulation.
No station is a hidden second brain. Select a station to see what it contributes. The moving pulse shows the intended circulation, not a claim that every return edge is already closed in production.
Perception becomes a disturbance
Vision, world state, text and audio enter through declared sensory surfaces. These surfaces stimulate the field. They do not reason, remember or select an answer.
04 · The five layers
One system, with strict boundaries between roles.
The loop above sketches the intended flow. The five layers show responsibilities and constraints; the diagram does not establish that each capability works.
05
Verification
May measure, control, ablate, refuse and publish a verdict either way.
May not touch the answer path, or count a clean measurement as progress.
04
Communication
May project a settled state, format it, and report abstention.
May not repair an answer, vote between answers, or hold a second reasoning path.
03
Core cognition
May hold operators, energy terms, organs, memory admission, leases and bounded candidate competition.
May not run a second optimiser, a hidden relax loop or an answer repair pass.
02
Perception and material
May encode an observation or a constraint into declared sensory slices. Sources: vision, worlds and games, text, audio.
May not reason, remember, route by task type or select an answer. Each source must prove its own path from observation to settled field material, and they are not equally deep.
01
Substrate and dynamics
Is state.S with the K, L and T operators inside one composed energy landscape, moved by the joint fixpoint pass.
Guarantees that every cognitive write is owned, recorded and auditable. An unauthorised write aborts the run.
Each empirical test must identify the capability or boundary it examines and the controls used.
Six research questions, one field.
These roles describe the intended organism. They are research questions, not capability ratings.
Vision
How can observations become useful structure in the field?
World
How can the system connect actions to changes in its surroundings?
Grounding
How can field structure acquire meaning through experience?
Memory
Which changes persist and influence a later outcome?
Q&A
How can an input lead to a field-supported answer or abstention?
Transfer
When does a learned mechanism help in a different situation?
07 · How a claim earns a place on this page
Every claim must be able to fail.
The registry is an audit trail, not a score. A large number of verified claims means the measurements were clean, not that the system improved. This section is the part of the architecture that decides what the rest of the page is allowed to say.
-
01
Register a claim
State the mechanism, the expected movement and the condition that would refute it.
-
02
Build the harness
Bind the claim to the real field path, its controls, its source material and its consumer.
-
03
Measure
Run positive and negative controls. Shuffle, impossible and ablation arms expose false wins.
-
04
Book the verdict
Verified and Failed are both final evidence. A Failed claim names the missing primitive.
-
05
Mirror it publicly
Publish the evidence and its limitations together. Public evaluation is coming soon.
What counts as progress
The research framework distinguishes local mechanisms, production integration, external outcomes and transfer.
- E0EvidenceA census, a measurement, a diagnosis.
- E1MechanismA local build with a causal ablation.
- E2Production chainOpportunity through to effect on a regular production surface.
- E3External outcomeA paired improvement of an external result cell.
- E4TransferThe same mechanism improves a second domain.
The rules that protect the difference
These are the standing rules that separate an attractive demonstration from evidence that the field itself did the work.
No teach reporting
A result is never reported as intelligence if the answer was shown immediately before scoring. Taught recall is mechanics. A warmup round that teaches sibling pairs is teaching too, and is labelled as such.
Report incorrect answers
Measure incorrect commits and abstentions alongside successful outputs, using controls appropriate to the task.
No smart decoder
The readout may project and format a settled state. A benchmark that a cleverer decoder could win is an invalid benchmark, whatever the score says.
The kill rule
If a mechanism reduces to an adapter side proposer, a host side enumerator or a smart decoder, it is rejected immediately and regardless of its numbers. Guard erosion is not traded against results.
Sieves before building
A build has to pass screens before it exists: can the proposed term move the field at all, does its material survive a frozen replay through the unchanged consumer, does the regular seam actually drive that consumer. Each screen is necessary and each one is explicitly capable of a false positive, so a green screen is a licence to try, never a promise of effect.
Predict before you run, and replicate sealed
Experiments should record their prediction before the outcome is known. Replication must preserve the relevant code, configuration, inputs and controls.
How the work is organised
Many experiments move at once without their evidence or verdicts overwriting one another.
Parallel lanes, isolated workspaces
Separate lanes investigate distinct claims and material families at the same time, each with a narrow question and a named acceptance set fixed before it runs, each editing and measuring in its own workspace.
Serial landing
Finished work enters the shared history through one guarded queue that checks that registry rows, sources and generated public mirrors survive the landing together.
Ownership and envelopes
A session ledger shows who owns each front and claim, so a verified result is consumed instead of repeated. Each verdict travels with its configuration, inputs, controls, metrics and file digests, which is what connects a public sentence to the evidence that earned it.
Continue the story
From architecture to evidence, then open fronts.
Explore the scientific questions and measurement principles. Public demos and evaluation are coming soon.