How Aby Thinks

Not a model.
A reasoning
architecture.

Aby is seven coordinated reasoning facets. Each one does a specific job. Together they produce intelligence that earns its knowledge, catches its own mistakes, and grows smarter from real experience.

What happens when you give HAL a command
Command
Natural language parsed by Claude
Phronesis
Ranks task by urgency and value
Hexis
Finds capability and confidence
Aby's Dream
Verifies physics grounding chain
Pseudos
Checks confidence is earned
HAL Acts
Verified, grounded, honest
Organum
Learns from the outcome
Leviathan
Hunts adjacent gaps
Epagoge
Promotes rules when evidence warrants
The Seven Facets
01
Hexis
The Knowledge Map

Hexis is HAL's structured capability ontology. Every skill HAL has is a node in this graph, connected to the physics principles it depends on, annotated with a real confidence score computed from actual evidence. When you ask HAL to do something, Hexis is what tells every other facet whether that capability exists, how confident Aby is about it, and what physical constraints govern it.

Live Hexis lookup · Catch
capability: Catch
confidence: 0.82
grounded_in: compliant_contact_mechanics
iso_limit: 150N (ISO/TS 15066)
verified_at: 0.75m, 1.0m drop heights
02
Pseudos
The Honesty Checker

Pseudos is Aby's self-doubt mechanism. Before any action, it scans the proposed capability claim and asks: was this confidence score actually earned through real verification, or was it assumed? If it finds a gap between what Hexis claims and what the evidence actually supports, it flags it immediately. This is the facet that stops HAL from being confidently wrong — the thing nobody else has built.

Pseudos scan result
claim: Catch at 1.0m, confidence 0.82
evidence_count: 6 real test cases
laundering_found: false
verdict: PROCEED — confidence is earned
03
Organum
The Farmer

Organum is the most important facet for the investor conversation. It is the only facet that generates genuinely new knowledge. After every real interaction, Organum forms a hypothesis about what was learned, writes a verification test, runs it against the physics simulation, and — if it passes — promotes a new grounded belief into Hexis with a real confidence score and a real citation. This is how Aby grows smarter from experience, not from retraining.

Organum cycle · post catch-at-2.3m/s
hypothesis: compliance model holds at 2.3m/s
test: run spring-damper sim at 2.3m/s
result: 139.4N — within ISO limit ✓
promoted: confidence 0.82 → 0.84
citation: hal-core simulation, 11 Aug 2026
04
Leviathan
The Gap Hunter

After every action, Leviathan compares what HAL just verified against everything else in Hexis. It asks: does this new knowledge tell us anything about capabilities we haven't verified yet? It finds structural similarities between known and unknown capabilities and proposes honest, discounted fills — but always flags them as unverified, always queues them for Organum, never auto-promotes.

Leviathan gap discovery
known: Catch verified at 2.3m/s
similarity_to: Grasp (Jaccard 0.29)
proposed_fill: Grasp at velocity — possible
proposed_confidence: 0.19 — heavily discounted
status: pending_organum_verification
05
Aby's Dream
The Grounding Checker

Before HAL acts on any capability, Aby's Dream traces it back through the knowledge graph to real, verified physics. Every capability must trace to real grounded science — not just to another assumed capability. If the chain breaks anywhere, the action is blocked. No ungrounded claims ever reach HAL's execution layer.

Aby's Dream grounding check
PickAndPlaceReach ✓Classical Mechanics ✓
ReachMaterials Science ✓
verdict: GROUNDED — action permitted
06
Epagoge
The Rule Maker

As Organum accumulates verified evidence across more and more interactions, Epagoge decides when that evidence is sufficient to promote a general rule. It uses Laplace's rule of succession — a principled statistical threshold. One real counterexample blocks admission entirely. Rules are never assumed; they are earned interaction by interaction.

Epagoge pattern admission
pattern: HAL can catch safely below 1.0m
instances_tested: 6 capabilities evaluated
laplace_confidence: 0.88
counterexamples: 0
admitted: true — standing rule promoted
07
Paradeigma
The Analogy Engine

When HAL encounters a problem it has never solved before, Paradeigma searches for a structurally similar problem it has already solved. If the deep structure matches — same physics, same constraint type, same motion class — it proposes a transfer for human review. Never auto-applied. Always a proposal. This is how Aby reaches beyond what it has explicitly been taught.

Paradeigma transfer proposal
new_problem: Catch at 90° approach angle
matched_to: Drone 108 km/h crossing intercept
similarity: 0.75 — structural match
proposal: Apply PN guidance geometry to arm
status: pending_human_review
The Platform

Aby is a SaaS platform.
Not a robot.

HAL's intelligence layer is hardware-agnostic by design. The same Aby reasoning architecture that runs on our simulation environment will run on any physical platform. A hardware partner brings the body. Aby brings the mind that makes it trustworthy.

Any Hardware
Aby connects to any robot platform via standard APIs. The intelligence layer is fully decoupled from the actuators.
Any Domain
Manufacturing, healthcare, defence, logistics. The reasoning architecture is domain-independent. The knowledge base adapts.
Verified by Design
Every deployment carries the full Porphyry proof framework. Clients can verify what the system has learned and when.