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.

FACET 01
Organum
Investigate

When Aby's confidence scanner flags a claim it cannot verify, Organum queues an investigation — non-blocking, so reasoning never stalls waiting for answers. Each investigation calls the reasoning engine for a real fix, which lands as pending_human_review.

FACET 02
Leviathan
Fill gaps

Real Jaccard similarity fill-by-analogy. Detects missing rows in the knowledge base and fills them by structural analogy with the most similar known case. Every suggestion is scored and queued for review — never applied directly.

FACET 03
Phronesis
Prioritise

Ranks the live backlog across all investigation queues using Expected Information Gain × value/cost. The highest-value investigation always runs first. The backlog is never arbitrary.

FACET 04
Aby's Dream
Ground

BFS over the knowledge graph, checking whether every composite capability traces back to real physics. Capabilities that float — not grounded in physical law — are flagged. The grounding ratio is always honest.

FACET 05
Epagoge
Induct

Laplace's-rule pattern admission over accumulated evidence. When enough instances of a pattern accumulate, Epagoge admits it as a structural regularity. First pattern found: all capabilities falling back to SafeKneel trace to Classical Mechanics.

FACET 06
Paradeigma
Transfer

Checks whether a new unresolved problem structurally matches a real already-solved case. When it does, the fix from the solved case is proposed — with similarity score and human review required before any application.

FACET 07
The Refutation Engine
Verify

The newest and most important facet. Two deterministic gates — no language model in any verdict. Gate 1: type admissibility (64 transforms, 11 classes, 14 bridges). Gate 2: Shannon surprisal against measured population frequencies. Every rejection retained permanently.

Zetesis — nightly autonomous search at Continuous
ROUND 1–3

Propose

Type-first generation. Same-type battery-family pairs explored first. Budget split evenly across 3 rounds.

GATE 1

Type check

64 transforms across 11 object classes. 14 theorem-bridges declared. Cross-type claims blocked as malformed.

GATE 2

Surprisal

Properties measured by running transforms on real inputs. Agreement on rare properties counts. Agreement on common ones counts for almost nothing.

VERDICT

Admit or reject

≥4 bits: CONFIRMED. ≥2 bits: WEAK candidate. Below: REFUTED. Sent to the graveyard.

CHAIN

Hash & store

Every verdict hashed and chained to the previous. Trigger fires on INSERT. Permanent and tamper-evident.

Facets live7 of 7
Zetesis roundsContinuous · 24/7
Graveyard30,232 verdicts
LLM verdicts0
Probe families7 measured
Facets live7 of 7
Zetesis roundsContinuous · 24/7
Graveyard30,232 verdicts
LLM verdicts0
Probe families7 measured