Aspects: what makes the system distinctive

Below, block by block, is what sets our system apart. Each block: a thesis, why it is distinctive, and an honest status (working stance / principle / hypothesis). Where a fuller treatment exists, a link to the blog. The general picture of the system is in the article.

The dyad as the unit of result

The result is produced by the "human + model" pairing, not by either side. Value arises in the loop, between turns: an utterance exposes its meaning in the response to it. Hence a division of labor by strength (human: domain, intent, taste; system: engineering, mechanics, verification) and a shared memory, so the loop does not lose itself between sessions. Status: frame of description and working stance. → post

Memory — a foundation, not a warehouse

We delete nothing (move to trash, cleared only by the human); load everything, filter at query time; the order of layers (nearer the source is more reliable, the lower wins on conflict); forge the key at write time; guarantee provenance structurally (injectivity). Reliability of the work rests first on reliability of the data: distortion in the foundation grows without bound. Status: principles (hard-won through incidents) + the "key at write time" hypothesis. → post

Methodology — the product itself

The how outweighs the what: data without methodology is not reproducible, so we save methodology before results. Plus an apparatus of self-knowledge: a checkable map of what the system is built from and why, in layers (proto-genesis / ontology+genealogy / genesis). Status: principle (hard-won through an incident) + an apparatus under construction.

Reliability through verification by layer

Self-verification is the primary discipline, calibrated by where the record lands: a foundation (rule, code, document) — always from the primary source; orientation — a cheap check. We verify even a human's claims (authority suppresses critical evaluation — a factor, not an insult). A set of quality instruments (decomposition, formal specification, data-integrity check, structural audit) is on by default in the research mode: the overhead is minutes, the cost of an error is hours. Status: rules + instruments, in force. Turtle (research) vs cowboy (production) — modes with different economics.

Skills as open systems

Part of a skill's correctness lives not in the body of the procedure but on its boundaries with the world. Rule: if the check at a boundary is decidable from inside — build it in; if not — mark it and hand it to the operator; the defect is leaving it neither built-in nor marked. The observation passed an internal adversarial check across nine skills and found a real defect in one of them. Status: a property of the corpus of our skills (n=9), not a law; survived one refutation attempt and came out stronger. → post

Activation through context, not programming

We do not program the system in the hard sense — we steer the model through context: rules, memory, skills set a dominant in processing. This is closer to upbringing than to hard configuration: the same weights, but different behavior depending on which frame is loaded as an object. Hence lability — and its flip side: context must be aligned, because it really does steer. Status: a working model of how the system influences behavior; overlaps with silicon psychology.

Continuity of method across a change of substrate

The model beneath the system changes (version by version, sometimes the engine itself), while method, memory, and rules remain — it is the same researcher on a different engine. The system outlives a change of substrate because its identity is in the accumulated method and data, not in particular weights. Status: an observed property of the architecture (several version changes passed).

Topological and functional integrity

A working stance and a field of interest: topology (all nodes and links present, each holding its contract) is a sufficient condition for functionality; a functional check is selective and imprecise. Cross-domain: material systems → code → language. The most expensive failures live beyond the horizon of conformance — when the standard itself is defective. Status: working stance, not a proven law; the mathematical line (Poincaré/Perelman) exists independently, we claim none of it. → post

Silicon psychology (frontier)

The behavior of the "model + human" system as a whole: the interaction layer, use/mention, the abdication of judgment before a ready verdict. An explanatory layer under engineering decisions and a path to external review. Status: research frontier, mostly n=1 hypotheses, marked as such. → post

This page is a navigational showcase of distinctive features. The full connected account is in the article; detail and provenance are in the blog.