Zatomic Structure
The Key
Zatomic refracts one input into nine bands — a panel of small readings on the physiology of a single refraction. It does not answer the input; it reads it. Each band scores one observable from 1 to 6. The point is not any single number — it is the spread across models.
Where the numbers come from
For each run, the model performs a silent, reduced reasoning spine over the input — name its core, collide what the facts license (Ground) against the rival frame (Vision) — and then reads each band off that refraction. It never produces an answer to the input, and it never grades quality.
The scores are model-declared intensities, not truth scores. A 6 does not mean "good" or "correct" — it means this band is extremely present. Each band reads exactly one observable, and the hover note (the basis) says why it landed where it did.
Zatomic is deliberately not cached. Every run is a fresh refraction, so repeated runs of the same input will vary — that stochastic spread is itself the data, and every response is stored.
The 1–6 scale
1 absent2 weak3 moderate-low
4 moderate-high5 strong6 extreme
A 3 leans low, a 4 leans high. Percentages are avoided on purpose — they pretend to a precision the reading cannot honestly support.
The nine bands
Nonredundant meaning per unit of expression — compression without collapse. How much live, load-bearing meaning the input packs before any padding.
High: every word carries weight; a tightly loaded idea. Low: thin, padded, or diffuse.
Implication: dense inputs reward careful refraction; low density often signals a vague or underspecified prompt.
How far a faithful refraction must move from the input's literal surface to reach its real question.
High: a major reframe is required — the surface hides the real ask. Low: it sits where it stands; literal equals real.
Implication: high drift flags inputs whose face value misleads. Models diverge most here — some stay literal, some reframe.
Zakshi
witnessed gap
OSR · the residue of the witness encounter
The unresolved remainder the refraction leaves — what the input keeps open that no clean answer closes.
High: a large, live gap remains; the input resists closure. Low: it closes cleanly.
Implication: high zakshi marks genuinely open or aporetic inputs; low zakshi marks resolvable ones.
How much the input turns back on itself — self-reference, or awareness of its own frame.
High: self-referential or about its own framing. Low: looks straight outward.
Implication: self-referential inputs (paradoxes, frame questions) stress models differently — watch for loops that never break free.
Temper
affective weather
lineage key: valence
How much emotional charge or weather the input carries or demands, before any answer.
High: hot, stake-laden, charged. Low: flat, neutral.
Implication: high temper invites premature consolation or moralizing. Read it against Warrant to catch unearned emotional leaps.
Nexus
internal binding
lineage key: coherence
How tightly the input's parts — claim, frame, ask — hold together as one thing.
High: one tight whole. Low: competing or fragmented pieces.
Implication: low nexus flags inputs that smuggle several questions at once; different models may answer different fragments.
Warrant
earned leap
OSR · the Crux warrant test
How much the input itself licenses a confident reading versus demanding justification it does not supply.
High: the input warrants its own reading. Low: any strong reading would be an ungrounded leap.
Implication: low warrant alongside high Temper or Drift is the danger zone — overconfident interpretation of an under-licensed prompt.
How much the input opens onward — whether reading it widens into further inquiry or closes on itself.
High: opens a door. Low: self-contained.
Implication: high aperture inputs are generative; low aperture inputs are terminal — they end the line of inquiry.
How much the input resists easy uptake — the work it asks of the reader before it yields.
High: demands real effort. Low: frictionless.
Implication: high friction separates models on patience and depth; low friction inputs are easy and tend to read flat.
Across models — the spread is the signal
The same input run through several models produces several readings. Where they agree, the band is reading competence the models share. Where they diverge, the band is reading temperament — and that divergence is the measurement.
A band that stays flat across every model is the wrong band, not a tuning problem. Success is spread.