Knowledge annealment

Annealment is the degree to which knowledge is bound into structure a machine can traverse and compute, and it is a property of the corpus rather than of the model. Repeated passes add verified structure to a corpus until it becomes measurable, citable and executable. K-onneal is short for knowledge annealment.

The metallurgy is the honest metaphor. Annealing is heat treatment that turns soft metal into a durable structure; annealment turns prose that a human must read and interpret into structure a machine can run. Models are open-weight and can be exchanged; the annealed corpus stays, and each pass adds further structure to it.

Capability by corpus representation

Measured directly: representations of the same corpus — the OIML R 60 family, in the reference deployment — indexed and graded on 18 witness-graded capability probes. Plain text answered two; the full annealed corpus, seventeen to eighteen. Representation quality and serving techniques multiply rather than add.

2/18
plain text
It can find topics, but it cannot cite a clause.
3/18
clause-anchored
It locates passages and cites the clause.
6/18
typed units
Tables and formulas are retrieved as objects.
13/18
machine models
Rules are carried as constraints and calculations.
computed
execution
Verdicts are computed from the rules.

Plain text answered two of eighteen probes. It can find topics, but it cannot cite a clause.
Clause-anchored documents answered three of eighteen. They locate passages and cite the clause.
Typed units answered six of eighteen. Tables and formulas are retrieved as objects.
Machine-readable models answered thirteen of eighteen. Rules are carried as constraints and calculations.
Execution returns computed verdicts rather than quotes.

Ten rungs, L0 to L9

Each rung is defined by which corpus primitives it requires, and graded by a witness — computed, never judged, at the upper rungs.

Rung	Capability	Representative probe	Witness
L0	Locate	“Where does R 60 address creep?”	clause anchor
L1	Extract	a verbatim value from prose	exact string
L2	Nomenclature	colloquial → defined term, across registers and languages	term + register anchor
L3	Geometry	a table cell by row × column, unit-aware	value + keys + unit kind
L4	Composition	joins within a document and across its parts	both anchors + relation
L5	Cross-standard	cited external standards; composition and binding	the linked id / clause
L6	Diachrony	current edition; deltas between editions; validity windows	edition pair + delta fact
L7	Perception	labels readable only from the drawing itself	pixel-only label
L8	Computation	conformance verdicts with the violation meaning	computed offline — deterministic
L9	Instance & process	process order and its consequences	order verdict — deterministic

A lane’s capability ceiling is the highest consecutive rung at ≥ 60 % witness-checked pass; its annealment index is the fraction of primitive tiers its representation materializes.

P1–P10

The taxonomy the ladder is defined on. Editorially anchored text reaches P2; typed document models reach P7; machine models carry P8–P10 — quantitative typing, computation, instance and process.

surface text

editorial anchoring

object geometry

nomenclature

relation

temporality

perception

quantitative typing

computation

P10

instance & process

Three eras of the assistant

The ladder places every answering system in one of three eras.

Era	Representation	The assistant is	Ceiling
1 · Legacy	plain text	an index: find and restate	L1–L3
2 · Now	typed units + machine models	a grounded instrument: cite, render, verify, compute	L7 · models L9
3 · Frontier	executed models, closed world	an operator: verdict, simulate, prove, personalize	beyond L9

The corpus as a closed, executable model.

The capabilities below exist because the corpus is a model with a closed world and executable semantics — no document representation can follow. All four serve the public today in the reference deployment.

Verdict as data

Pass, fail or void — the reason in the standard’s own words, and the full clause chain — produced by execution and verified by execution.

Counterfactual simulation

Constraints and lookup formulas run on hypothetical parameters. Documents state facts; models evaluate hypotheses.

Provable absence

The closed world turns “does the standard require this? — no” into a proof rather than a refusal: N nodes enumerated, zero matches.

Self-verification

The model plane is machine-checked consistent, so the assistant can verify its own model-grounded claims by execution.

The full instrument — lanes, the golden set, rigor rules and the four-corpus comparison — is documented at the reference deployment:OIML SMART AI — knowledge annealment ↗.