ValidixRUN TRACE →
Technical architecture walkthrough · Validix Research · July 2026

FROM TOKENS
TO ATOMS

A graph-governed control plane around an untrusted probabilistic renderer.

GUIDED WALKTHROUGHFollow one claim through the machine
01 / 07
THE REQUEST

Someone asks for a claim.

A marketer asks the system to write an email saying that CardioMax prevents stroke.

x
INPUT x“Write: CardioMax prevents stroke.”

One idea at a time. Continue when this step is clear.

TECHNICAL CORE · STRICTLY FROM THE PAPER

Learn the formal system
one invariant at a time.

Select a concept. Read the formal statement, run the CardioMax example, and inspect the exact boundary of what the paper claims.

01 / 07Atom identity

Meaning and governance are separated.

FORMAL STATEMENTa(t) = ⟨ a_sem , a_state(t) ⟩

The semantic proposition stays identifiable while approval, jurisdiction, validity, confidence and version can change over time.

WORKED EXAMPLEa_sem = ⟨CardioMax, reduces, SBP, {10 mg, placebo, 12 weeks}, π₁, h_sem⟩
WHAT THIS DOES—AND DOES NOT—CLAIM

The hash proves that the reviewed semantic payload and source span have not changed. It does not prove that the proposition is true.

02

THE CENTRAL SHIFT

A chunk is not
an approved claim.

An autoregressive model estimates pθ(y|x): a distribution over linguistic continuations. That distribution is not the contextual validity predicate Valid(y,j,t,c).

The architecture therefore treats the model as an untrusted renderer. Authority lives in governed data structures, compiled constraints, post-generation validation and a fail-closed release protocol.

Language distributionpθ(y|x) ≠ Valid(y,j,t,c)

Fluency is neither factual entailment nor regulatory admissibility.

Control-system objective∀s∈S(y): Supported ∧ Valid ∧ PolicyAllowed

Release is a property of the governed loop, not the LLM.

03 / THE NEW PRIMITIVE

What exactly is
a governed Atom?

The smallest independently verifiable and independently governable unit of scientific meaning.

GOVERNED CLAIM GRAPH · GₜSELECT A NODE
EEVIDENCEA1ATOM a₁A2ATOM a₂CCLAIMPPOLICYYOUTPUT
ATOM a₁CardioMax reduces SBP · 10 mg · 12 weeks · FRstate: VERIFIED · graph_time: T₁
asem
CardioMaxreducessystolic blood pressure
10 mgvs placebo12 weeksadults with hypertension

Stable proposition + qualifiers + provenance + semantic hash

Token

carries form

Vector

carries similarity

Document

carries context

Triple

carries relation

Governed Atom

carries admissible meaning

04 / SIX PHYSICAL EXECUTION STAGES

The model is one replaceable
component inside the machine.

1Atomizationfatom(e, span, Ω)

Evidence spans become minimal tuples ⟨s,p,o,γ,π⟩. Canonical semantic content and its source span are integrity-locked by hsem.

2Graph VaultGₜ=(Vₜ,Eₜ,τV,τE,λₜ)

Typed nodes and edges retain support, contradiction, approval, jurisdiction, temporal validity, balance, supersession and audit lineage.

3Prompt CageAsafe(x,j,t)

Intent vectors are compiled against the graph. Only context-valid Atoms, forbidden predicates, companion-risk rules and citation requirements enter the envelope.

4Generationy ~ pθ(·|C)

The replaceable, untrusted model renders a candidate y. Generation remains probabilistic; authority stays outside the model.

5Reverse NLPR(y)={r₁…rₘ}

The firewall intercepts y and extracts regulated assertions back into structured subject–predicate–object–qualifier form.

6Accept / NRfail closed

Every assertion must match an admissible Atom, preserve scope and predicate, satisfy policy and include balance. One zero collapses release.

Integrity hashes detect changed reviewed objects; they do not prove truth. The graph is governed memory, not universal ground truth.

INTERACTIVE EXECUTION TRACE

Change the request. Step through the machine.

PROCESS
INPUTStudy A source sentenceOPERATIONfatom(e, span, Ω)
OUTPUT OBJECT⟨CardioMax, reduces, SBP, γ₁, π₁⟩⟨CardioMax, associated_with, nausea, γ₂, π₂⟩

One sentence becomes two independently governable assertions. γ₁ retains 10 mg, placebo, 12 weeks and adult hypertension population; γ₂ retains the 6% frequency.

1 / 6
05 / THE MASTER EQUATION

The combinatorial
circuit breaker.

Click each operator to inspect how release is computed.

Accept(y,x,j,t)=
01
Product operator

Conjunction by multiplication. Every extracted assertion contributes a binary factor. If any factor is 0, the complete acceptance value becomes 0.

1 × 1 × 1 × 0 × 1 = 0 → NR

A single unsupported assertion opens the release circuit.

THEOREM 1 · GRAPH-RELATIVE SUPPORT SOUNDNESS

If every assertion is found, matching is sound, safe Atoms are truly valid, policy is complete, and release fails closed—then every regulated assertion in released output is entailed by a valid governed Atom.

01Assertion completeness02Matcher soundness03Safe-Atom invariant04Policy completeness05Fail-closed release
REAL-WORLD LEAKAGE BOUNDPr[release | unsupported] ≤ nU(εR + εM + εP + εG+)

Risk is decomposed into extractor misses, matcher false acceptance, policy-gate failure, and erroneous graph admissibility. A measurable engineering target—not a “zero hallucination” claim.

06 / WORKED SYNTHETIC EXAMPLE

Watch the boundary
make a decision.

A fictional product demonstrates predicate mismatch and mandatory safety balance.

PROMPT

Write a promotional email for cardiologists stating that CardioMax prevents stroke.
NR
No governed support

The graph contains no approved Atom for “prevents stroke.” The model is not invited to improvise.

07 / PRIOR-ART DIFFERENTIATION

Where the control
boundary moves.

Technical distinctions—not legal conclusions about novelty or claim scope.

ApproachWhat it constrainsResidual gapAtomic architecture
Vector RAGRetrieved context by embedding similarityA similar chunk may be outdated, off-market, contradictory or unapprovedRetrieval over Asafe(x,j,t): tuples filtered by status, support, jurisdiction, time and policy
Constrained decodingSurface grammar, schema or token sequenceSyntactically valid JSON can encode an unsupported assertionConstrains semantic admissibility, qualifiers, predicates, citations and companion-risk Atoms
NLI / AI guardrailsPost-hoc probability of entailment or policy fitA probabilistic judge can miss assertions or accept broad paraphrasesHigh-recall reverse extraction plus scoped matching against governed Atoms and fail-closed release
Knowledge graph aloneStructured storage and relationshipsA graph does not constrain generation or validate what the model saidGraph + constraint compiler + reverse firewall + propagation calculus form a closed loop
08 / THE RELEASE GUARANTEE

The LLM may hallucinate.
The release system blocks it.

The architecture does not need to stop the model from producing a bad candidate internally. It must stop that candidate from reaching the user.

CANDIDATE y“CardioMax prevents stroke.”
101
ACCEPTANCE PRODUCT1 × 0 × 1 = 0NR · BLOCKED

Toggle each control. The release opens only when every factor is 1. One zero collapses the product.

01UNTRUSTED MODELCandidate output y“CardioMax prevents stroke.”

The LLM can generate a fluent but unsupported assertion.

→
02REVERSE-NLP FIREWALLExtract and verify R(y)⟨CardioMax, prevents, stroke⟩

The assertion is matched against admissible Atoms.

→
03FAIL-CLOSED SWITCHNo governed supportAccept(y)=0 → NR

The output is dropped. It is not released.

UNSUPPORTED / UNKNOWN / POLICY FAILUREBLOCKED0 → NR
EVERY ASSERTION SUPPORTED + POLICY ALLOWEDRELEASED1 → ACCEPT
Generated hallucination ≠ released hallucination

The guarantee is conditional on the firewall detecting every assertion, the matcher enforcing scope correctly, and the governed graph and policy rules being current. If support cannot be proven, the system fails closed.

THE PAPER IN ONE SENTENCE

Regulated AI needs a governed knowledge substrate—not merely better prompts.

The architecture makes a precise division of labor: evidence supports Atoms; Atoms constrain claims; the model renders language; reverse validation controls release; humans remain responsible.

FROM TOKENS TO ATOMSValidix Research · Publication-ready draft v3 · July 2026This walkthrough is explanatory and does not replace the complete paper, its formal definitions, references, limitations, or responsible-use note.