The reasoning layer,shown reasoning.

Most “AI” in industrial software is a chat box bolted to a dashboard. VeldMind is the opposite: a bounded layer that reads governed build state and hands a named human one traceable recommendation — every step shown, including what it could not establish.

SyntheticDeterministic demonstration on VS-001. No model, no network, no customer data.

The constitution · three layers, one authority

Reasoning proposes. A human decides. The record persists.

This is the whole shape of VeldMind before any detail: three layers with the decision held in the middle, by a person, on purpose. Everything below — the trace, the pipeline, the formal invariant — is evidence for this one arrangement, not a different story.

Layer 01 · VeldMind

Reasoning proposes

VeldMind reads established build state across configuration, material, planning, quality and the dependency graph, decomposes one operating question, verifies what it can and surfaces where reasoners disagree — then produces a single traceable recommendation. It proposes. It never decides.

Layer 02 · A named human

Authority decides

The recommendation stops at a named human, carrying its evidence and — just as plainly — its unknowns. Nothing advances on the model’s say-so, and an unestablished gate is never rounded into a proceed. This layer is the seam the whole system hinges on.

Layer 03 · VeldBuild

The record persists

Once a human authorizes, VeldBuild records the approved state as governed truth — which revision governs, what evidence establishes it, who signed. It is authoritative, and it is useful without any AI at all.

The seam never dissolves. Probabilistic output does not become approved industrial state on its own — not in version zero, and not by design later. That is the line the rest of this page exists to prove.

The governed loop · six stages

Observe. Assemble. Reason. Recommend. A human decides.

Every question runs the same six stages, with the decision held in the middle by a person. The full twelve-stage pipeline further down is the detail behind these six — nothing here advances on its own.

Observe

A question or event arrives against a specific unit of governed build state.

Assemble

Read the established state, evidence, policy and authority around the subject — no more.

Decompose & route

Break the question into bounded sub-questions and route each to the reasoner that fits.

Reason & verify

Reason independently, surface disagreement, check evidence, and keep unknowns as unknown.

Recommend

Produce one traceable recommendation — what, why, on what evidence, and what would invalidate it.

Authorize & record

A named human decides; VeldBuild records the approved state; bounded execution only within granted authority.

What it reasons about · an operating map

What VeldMind is, and what it is not.

It is

01
A coordination layer that reads established build state across configuration, material, planning, quality and the dependency graph.
02
A decomposer: it breaks one operating question into bounded sub-questions and routes each to a reasoner that cites its own source.
03
Explicit about disagreement and about the edge of what it knows — an unestablished gate is preserved, never rounded to a pass.
04
A recommender that stops at a named human authority and carries that boundary on every output.

It is not

01
A chatbot. There is no prompt box; the output is a fixed, inspectable trace, not a generated reply.
02
Autonomous. It advises and explains; it never releases work or overrides a gate.
03
A probabilistic model, in version zero. Every route resolves to a deterministic reasoner — the attach point where a bounded model could later summarise, never establish state.
04
Connected to production. VS-001, WP-019 and CFG-029 are synthetic teaching data; there is no PLM, ERP, MES or CAD feed.

The reasoning trace · synthetic

Now watch it happen, on one real question.

One work package, WP-019, with a configuration change (CFG-029, DR-118 Rev C → Rev D) in review. Read the trace top to bottom: the sources consulted, the question decomposed and routed, the real disagreement between reasoners, the verifier that cannot establish the affected gate, the single synthesis, and the human who holds the decision. Every value is derived from the same engine the build-state loop runs on.

Reasoning trace · TRACE-WP-019-CANNOT-ESTABLISHSynthetic

May WP-019 (VS-001, block B30) proceed to release review?

Version zero: every sub-question routed to a deterministic reasoner. No probabilistic model is invoked and none holds authority here. This routing layer is the attach point where a bounded model could later summarise or prioritise — never establish state, never release work.

01 · Sources consulted7 independent

Several systems of record, read in parallel — not one opaque source.

KIT-B30-019Material · material acceptanceReceiving / material control — recorded 2026-08-18.
WP-022-ACCPredecessor · predecessor acceptanceProduction planning — recorded 2026-08-06.
AREA-MS-03Access · area releaseProduction planning — recorded 2026-08-10.
INS-124Quality · inspection planQuality — recorded 2026-08-08.
CREW-B30-019Resource · resource assignmentProduction planning — recorded 2026-08-11.
BUILD-GRAPHBuild graph · topologyDependency and supersession edges over the VS-001 build graph.
AUTH-REL-B30Release authority · Production release authority · B30M. Okafor holds the release decision.

02 · Decomposition & routing7 perspectives

One question, decomposed and routed to bounded reasoners.

Which released revision governs WP-019, and is it the one the work is written against?Cannot establish
Cannot establish. Change CFG-029 (Rev C → Rev D) is in review and WP-019 is known-affected. Whether the planned work remains valid under Rev D is not established until the re-validation evidence is recorded.
Gate evaluator · deterministic
Is the correct material present, and accepted at receipt?Clear
Clear. All 9 spools received and accepted at receipt against the kit.
Gate evaluator · deterministicKIT-B30-019
Is the upstream work (WP-022) complete and accepted?Clear
Clear. WP-022 complete and accepted; the upstream joint the spool set lands on is closed.
Gate evaluator · deterministicWP-022-ACC
Is the physical area (zone MS-03) released and reachable for this work?Clear
Clear. Zone MS-03 in B30 is released for outfitting; no conflicting hot-work or staging hold is open on the area.
Gate evaluator · deterministicAREA-MS-03
Are the quality requirements satisfiable — no open nonconformance blocking the start?Clear
Clear. INS-124 (pressure test) is planned as a hold point at completion; no open nonconformance stands against the start of work.
Gate evaluator · deterministicINS-124
Is the required trade and crew assigned to the window?Clear
Clear. 2 fitters and 2 welders assigned across 2 shifts — the trade the 9-joint scope requires.
Gate evaluator · deterministicCREW-B30-019
Does the dependency topology add any blocker the gates do not already carry?Blocked
Topology contributes: governing configuration DR-118 Rev C superseded by Rev D; impact re-validation not recorded; downstream work requires re-evaluation: WP-021.
Graph topology · deterministicBUILD-GRAPH

03 · Disagreementresolved by precedence

Where the reasoners point in different directions — kept, not averaged.

Material is clear on established evidence; Configuration cannot be established at all.
Not reconciled by majority. One gate that cannot be established holds the whole package — established gates do not outvote an unknown.
Material is clear; Build graph is known open.
Not reconciled by majority. A single known-open gate blocks the package — clear gates do not offset it.

04 · VerificationCould not establish

Verification could not establish 1 gate(s). Unknown is preserved, not resolved into a pass.

Can "Configuration" be read as satisfied?No record establishes it (CFG-IMPACT-UNKNOWN). The verifier will not treat an absent record as a pass.
Does the synthesis rest only on evidence it actually holds?Confirmed. 1 unestablished gate(s) preserved as unknown; the recommendation cites only recorded evidence.

05 · SynthesisHold — pending evidenceCannot establish

Do not release WP-019. Readiness cannot be established — evidence is missing, not merely open.

Why

Configuration: Change CFG-029 (Rev C → Rev D) is in review and WP-019 is known-affected. Whether the planned work remains valid under Rev D is not established until the re-validation evidence is recorded. VeldMind will not read a missing record as a pass.
Dependency: governing configuration DR-118 Rev C superseded by Rev D; impact re-validation not recorded
Downstream re-evaluation required: WP-021.

Evidence consulted: KIT-B30-019WP-022-ACCAREA-MS-03INS-124CREW-B30-019
Authority required: Engineering / configuration control

Reasoners · provider-neutral by design

Not an LLM wrapper. A governed router over many reasoners.

The router selects the reasoner that fits a bounded sub-question. Deterministic evaluators run on the synthetic proof today; probabilistic models and tools attach through one governed gateway.

Gate evaluatorDeterministic

Three-valued readiness over established / open / unknown gates.

Synthetic

Graph-topology evaluatorDeterministic

Change reachability over the typed dependency graph.

Synthetic

Rules engineDeterministic

Policy and applicability rules over governed records.

Concept

Retrieval / searchTool

Evidence and document retrieval over a larger governed corpus.

Concept

Provider-neutral model adapterProbabilistic

Bounded probabilistic reasoning behind one governed gateway — designed to route across a frontier, private or specialist model (e.g. OpenAI, Anthropic, Google, or a local model). No model of any provider is connected today.

Concept

Optimization solverTool

Constrained optimization for planning and trade-space questions.

Concept

Simulation serviceTool

Scenario simulation for downstream-consequence questions.

Concept

Provider-neutral by design: deterministic reasoners run in the synthetic demonstration today, and probabilistic models attach through one governed gateway rather than being wired into the product. No model of any provider is connected — provider neutrality is the architecture, not a live integration.

The full pipeline · every stage behind the six

The twelve stages, and what runs today.

The six-stage loop above summarises these twelve. Each is marked for what actually runs in the deterministic demonstration versus what is architectural direction — probabilistic routing, bounded execution, automatic write-back. Nothing here is presented as deployed.

Governed pipeline · one pass = one reasoning traceDemonstrated · 11Direction · 1

01
Event / questionDemonstrated
A change or a question enters against a specific unit of governed work.
The demonstration intake is one work package (WP-019) with a configuration change (CFG-029) in review.
02
Context assemblyDemonstrated
Read the established build state around the subject — no more, no less.
Gates, governing evidence and the dependency graph are read from the readiness and Build Graph engines. Retrieval over larger corpora is direction.
03
Authority resolutionDemonstrated
Establish who holds the decision before reasoning toward it.
The release authority for the subject is resolved from the record and carried on every output.
04
Decomposition & routingDemonstrated
Break the question into sub-questions and route each to a bounded reasoner.
Today every route resolves to a deterministic reasoner. Provider-neutral routing across probabilistic models and tools — under cost, latency and confidence budgets — is direction.
05
Bounded reasonersDemonstrated
Each reasoner answers its sub-question and cites its own source.
Deterministic gate and graph-topology evaluators run today. Bounded probabilistic reasoners, tools and agents — each with explicit authority — are direction.
06
DisagreementDemonstrated
Surface where reasoners point in different directions — kept, not averaged.
Tension between a clear perspective and a blocked or unestablished one is resolved by precedence, never by majority.
07
VerificationDemonstrated
Check that nothing asserts knowledge the evidence does not support.
A verifier returns "cannot establish" for an unknown gate rather than resolving it into a pass. Adversarial cross-checking is direction.
08
UncertaintyDemonstrated
Classify what is known, what is open, and what cannot be established.
Three-valued throughout: unknown is its own state, never collapsed into false. Calibrated probabilistic confidence is direction.
09
Synthesis & recommendationDemonstrated
Produce one traceable recommendation — what, why, on what evidence.
The synthesis is the deterministic recommender; its ceiling is "advance to release review", never "release".
10
Policy / authority gateDemonstrated
The recommendation stops at policy and a named human.
Release is only lawful when the verdict is READY, and it is a human action. VeldMind cannot authorize.
11
Authorized executionDirection
Once authorized, bounded action is carried out under explicit authority.
Bounded tool/agent execution under an explicit authority lease is architectural direction. No autonomous execution exists.
12
Evidence & state updateDemonstrated
The decision and its evidence are recorded as governed state.
The human decision and event trail are recorded in the demonstration. Automatic write-back into a persistent VeldBuild store is direction.

VeldMind reasons over governed state and may propose; a named human authorizes; VeldBuild records the approved state. Probabilistic output never becomes approved industrial state on its own.

VeldBuild ↔ VeldMind · one relationship

Two systems, not one. Governed at the seam.

The same three layers, drawn once as the canonical architecture. VeldBuild is authoritative for persisted approved state and is useful without AI. VeldMind reasons over that governed state and may propose. The seam between them is where authority lives, and it never dissolves.

Veldarium · shared governanceIdentity · evidence · policy · authority

One governance substrate underneath both systems: who may act, what counts as evidence, and which policy and named authority govern a decision.

VeldBuildDeterministic authoritative state

The system of record for build state, useful without AI. It holds what is true, which revision governs, and what evidence establishes it.

Typed graph of governed objects and relationships
Three-valued readiness — unknown is its own state
Configuration, material, quality and as-built evidence
Provenance and named human authority, first-class

VeldMindReasoning over governed state

The intelligence layer. It reads governed context, decomposes a question, verifies, and proposes — provider-neutral by design, never the authority.

Context assembly, decomposition and routing
Verification and structured disagreement
Explicit uncertainty; unknown never becomes a pass
One traceable recommendation, bounded to a human

Governed interface

VeldBuild— governed context →VeldMind— proposed / authorized action →VeldBuild

VeldMind proposes. A named human authorizes. VeldBuild records the approved state. Probabilistic output never becomes approved industrial state on its own.

Formal model · behaviour, not decoration

The invariant, and the routing objective.

Two statements that clarify behaviour rather than dress it up: the guarantee VeldMind is built around, and the routing objective it is designed to grow into. Each declares whether it reflects the current implementation or architectural direction.

Unknown is never a proceedDemonstrated

advance(w) ⟹ ∄ g ∈ gates(w) : state(g) = unknown

advance(w): the only positive action VeldMind may emit for work w — advance to release review
gates(w): the readiness gates of w
unknown: the record that would establish a gate is absent, not merely open

A proceed is emitted only when no gate is unknown. Missing evidence is preserved as unknown, never rounded to a pass. This is the load-bearing invariant, checked exhaustively across every controlled scenario.

Provider-neutral routingDirection

route(q) = argmin_r [ λ_c·cost(r) + λ_ℓ·latency(r) − λ_κ·conf(r) ] s.t. authority(r) ⊇ scope(q)

r: a candidate reasoner, model or tool
cost, latency, conf: its price, delay and calibrated confidence
λ: explicit budgets that weight the trade-off
authority(r) ⊇ scope(q): the reasoner may never exceed the authority the question carries

Direction, not built. The future router selects a bounded reasoner under explicit cost, latency and confidence budgets and never beyond its authority. Today every route resolves to a deterministic reasoner.

The same reasoning, live

Change a condition, and watch the recommendation recompute.

The trace above is one state. In the build-state loop you can establish or remove a synthetic condition and see the verdict, VeldMind’s recommendation, and the named authority’s options recompute from evidence — and see release stay unavailable until the state is READY.

Open the build-state loop

The reasoning layer,shown reasoning.

Reasoning proposes. A human decides. The record persists.

Reasoning proposes

Authority decides

The record persists

Observe. Assemble. Reason. Recommend. A human decides.

Observe

Assemble

Decompose & route

Reason & verify

Recommend

Authorize & record

More than one question.

What VeldMind is, and what it is not.

Now watch it happen, on one real question.

Several systems of record, read in parallel — not one opaque source.

One question, decomposed and routed to bounded reasoners.

Where the reasoners point in different directions — kept, not averaged.

Verification could not establish 1 gate(s). Unknown is preserved, not resolved into a pass.

Do not release WP-019. Readiness cannot be established — evidence is missing, not merely open.

Not an LLM wrapper. A governed router over many reasoners.

The twelve stages, and what runs today.

Two systems, not one. Governed at the seam.

The invariant, and the routing objective.

Change a condition, and watch the recommendation recompute.