Governing the Model Vault
Governed Reuse of Validated Models
James Wolstencroft · June 2026
The Tekmerium · Paper V of VI
Abstract
A validated model is one of the most expensive things an engineering organisation owns: slow to earn, costly to prove, wasteful to rebuild from scratch each time. The discipline for keeping what an organisation has proven and reusing it safely is the subject of this paper. It specifies the Vault: a governed store of validated models of record for a single system, with the rules under which an engineer or an automated process may draw one. Its spine is four disciplines (admit, stamp, scope and gate):
A model enters only with an owner, version, provenance and a graded assurance stamp.
It is reused rather than regenerated.
It is valid only within a sanctioned envelope.
A use-gate refuses any request outside that envelope, routing the gap to a human.
Reuse alone justifies building the Vault today; machine authorship removes the choice. What reuse recommends, AI makes mandatory.
1. Keeping what you have proven
A validated model is sunk effort that an organisation should never have to spend twice, but reuse only delivers if what is reused can still be trusted.
Proving that a thermal, fatigue or control model behaves as claimed takes time, instrumentation and judgement; once earned, that proof is an asset worth governing, versioning and drawing on again rather than reinventing. But a model pulled from a colleague’s folder, or summoned to fit a brief, carries no pedigree, no record of how far it was proven or under what conditions. Safe reuse therefore demands that each model carry its provenance, assurance level, and valid bounds, and that something refuses to hand it out beyond those bounds. This paper addresses the governed-reuse problem; who remains accountable when a machine does the drawing is left to the Accountable Author.
2. The Vault as a distinct component
The Vault is the governed store of models from which a system is built, distinct from the audit trail above it and the System of Record beneath it.
It contains the validated, versioned, assurance-rated artefacts that an AI may use but never create on demand (The Continuum, Ch. 23). It is not the audit trail, which records reasoning; rather, the Vault holds the fundamental components that reasoning chooses from. It is also not the System of Record for the substrate, where the authoritative bytes of an artefact are stored and referenced. The System of Record answers, “Where does the authoritative copy live?” while the Vault asks, “Is this model suitable for use, and within what boundaries?” One concerns location and currency; the other concerns assurance and permissions. Positioned on the substrate, the Vault does not replicate it.
3. The assurance ladder
A model in the Vault carries a graded verdict on how far it has been proven: a ladder, not a tick-box.
The stamp has three rungs (proven mathematically, against real-world data, or in real-world use), and they are not interchangeable: a model validated only in mathematics counts for little beside one proven in service (Ch. 23). The stamp travels with the model and the evidence that justifies it, so “validated” is never an unqualified assertion but a claim with a level and a basis. This matters most where speed is greatest: when a machine composes a design in minutes, each ingredient’s assurance level is the only thing that tells the accountable engineer whether the composite can be trusted. An outdated stamp that no longer serves as proof is worse than having no stamp at all, because it grants unwarranted confidence.
4. Reuse over regeneration
The model is a proven artefact with a stamp, not a freshly generated approximation
When a generative system reaches for a thermal model, it must receive the validated one from the store, assurance rating attached, not a plausible but unproven guess that arrives with a confident interface and no pedigree (Ch. 23). Which version is canonical is not the Vault’s call (that judgement belongs to the Digital Backbone of ‘Engineering the Decision Trail’) but holding that model under governed conditions and enforcing its terms of use is. The reuse discipline closes generative design’s most seductive shortcut: that a model which “worked beautifully” once can be summoned again at no cost, when what is summoned must be the proven artefact, not a fresh guess wearing its name.
5. The sanctioned boundary
Each model is proven only within a sanctioned envelope; the use-gate refuses any request beyond it and hands the gap to a human.
Validation is never unconditional: a thermal model proven to 60°C is not sanctioned at 90°C, a fatigue model proven for one load spectrum is not proven for another, and the envelope makes those bounds explicit and machine-readable (Ch. 23). The envelope only matters because it is enforced: when an automated process requests a model outside its range, the gate declines rather than let the AI silently extrapolate into territory it was never proven for. This is what separates a Vault from a library: a library lets you check out any book; the Vault checks the request against the model’s proven bounds and converts a silent, invisible extrapolation into an explicit, logged refusal and a task for an engineer.
6. The Vault Gate
One protocol enforces the disciplines in six steps: admit, stamp, scope, gate, record and re-qualify.
Each step is a rule the gate enforces:
Admit nothing without an owner, version, provenance and a validation level.
Stamp the assurance level and attach its evidence.
Scope the model to a sanctioned envelope.
Gate the reach so automation only accesses the Vault within each envelope.1
Record every use back into the audit trail of ‘Engineering the Decision Trail’.
Re-qualify, expire and audit, since validation is perishable.
A worked case makes it concrete (numbers illustrative): a generative tool reaches for a thermal model proven only to 60°C and lays out power electronics for an 85°C market, the packs throttle and fail, and nothing records the overreach. Run through the gate: the 85°C request is refused; a higher-temperature model is commissioned and admitted; and the use is logged. A model whose proven level and envelope cannot be stated is not an asset but “a liability with a confident interface.”
7. Reuse recommends the Vault; accountable AI makes it mandatory
Reuse is reason enough to build the Vault. Generative AI removes the choice.
A validated model is expensive to build and wasteful to rebuild, so an organisation that has proven one ought to keep and govern it; reuse alone justifies a Vault now. But an authoring machine will soon compose systems from stored models faster than any human can vouch for them, and it cannot be let loose on a store that is not already governed: the Vault that reuse recommends, AI makes mandatory. The guardrail is plain: for generative architecting to be trustworthy rather than merely fast, its ingredients must come from a managed, validated store, and its choices must be recorded in the trail. Without the Vault, an organisation accumulates fast, plausible, untraceable designs it cannot defend; with it, the same speed produces work a human can still sign for.
8. Limitations and open problems
The Vault governs one system, presumes honest inputs, and grades trust without quantifying it.
The single-system boundary is deliberate: reusing a validated model on different projects merges into the Hyper-Continuum, with family-scale reuse and product line engineering reserved for a companion volume. Enforcement requires complete capture: a model with an overstated stamp or missing a dimension defeats the gate. The assurance ladder is graded but not yet quantified: how a stamp is scored, how confidence decays, how partial validation is represented, all open. The figures quoted are illustrative, not measured. The Vault’s strength depends on its substrate, audit trail, and Digital Backbone, all developed elsewhere in the series.
References
Regulation (EU) 2024/1689 – EU Artificial Intelligence Act, 12 July 2024.
Directive (EU) 2024/2853 – EU Product Liability Directive, 23 October 2024.
Tabassi, E. – AI Risk Management Framework (AI RMF 1.0), NIST, 2023.
ISO/IEC 42001:2023 – Artificial Intelligence Management System.