Elastic Assurance: Compute Broadly, Disclose Narrowly

How a second, AI-native assurance plane lets an organisation examine everything, disclose only what deserves human judgment, and file every review as dated, defensible evidence beside the formal report.

Scott Farrell · LeverageAI · A worked example on the public record of Transgrid and the Australian Energy Regulator

The green light is not reality

A large organisation contains thousands of local facts, judgements, exceptions and trade-offs. Management cannot read them all, so governance compresses them: a handful of KPIs, a few control attestations, an amber/green/red status, a short commentary box, and an escalation only when a predefined threshold is crossed. That compression is necessary. It is also lossy.

The green light is not reality. It is a deliberately low-resolution representation of reality, designed around the amount of attention available at the management layer. Green means the area passed the limited tests chosen for escalation. It does not mean nothing else interesting, risky or consequential is happening.

The soft-data exhaust — reports, correspondence, meeting records, operating chatter — routinely carries dimensions the reporting model cannot: implementation differs materially between two sites running the same control; the same obligation consumes radically different human effort; field staff quietly qualify a number the final report presents with confidence; workarounds have quietly become normal practice; a risk category is emerging for which governance has no name. The organisation is not concealing these things. Its reporting structure simply has nowhere to put them.

The compression inversion

Traditional governance has to compress first, then review — complex reality, then selected metrics, then a traffic light, then management attention. That order is forced because humans cannot inspect the underlying estate at full resolution.

AI lets the order invert: complex reality, then broad machine review across many axes, then selective findings, then management attention. The system reads at high resolution and reports at low resolution. This does not mean flooding executives with more detail — human attention is still scarce. It means the organisation no longer throws most of the nuance away before anybody has examined it.

Compute broadly. Disclose narrowly. Drill down on demand.

Two assurance planes

You do not ask the AI to overrule governance. You run two planes side by side. The Formal Assurance Plane is the established control system — mandatory KPIs, approved procedures, regulatory reports, formal status, recognised escalation paths — and it keeps its regulatory meaning untouched. The Exploratory Assurance Plane is the AI-native layer: it reads the soft estate, compares implementation across teams and sites, searches for contradictions and known absences, tests alternative interpretations, and reports findings with evidence — without automatically changing formal status.

That is better than forcing every weak signal into amber. The traffic light keeps its meaning; the organisation still admits it has discovered something outside its existing projection. Management does not receive another giant dashboard — it receives an Assurance Finding Card: what unusual shape or omission was identified (Finding), why it may matter despite the formal status (Significance), the receipts (Evidence), what is known versus inferred (Uncertainty), what to ask next (Question), and what to do — review, resource, monitor, or nothing (Response).

Why "elastic"

Human assurance works through scarce teams: audit samples, a few chosen KPIs, specialists on known risks, committees on what is already escalated. AI assurance can fan out across every site, every control family, every reporting period and many independent questions at once. It is not literally unlimited — it is bounded by budget, access, model quality and evaluation — but compared with human review capacity it is elastic. Once the ingestion and governance plumbing exists, adding another site, another lens or another standing question is an incremental compute cost, not another multi-month audit programme. The bottleneck moves from "how many things can we afford to examine?" to "which findings deserve scarce human attention, and which questions should the institution keep asking?"

Un-invested, not objective

Staff and management are not usually dishonest, but green has organisational value: it avoids escalation, protects schedules and budgets, and spares an already-stretched team more scrutiny. So people can rationally work harder, compress uncertainty and absorb problems locally to keep the status green. The AI reviewer sits in a different incentive position — it gets no bonus for keeping a project green and is not defending a decision it made last year. But be precise about the claim: the AI is not objective. It is un-invested. It is biased toward what was recorded, what it can access and the lens it was given — and those biases can be disclosed and changed, which human investment in a narrative usually cannot. That is why multiple declared lenses — delivery, engineering-risk, workforce-capacity, regulatory-prudency, community — are allowed to disagree inside the same review.

Six boundaries make this both radical and deployable:

The principle that holds it together: record the review, but never let the review outrank its exhibits. This is a near-perfect deployment of what we call the Lane Doctrine — batch the review, ship the package, and let existing governance decide. The AI does not prosecute the project team. It builds the evidence room; humans decide what the evidence means.

A worked example on the public record: the Transgrid wedge

Transgrid, the operator of the New South Wales high-voltage transmission network — roughly 11,500 kilometres of line and 136 substations and switching stations⁹ — is close to a perfect environment for this. It is finishing one badly over-budget megaproject, building another under intense social-licence scrutiny, redesigning the grid's technical foundations as coal retires, and fielding extraordinary new data-centre demand — while its regulator raises the evidentiary bar on every dollar. (This is an outside-in reading of the public record; nothing here implies any engagement.)

The usual AI conversation in a business like this is drones for line inspection and copilots for document processing — the eighteen-project deck, all traffic lights, zero terminal value. That is horse optimisation. The higher-value move uses a two-project wedge.

Pilot 1 — a completed but contested project as a labelled corpus

On the AER's like-for-like basis, Transgrid is seeking an additional $1.142 billion (2022–23 dollars) on Project EnergyConnect, on top of the previously approved $2.121 billion for the NSW component — about $3.263 billion in real terms — a request the regulator says would add $173 million to revenue in 2027–28 and roughly $18 to a residential bill that year.¹ Meanwhile the physical project has raced ahead of the argument: Transgrid announced in June 2026 that its 700-kilometre NSW section was complete and energising.²

The infrastructure may be physically complete, but the argument over why it cost what it cost — and who should pay — is only beginning.

The regulator is now weighing which events were genuinely unforeseeable, which costs arose from contractor failure, which risks should have sat with shareholders, and when the likely cost trajectory first became visible internally.¹ That is not a document-drafting problem. It is a demand for a causal record of the project — contract decisions, variations, risk movements, engineering judgements and regulatory representations, joined chronologically with receipts. Use that completed, contested project as a labelled historical corpus and ask a single question: what was the emerging shape of the overrun, when did it become visible, and which organisational signals preceded formal recognition?

Pilot 2 — the same questions on a live, still-moving programme

Then apply the learned questions prospectively. Transgrid's system-strength preferred portfolio was estimated at about $6.3 billion; in April 2026 it notified the AER of a material change because synchronous-condenser procurement costs had risen more than 30%,⁴ and by July it had published a revised portfolio giving grid-forming batteries a larger role.⁵ Separately it is seeking approval of a $1.185 billion System Strength Project for 2026–31, where the regulator has isolated risk costs and labour/indirect costs as its two focus areas.⁶ This is a living decision system — prices, retirement dates and technology credibility all moving underneath the plan.

HumeLink shows why the findings live in the joins. The AER approved $3.965 billion in Stage 2 capex — cutting $314.4 million from the application — and tied its decision explicitly to landholders, communities and social licence.³ Cost, schedule, environment, land access and community commitments are reported in separate lanes, but the emerging failure crosses them: do repeated landholder complaints later track to access delays; does schedule pressure make consultation language more certain while the underlying issue stays unresolved? No ordinary dashboard owns those questions. And notably, Transgrid's own operational-technology case admits the constraint directly: its public material warns operators may become overburdened confirming information across multiple sources, and it seeks $163.5 million to help⁷ — an explicit, first-party statement of the management-attention problem this whole architecture addresses. Nearby, 14 GW of speculative data-centre enquiries within 12 kilometres of Sydney West⁸ pose a soft-data question — which of these are credible? — worth billions in avoided augmentation.

The Soft Attestation Package

Each review is filed as a Soft Attestation Package. The crucial technical distinction: the package attests that a defined review occurred against a known evidence state — not that the AI's conclusion is true. That keeps the AI a witness, not an oracle. A package records its subject and the formal status reviewed; the evidence state examined (with timestamps and source pointers); the review specification (standing questions, lenses, model versions); ranked findings, not asserted root causes; verbatim exhibits and resolvable pointers; contradictions and absences; any hard-data checks; a counter-case; a workforce implication; the human disposition (accepted, rejected, investigated, monitored, or promoted); and a seal and version with the diff from the previous review.

Then it is refiled into the project's soft estate as new, dated, derived data — the same write-back move Karpathy made explicit for knowledge bases in April 2026.¹¹ It ranks below primary evidence, cannot cite itself as proof, points every finding to exhibits, and goes stale when its supporting evidence changes. So the review compounds, but never promotes its own earlier opinion into ground truth.

How was green produced?

Attach a package to every formal report and the most important question stops being "was the project green?" and becomes how was green produced? A green state has several very different underlying compositions:

This makes the project record bitemporal: what appeared true at the time versus what the organisation knew at the time — the difference between honest uncertainty and organisational blindness. It also protects honest teams: it can show a concern was considered and rationally rejected on the evidence then available, or that management funded extra capacity the moment the human reserve margin fell. The system does not merely catch failure. It preserves evidence of reasonable contemporary judgement — and it means you ask the system what the soft data showed at the time, instead of doing archaeology on the failure years later.

Standing Questions, and your first plane

The human panel's job is not to read the exhaust. It is chairmanship: decide which questions matter, set the North Star and boundaries, challenge evidence, and own the decision to act. When a good question proves useful — "why do two sites running the same control need radically different discussion and exception handling?" — it is promoted from a one-off investigation into a versioned Standing Question with a baseline, escalation criteria, a cadence, an owner and retirement conditions. It is a Question Ledger entry promoted into an operational review lane.

You do not build the enterprise version first — that would fossilise before completion. Begin with one bounded vertical slice: one operational control, two sites that implement it, one year of reports and correspondence, and a small human panel who knows the work. The purpose is to learn which sources carry signal, which comparisons produce insight, and what evidence humans need before taking a finding seriously. From there scaling is radial — same control across more sites, related controls, the same question across control families, more sources, more lenses, enterprise-wide standing questions.

This is not a faster dashboard. It is a new institutional organ — one that reads before it compresses, compares before it samples, and asks many questions before choosing which one deserves a human's scarce attention. For a regulated entity whose entire business is arguing why to a regulator, that is a compounding asset, not a productivity toy.