Dashboard Lineage Audit

Every number on a trading dashboard implies a chain: a component rendered it, a loader fetched it, an artifact stored it, a publisher wrote it, a computation produced it. On most desks nobody has ever walked that chain end to end — the dashboard’s numbers were checked by looking plausible. The Dashboard Lineage Audit is a service engagement that walks the chain for every number on the glass, and reports what it finds in a vocabulary that includes the finding most audits are too polite to name: FABRICATED.

I did not invent this method for clients. I invented it for myself, and I have executed it twice on my own cockpit. Both passes caught real fabrications — hardcoded statistics presented as computed values, and telemetry generated from random numbers rather than from the system it claimed to describe. A method that has never caught anything is a ritual; this one has caught my own.

And the present state of my own surface belongs in the same paragraph, because an audit product that hides its own reds is selling the ritual: my advisory lineage rails are red today. I publish that here rather than let it be discovered. It is the reason the deliverable is a per-number ledger rather than a clean bill of health — including when the surface under audit is mine.

What it does

The audit is run by someone with no stake in the answer being yes and no memory of why a join was written the way it was.

• A full inventory of the display surface. Every number, gauge, and status indicator on the audited UI is enumerated. Nothing is sampled, because fabrications do not distribute themselves conveniently across a sample.
• A five-hop trace per number. Each displayed figure is traced through component → loader → artifact → publisher → computation. The trace terminates either at a real computation over real inputs, or at the point where the chain breaks — and the break point is itself the finding.
• Findings with honest statuses. Each number receives a status. VERIFIED means the chain closed. The statuses short of that are stated plainly, and FABRICATED — a value that is hardcoded, generated, or otherwise not produced by the computation it claims — is a first-class outcome, never softened into “requires further review”.
• The deliverable: a lineage ledger. The engagement produces a per-number ledger of the traced chain and its status — a document your risk function, your allocator, or your next audit can pick up and re-walk. Findings from client engagements: NOT YET PUBLISHED, and individual client results stay private to the client.

Five findings from the most recent pass over my own surface

In-house, 4 August 2026. Client findings stay with the client; these are mine, and each names a class of break rather than a one-off, because the classes are what recur on someone else’s glass.

• A count that meant three different things. A release note stated that “the 448 republished artifacts carry a new evidence block”. Traced: 448 was the number of changed paths — which includes a .sha256 sidecar beside every file. The actual artifact count was 239. The number that carried the evidence block was 62. One label, three denominators, and the only one that had been checked was the largest. The class: a count whose denominator changes silently between the claim and the check.

• Capability landed, wiring did not. Forty-four driver files carried the code to stamp provenance onto their output. Eight of twenty-one published families actually showed the stamp. Three drivers carried the code while their artifacts did not — they had been edited and never re-run. Reading the source would have confirmed the claim; reading the artifacts refuted it. The class: the code and the artifact disagree, and only the code was read.

• A style that never reached the screen. Three rows of a governance panel referenced a layout object with a one-character mistake that turned it into a differently-shaped object. The type checker had been reporting it for weeks. Every check run over that file excluded it as “the pre-existing errors” — so an error that was true had been reclassified as background. The class: a known-red file that every check excludes is a blind spot with a name on it.

• Absence rendered as good news. A performance figure that had never been computed was displayed in the colour reserved for a favourable value, because the code substituted zero for the missing measurement and zero cleared the threshold. Not merely a fabricated number — a fabricated number that read as encouraging. The class: the costliest fabrication is the one that flatters, because nobody investigates it.

• The defect was authored at the type, not the call site. A declaration promised a numeric field the upstream source can genuinely omit. Under a strict type checker, substituting zero was the only edit that satisfied the compiler — so the shortest correct-looking fix and the fabrication were the same keystroke. Patching the display would have left the cause in place. The class: when the type asserts something the data does not guarantee, every consumer downstream is pushed into inventing a value.

What it does not do

The Dashboard Lineage Audit does not certify a dashboard whose chains it could not close. A number whose lineage terminates in a broken hop is reported as exactly that — unverifiable — not granted the benefit of the doubt because the rest of the panel checked out. And the audit does not issue a clean bill of health as a summary judgement: the deliverable is the per-number ledger, and the dashboard is as honest as its worst-traced number.

Who it is for

Desks that inherited a dashboard nobody fully understands; teams whose UIs were built quickly, by contractors, or by agents; allocators and risk functions who want “where does this number come from?” answered with a trace instead of an assurance.