the blog
Notes on keeping AI-built data pipelines honest: why dashboards lie so quietly, and what it feels like when yours can finally prove itself.
September 2026 · Jeff MacDonald · build it yourself
Sixty lines of Python that sign every pipeline stage and catch an edited file at the exact link. Then the five places the toy breaks (re-saves, magnitude, keys, unwrapped stages, verifying elsewhere) and how Tamper Signal handles each.
read the post →September 2026 · Jeff MacDonald · checklist
One number on the AI-built dashboard looks off. Five checks against the export (row counts, column totals, spot checks, date edges, repeated keys) that explain the gap to the cent, then receipts so the next refresh checks itself.
read the post →September 2026 · Jeff MacDonald · tutorial
Pandas never raises when a NaN filter, a merge on duplicate keys, or a coerced column eats your rows. First the fixes that need only pandas, then a signed receipt for each step that shows exactly where the rows went.
read the post →July 2026 · Jeff MacDonald · 2.1.0
Tamper Signal 2.1 puts the furniture away: the status light stays byte-identical, and the badge, verified table, and inspector console behind it become one surface, the Signal Room. Green earns silence; red leads with the break in business numbers; and one attach call now ships both halves, so the light can never dead-end in raw JSON again.
read the post →June 2026 · Jeff MacDonald
Tamper Signal 2.0 adds a live-source watcher: point it at a feed and it keeps that source on the same signed chain, appending new data on its own, but a retroactive change to an already-settled value is never signed unattended. It is withheld for a human, with a chain-of-custody console and a live weather demo to prove it.
read the post →June 2026 · Jeff MacDonald
Content provenance is dying in the social feed, but in a delivery pipeline you control the channel. A chain-of-custody SOP for social-media data, plus how to wire Tamper Signal into an AI vibe-coded visualization, with the implementation in Python or Node side by side.
read the post →June 2026 · Jeff MacDonald
AI can write a working pipeline in minutes, but it can't prove the pipeline didn't drop rows, double-count, or coerce a column on the way to your dashboard. Why your tests catch only the failures you already imagined, and how a signed receipt catches the one you didn't.
read the post →June 2026 · Jeff MacDonald · 1.7.0
Every BI tool has a Download button; none of them give you the proof. A verified bundle takes the data out with its receipts attached, so a CSV you export re-verifies as JSON and the light stays green on someone else's machine. And you can bring an updated file back the same way.
read the post →June 2026 · Jeff MacDonald
A chart asks you to believe; a table lets you check. The case for shipping a Data tab next to every chart: the raw verified rows the numbers came from, re-hashing themselves in the browser, so a doubted number is one click from the data instead of a screenshot in a Slack thread.
read the post →June 2026 · Jeff MacDonald · 1.6.0
Real reporting data drifts day over day as it matures, and a trust tool that fires on normal drift gets muted. Tamper Signal 1.6.0 adds signed tolerance bands and run history, so the light learns what normal looks like and stays believable on a pipeline that refreshes every morning.
read the post →June 2026 · Jeff MacDonald
The longer version of the homepage: AI-written data code fails silently. It drops rows and double-counts without ever crashing. A signed receipt at every step proves your numbers still descend from the source, and names the exact link if they don't.
read the post →
