Audit-Proof AI Architecture: Documentation & Re-Runnable Demos | Mycroft

Introduction: Documentation, Auditing & Re-runnable Demos

[0:00]Hi, I'm Am Molly. This video is about the last week of the build. Writing the documents that explain the system, discovering a stated requirement I had met zero times, and turning the demo into something that reruns instead of something that was true once. The code was finished. What was missing was the part that let someone else understand it without me sitting next to them. Three things this week, and none of them is code. the documents that explain the system, a literature requirement I had written down and never met, and a demo that reruns instead of one that was true on the day I recorded it. Every count here is measured against the last commit, not remembered. The plan named

Auditing Plans vs. Commits: Uncovering Missing Core Docs - Honest Prior-Art Literature Auditing & Citation Tracking - 3-Layer Architecture: Immutable, Append-Only, and Rebuildable - 8-Link Evidence Trace: From SEC Filing to Published Headline - Re-runnable 5-Act Demo Script with Zero Write Statements

[0:35]Six documents rather than assume I checked each one against the last commit. Three of them did not exist at all, not thin, not stale, absent. The proposal, the system architecture, and the data architecture, the audit found them. The plan had not predicted them. Then the one that stung. The plan says the published documents must site the prior art literature honestly. I gpped for every name zero and the honest version of that is narrower than it first looks. Two published documents were genuinely at zero. The third was the proposal which did not exist yet. So its zero is a missing file and not a failed requirement. A citation count is not a literature review and most of those 22 are context. One is loadbearing. Gornall and Strebulie established that funds write up every share class to the latest round price.

I reproduced that on this cohort and it is the reason dispersion is measured per company with the share class recorded. It overturned an earlier drafts rule. The architecture document spine is one idea. 15 tables in three layers and each layer is allowed to change in a different way. What the SEC filed is immutable. What a human decided is append only. What the code derived is rebuildable. Every filed holding carries exactly one match decision which is why the middle layer has the same row count as the first. The data architecture document ends with a trace. One headline, 25% of consecutive observations unchanged, followed down eight links. The published sentence, the generated figure, the function with its guards, the panel, the match decision, the named human behind it, the filed position, and the archive file. The SEC published. Break any link and

You have an output, not evidence. The plan said record a demo. I wrote a script instead. Five acts. Every figure quered live. So, a transcript that disagrees with the database is a bug rather than an old file. It writes nothing. Zero write statements across the demo and the server it calls. Act two shows a decision a named human already made and deliberately does not make a new one. Four things this week does not establish. The audit covers the six documents the plan names and no others. A citation count is not a literature review. The right statement count is a static scan of five files, not a proof. And the run log that makes the chain datable has two rows in it which is a start and not a history. Week 11 on one page. Six documents named in the

Plan, three of which did not exist, now 444 lines written from nothing. A stated requirement that had been met zero times and two published documents, now cited 13 times, one of them loadbearing enough to overturn a rule. 15 tables in three layers, each with its own rule for changing one published number traced down eight links to the archive file the SEC published. And a demo in five acts with zero write statements. Your turn. Paste this into Claude. Take your own plan. and list every requirement it states as a must and grep for evidence of each one against your last commit rather than against what you remember. Count the ones at zero. Mine was one and I had written it down myself. Documenting the system and making the demo rerunnable. Week 11 of the private eye valuation agent amali for humanitarians.