AHF catches inconsistencies across a set of documents and pinpoints when a stated position or policy changes.
The real target is internal material: documentation, email, JIRA and Slack. We don't expect access to that for a demo, so we use public filings instead.
How it works: AHF reads documents one at a time, in the order they were published. Each one adds to a running picture of what has been stated, its ground truth. When a new document changes or contradicts that picture, AHF flags it, and a human reviewer decides whether it matters.
For every check it gives one of three answers: Yes, No, or Can't tell — and, once the evidence behind an answer is withdrawn, no longer knowable. It keeps full history, so you can see exactly what was known on any past date.
About 15 minutes. Each demo has a status bar at the top showing the system's answer at every step. If you watch one thing, watch that.
A company reports a total of 100. Two teams report their parts: 40 and 60. Over several weeks, new information arrives. Step through with the arrows at the top or your arrow keys.
Then go back to step 3. You see exactly what was known on that date, not today's view with later data hidden.
Six documents. The question: what did the bank report for Q1 2012, and does each number still stand?
Why this example. It's familiar, so it needs no background. And behind these six documents sits the bank's full record, 53,686 figures across 3,662 filings, run through the same rules.
One finding from the full record. Three filings report a 2011 figure as positive. The 2012 amendment reports it as negative. If the system sees only the 2012 filings, it treats this as a normal restatement. With the full record, it flags the change as unexplained: three earlier filings say otherwise, and none of them has been retracted. The second answer is correct. Takeaway: limited data can miss problems, and can also produce confident conclusions that more data overturns.
There's no AI model comparison here on purpose: the episode is well known, so a model might answer from memory rather than the documents. In this demo the system itself uses no AI in its reasoning.
Five years of one bank's filings are read first. Then its 2024 filings arrive one at a time. Press “Send next filing” at the bottom to send the next one, or click any filing in the strip near the top to jump to it, forwards or back. After each filing the screen shows:
Marking a question. Under each question, Raise to owner and Mark fine record what a reviewer decided. Every mark can be undone, and Reset marks clears all of them for that bank. ⚠️ Marks are kept on this server and everyone who opens the page sees them; this demo has no logins. A mark changes nothing the system concluded.
Finding another bank. The search box at the top finds the banks built on this server, by ticker or name. A ticker that is not built shows where a new one would be ingested — its SEC filings, earnings-call transcripts and investor presentations. That is not available yet: building a bank is a background job, not something the page can do while you wait.
AI in this demo. The AHF reviewer uses a language model (Gemini) under AHF's rules to read the filings, pull out statements, find the tables and ask the questions a careful reviewer would ask. AHF's code checks every quotation against the filed text, and a human reviewer decides whether to raise it. One of the four matches came from the AHF reviewer. Figures restated in a later filing, which are usually routine, are listed under “Watching” rather than raised. The method was developed on this bank, with its letter open, and the screen says so.
Each card also has folded Developer notes, for the people who build this: identifiers, and a link to Inspect, the raw store behind the cards. The earlier layout of this screen is still at Classic view.
Two more banks from a set of twenty run without seeing any letters: the one with the most raised and the one with the fewest. Whether those match their SEC letters has not been scored yet.
| Does it find real errors? | Partly shown. On the bank in Demo 3, 4 of 34 SEC questions match something it flagged months before the letter — but that bank is the one we developed on. On twenty banks we did not develop on, the fixed rules matched none of 54 SEC comments, and the AI's questions are not scored yet. In Demos 1 and 2, all scores come from planted problems. |
| Is it better than a good AI model? | On accuracy, slightly: a strong model with the same data and rules scored 73/75; this system scored 75/75. The bigger advantages: every answer traces to an exact source, a rule check costs nothing to re-run, and it gives the same answer every time. |
| How much of it is AI? | In Demos 1 and 2, none in the reasoning or calculations; an LLM wrote the plain-English captions afterward, working only from numbers that were already calculated. In Demo 3, AI reads the filings and suggests questions; code checks every quotation and does every sum, and a human reviewer decides. Reading one bank's three annual reports costs about $1–2. |
| Can we use it on our data? | That's a pilot, and it's the next step. |
| Who can see what? | Everyone sees everything. There are no permissions yet. |
The demo is read-only. Nothing you click changes what anyone else sees.