Under the hood
For builders, and for anyone weighing whether to run this with their own AI: what the system is, how it's checked, and the numbers so far, with their dates.
The architecture, briefly
- A set of valuation tools that an AI model calls (an MCP server with 22 tools): filings, financial history, industry benchmarks, the cost of capital, the valuation engine, the solver, sensitivity grids and simulations. The tools return numbers with their sources; the model returns judgment.
- A frozen case per valuation: the story, every driver with its reason, and the data it rests on, saved before any value is computed. A change makes a new revision with its reason; nothing is edited in place.
- An approval gate: no value appears until a person approves the case, and the approval is recorded with its time.
- A reviewer: a second AI session that reads the cases, the record and the filings and returns questions only. It can't see a value, change a case or run a command.
How the AI's judgment measures up
The agent was run on 14 of Aswath Damodaran's published valuations from 2021 to 2024, each at his date and on the data available then, three runs a case, on Claude Opus 5.5 (effort high) at commit 706736d (November 8, 2026). The measure is the gap between the agent's value and his, case by case.
| Set | Median gap |
|---|---|
| The 11 cases no change was tried on (the headline) | 42.4% |
| The 3 cases used to tune it | 34.3% |
| All 14 | 40.9% |
What it shows: the agent comes in below him in 11 of the 14 cases, leaning low on growth in years 2 to 5 and on sales to capital. Runs agree within about 10 points of each other, so a change smaller than that needs three runs a case before it counts.
A later round (November 8, 2026) found that rewording the agent's instructions didn't move its first draft; what moved the final result was the reviewer's questions, which brought the held-out gap from 42.4% to 34.3% on one run a case. The next attempts go into the tools, not the wording.
A post written by the system, and what it taught
The Comcast post went through the post-writing workflow on the system's own recommendations at every step. Reading it back found what the checks had let through: a price below both stories, which reads as a verdict without saying one, and raw tool output in the appendix. Both are now checks that stop a draft; the ADP post, written after, passes every check with no warnings.
What's checked on every build of this site
- No verdict words and no raw tool output on any page, by the same checks the posts use.
- Only fields on a written allowlist leave the laptop; nothing ranks the companies.
- Every figure in the animation is read from the session's own files; a number typed by hand stops the build.