EDGAR Intelligence, behind the scenes

Ten articles, one server, built in public.

The EDGAR Intelligence MCP was written up as it was built, one lane at a time, in a ten-part series on Medium: the summary tree, the numbers, the slide decks, the redline, the retrieval, the identity layer, the tool contract, the trust boundary, the failures that report success, and the methods revisited. Each part below carries its title, a short abstract, the link to the article, and its explainer video, which plays in place. The ten explainers run 1:28:18 end to end. Nothing is fetched until you press play.

The explainer videos are AI-generated: each one was created with Google's Gemini Notebook, formerly NotebookLM, as an overview of the article it accompanies, and it carries the same professional-review caveat as any AI-generated summary on this site. The articles themselves are written.

Head
1:28:18

Poster for the Part 1 explainer, titled Fast Trees: RAPTOR Retrieval, with three line drawings: an eye over a stack of files, a clock in motion, and a tree of nodes. — Explainer, 7:23. A generated overview of the article, played here.

Part 1 Fast Trees: Building RAPTOR Retrieval over SEC Filings with a Diffusion LLM

Read it on Medium Explainer 7:23

RAPTOR builds a tree of summaries over a filing so that a thematic question, such as Apple's overall posture on supply-chain risk, is answered from a synthesis rather than from whichever chunk happened to rank first. The catch is that every internal node is a summarizer call, run one after another, so the summarizer's speed and price are the budget. This part swaps in Inception's Mercury-2, a diffusion language model, and runs a cold build over Apple's fiscal 2025 10-K to see what that does to the cost math.

Poster for the Part 2 explainer, titled Deterministic AI: a question mark wired through a book labeled SEC data architecture to a card reading 2023 revenue. — Explainer, 9:25. A generated overview of the article, played here.

Part 2 $416,161 Million Is Not a Sentence: Keeping SEC Financials as Numbers, Not Text

Read it on Medium Explainer 9:25

Apple's fiscal 2025 net sales of $416,161 million are not a sentence, so they never enter the text index. Revenue, EPS, share counts and debt maturities come from XBRL, flattened deterministically from the SEC's companyfacts data into Postgres rows, and arithmetic over them runs in a calculator tool rather than in the model. The same discipline covers who owns whom: Section 16 insider trades, 13F institutional holdings and Schedule 13D/G stakes are structured tables too, and the language model only routes the question to them.

Poster for the Part 3 explainer, titled AI Blind Spot in SEC Filings: a magnifying glass over financial documents beside a robot with crossed-out eyes, labeled engineering fix. — Explainer, 10:58. A generated overview of the article, played here.

Part 3 Slide 191 of 442: What to Do With Tens of Thousands of JPGs

Read it on Medium Explainer 10:58

SpaceX listed in June 2026 and, when this was written, had filed no 10-K or 10-Q; its risk disclosure lived in an S-1 and a 424B4 whose pages are images. This part builds the lane that gets words out of tens of thousands of JPGs, with three rungs chosen at ingest by looking at the content: plain OCR for text-dense pages, structured OCR for text trapped inside diagrams, and a vision model only for pictures of numbers, so that the supply-chain passage on slide 191 of 442 can be found and cited by its link. It also records what a well-known legal-AI product did with the same question, once, on one day, and gives it credit for not fabricating an answer.

Poster for the Part 4 explainer, titled The 10-K Mystery: a column of filing text, an equals sign, a question mark, and a brain crossed out. — Explainer, 8:14. A generated overview of the article, played here.

Part 4 Seven of Ten: "New" Risk Factors that Were Not New

Read it on Medium Explainer 8:14

0 of the ten highest-ranked new clauses were a year old

The first honest redline over Apple's risk factors reported twenty clauses added and twenty removed between two annual reports whose Item 1A differs in length by a rounding error, because Apple had reflowed the section and every chunk boundary moved. Seven of the ten highest-ranked new clauses existed word for word a year earlier. The fix was not a smarter diff: run the section summaries on both sides first for salience, then use the clause diff only to pin the exact language for the topics the summaries flagged.

Poster for the Part 5 explainer: a slide headed The Same Query, Twice, One Parameter Changed: Alpha, with the caption one query, two retrievers, opposite answers. — Explainer, 12:04. A generated overview of the article, played here.

Part 5 Buying Chips, Selling Compute: One Query, Two Retrievers, Opposite Answers

Read it on Medium Explainer 12:04

One query, run twice against the same index with one parameter changed: pure keyword search found SpaceX's chip-procurement risk factor three times over, and pure dense retrieval ranked a $920 million a month revenue contract first because it was also about GPUs, the wrong side of the transaction. Phrase the question the way a lawyer would and the two retrievers trade places. This part is the hybrid lane: reciprocal rank fusion, summary-augmented chunks on the dense side only, context expansion computed at ingest, a reranker allowed to disagree with all of it, and a blend dial handed to the language model, along with the one row of the dial's own cheat sheet that it turned out not to keep.

Poster for the Part 6 explainer, titled Corporate Identity Mystery: a folder under a magnifying glass, surrounded by org charts, handshakes and SEC documents. — Explainer, 9:02. A generated overview of the article, played here.

Part 6 Broadcom Was Never Broadcom: The Hand-Curated Layer That Decides Who You Are Asking About

Read it on Medium Explainer 9:02

The chip company trading as AVGO is legally Broadcom Inc., and its direct lineage runs through Avago, not through the original Broadcom Corporation it bought in 2016: three CIKs, one continuous business, and an empty former-names field at every step. This part is the hand-curated identity and lineage layer, six Postgres tables and no graph database, that decides which company a question is about and shows its judgment calls in the tool output rather than folding them into a score. Every retrieval lane answers faithfully about whatever CIK it is handed; none of them can tell you the CIK was wrong.

Poster for the Part 7 explainer, asking how do you make an AI follow rules it was never trained on, with sketches of financial databases, an agentic workflow and tool text. — Explainer, 8:15. A generated overview of the article, played here.

Part 7 Reflections on Tools: Getting Agentic RAG to Take Three Steps

Read it on Medium Explainer 8:15

Part 4's three-step method, summaries for salience, diff for evidence, counting for verification, was inert as an article, because a model choosing tools has never read it. What it reads is the tool list. This part is about where the few hundred characters that make a model take those three steps in order actually go: a guidance tool that returns markdown, docstrings that carry the workflow at the point of temptation, return paths that name the next call, and the sequencing that the MCP specification has no home for.

Poster for the Part 8 explainer, titled R2-D2 and the AI Trust Gap: a robotic hand plugging a network cable into a small box, framed by cartoon documents, padlocks and gavels. — Explainer, 8:39. A generated overview of the article, played here.

Part 8 R2-D2, You Know Better Than to Trust a Strange Computer: What It Takes to Trust One Anyway

Read it on Medium Explainer 8:39

A tool description is the one channel a model cannot refuse, which is exactly what OWASP describes as MCP tool poisoning. This part follows the defense from its origin as a detect-secrets plugin built for a USPTO server, through a colleague's fair correction that a commit-time scanner does nothing about injection arriving inside a retrieved filing, to the runtime scanner that labels and never strips, and the invisible Unicode channel that hides a directive inside an emoji. Then it keeps going, to the questions a firm asks before it will point this at a client matter: what the logs hold, where the query goes, what vendors see, and who is allowed to sign in. Those turn out to be one argument, not five.

Every call well formed. Every tool used precisely as documented. Nothing anomalous in the logs, because the documentation was the attack.

From Part 8

Poster for the Part 9 explainer, titled MD&A Hallucinations: a sketch of a Western false-front building braced with timber scaffolding, surrounded by a chain link breaking apart, a checklist, a magnifying glass over a question mark, and crossed-out tools. — Explainer, 6:55. A generated overview of the article, played here.

Part 9 The Whole MD&A Was a Hallucination: Failures That Arrive Looking Like Completed Work

Read it on Medium Explainer 6:55

Exxon's Item 7 is one sentence pointing to its Financial Section, and McDonald's is one row of a lookup table on the last page of its 2019 10-K; handed pointers like those, the summarizer wrote a fluent MD&A anyway, putting Exxon's 2022 revenue at $84.9 billion against a filed $413.68 billion. A two-probe audit of the whole production index found 533 section summaries of that shape across 326 of 477 issuers, and a regex and a length check that run before and after the model now report a pointer as a pointer. The rest of this part catalogs failures that arrived looking like completed work: jobs that succeeded having done nothing, a healthy container with no workers in it, a keyword tokenizer that discarded every digit, and exhibits routed but never read. The habit that found them is to count what a thing did whenever it says it worked.

Poster for the Part 10 explainer, titled The Illusion of Change: a magnifying glass over a handwritten document, surrounded by scales of justice, rising charts, a funnel of papers, circuit lines and a handshake. — Explainer, 7:23. A generated overview of the article, played here.

Part 10 EDGAR and Jev: After Seven of Ten, Forty-Seven Became Twenty-Eight

Read it on Medium Explainer 7:23

Part 4's Apple redline counted 47 substantive changes between two annual reports that had mostly been reflowed; the same call over the same filings now reports 28, because a section-level judge ruled that 19 pairs were the same language re-wrapped, and that judge can only take noise out, never add a change or delete a pair. This part brings in Jev, TypeSafe's System One model, which returns typed decisions and probabilities rather than text: first as a read-time backup judge whose reflow verdict passes only if every sentence already existed verbatim, then as a planner that proposes the first calls for a question while code writes every sentence of the plan. It ends where the series began, with Part 1's question about Apple's supply-chain posture, now answered across both text lanes and the company's history from one call.

After Where the server is, and how to work it

About EDGAR Intelligence what it reads, what it answers, what it will not do, and what it costs
The walkthrough: nine prompts, one session the prompts that take an assistant through a multi-lane analysis end to end

EDGAR Intelligence is a research tool over public SEC records. Summaries and rankings are AI-generated and require professional review; verify against the linked source filing before relying on any output. The explainer videos on this page are AI-generated as well, each created with Google's Gemini Notebook, formerly NotebookLM, as an overview of the article it accompanies. Nothing here is investment advice or a recommendation regarding any security.