Upload a 大量保有報告書 to Claude and ask "What is IAT Partners' ownership stake in Toyo Engineering?" You'll get something like: "Based on the filing, IAT Partners appears to hold approximately 32.7% of shares."
That answer is probably right. But probably right is not good enough.
The actual figure — extracted from the typed XBRL element RatioOfSharesHeld in filing S100XNCJ — is 32.72%. The difference looks small. But the difference in how you got there is everything.
The grounding trap
There's a popular belief in AI-native finance that RAG — retrieval-augmented generation — solves the accuracy problem. Upload your documents, ground the model, get better answers.
It doesn't solve it. It reduces hallucination rates. It doesn't eliminate them. And in institutional finance, the gap between "usually right" and "provably right" is the gap between a useful tool and a compliance liability. Directionally correct is not enough. Small errors compound, and trust is everything.
The problem isn't that the LLM is stupid. The problem is structural. When you ask a language model to extract a number from a PDF, it's doing pattern matching across a probabilistic token space. It doesn't know what an XBRL element is. It doesn't know that the same filer name appears in three different encodings across nine filings. It doesn't know that filing S100XNFM is an amendment that supersedes S100XNCJ.
The architecture decision
Axiora separates AI reasoning from data retrieval. Completely.
AI handles one thing: interpreting what the user is asking. Natural language in, structured query out. When you ask "who's been selling Toyo Engineering?", the AI layer parses intent and routes to the correct tool. It never touches a number.
Code handles everything else. The XBRL parser extracts values from typed elements — deterministically, reproducibly, the same way every time. The entity resolution engine groups ＩＡＴパートナーズＬＬＣ, ＩＡＴ Ｐａｒｔｎｅｒｓ ＬＬＣ, and IAT Partners LLC into a single canonical entity. The amendment chain resolver knows which filing supersedes which. The trajectory builder stitches nine separate filings into a single ownership arc: 32.72% → 20.16% over 21 days.
None of this involves a language model. None of it is probabilistic. Every step is auditable.
What the domain knowledge layer does
Raw XBRL extraction is the foundation. But extraction alone doesn't tell you what's happening. On top of the deterministic data layer, we run domain-specific skill modules — 41 rules across six categories:
Ownership signals (11 rules). Accumulation streaks, large step-ups, exits below 5%, activist escalations, pace acceleration, purpose drift, coordinated accumulation. Each has explicit thresholds. An accumulation streak fires after 3+ consecutive increases. A large step-up fires at ≥3 percentage points in a single filing. Every threshold is introspectable via the API.
Earnings signals (7 rules). Beats and misses against prior management guidance, forecast revisions, record revenue, margin expansion and contraction. Computed from 決算短信 XBRL, not from press release text.
Health scoring (8 components). Stability, profitability, and cash flow — scored 0-100 with industry adjustment and risk flags. A company with negative operating cash flow for two consecutive years gets flagged. A company with equity ratio below 20% gets flagged. The scoring formula is public.
Capital allocation (4 classes). Returner, Hoarder, Reinvestor, Mixed — classified by how free cash flow is deployed. If >50% of FCF goes to dividends and buybacks, you're a Returner. If cash accumulation dominates, you're a Hoarder. The classification traces to specific line items in the cash flow statement.
Purpose classification (8 categories). Every large shareholding filing includes a stated purpose — pure investment, activist, cross-holding, management participation, tender offer. We classify from the Japanese disclosure text using deterministic pattern matching, not LLM inference.
Conditional probabilities (3 thresholds). Given a filer at 7% with 3 consecutive increases and a "pure investment" purpose, what's the probability they reach 10%? We compute this from all historical trajectories.
Every one of these rules is deterministic. Every threshold is documented. Every output traces to source filings.
Why this matters now
The real competitive threat in 2026 isn't Bloomberg or FactSet. It's Claude with a PDF. Any analyst can upload an EDINET filing to an LLM and get a plausible answer in seconds. The quality is good enough to feel trustworthy. That's exactly what makes it dangerous.
The analyst who uploads one filing gets one answer. They don't see the amendment chain. They don't see the entity that filed under three different name encodings. They don't see the trajectory across nine filings over 21 days. They don't see that the same filer changed their stated purpose from "pure investment" to "activist" two months before the exit accelerated.
Axiora sees all of this. Not because we use a better model. Because we don't use a model for this at all.
The line
AI interprets language. Code retrieves data. They never cross.
This is the line. On one side, a language model reasons about intent, routes queries, and generates natural language. On the other side, typed XBRL elements, deterministic parsers, entity resolution tables, and amendment chain resolvers produce numbers that trace to source filings.
The answer to "What is IAT Partners' stake in Toyo Engineering?" is not approximately 32.7%. It is 32.72%, from S100XNCJ, filed on February 18, 2026. And it was 20.16% three weeks later — across nine filings that a PDF upload would never connect.
That's not a better answer. It's a different category of answer.
Axiora provides structured financial data for 4,000+ Japanese listed companies via REST API and MCP. Read how the trust architecture works or explore the docs.