Provenance
Talk to the lab
RESEARCH LAB · BACKED BY Y COMBINATOR

Provenance

A research lab building the live economic map.

The map covers every company, fab, product and material in the semiconductor value chain, and what actually moves between them. We build it from the public record and source every claim down to the sentence. Ask it what you would otherwise pay an expert network to tell you. Implement it into your own agents over MCP. When the record runs out, it tells you the answer is not established. It does not guess.

Read the method
133,107 ENTITIES · 1,188,498 SOURCED EDGES · 476,784 CLAIMS · LIVE
every claim carries its source · we abstain when the record is thin · corrections append, the history stands
01 · WHAT THE MAP HOLDS MEASURED 5 AUG 2026
THE QUESTION
What moved, how much, and how do you know

Take any two companies in semis. We want to tell you what moved between them, in what quantity, at what price, under what relationship, over what period, and from which documents. That one question is the whole product. Every answer opens back to the sentence it came from.

476,784 CLAIMS · 99.9% CARRY THEIR SPAN
THE COVERAGE
The entire value chain

EDA and IP, design houses, foundries, memory, OSAT, equipment, subsystems, materials, gases, wafers, substrates, distributors, hyperscalers, end markets and the governments that intervene. Plus the fabs and products that connect them. If it shapes what a name earns, it belongs on the map.

133,107 ENTITIES ON THE MAP
THE DISCIPLINE
We tell you what we do not know

We label numbers measured, derived, assumed or unknown, and we never collapse an unknown into a tidy point estimate. Where the truth is private, you get a sourced range or an honest blank. A model may find a span and map vocabulary. It may never invent a magnitude. We audit for that and quarantine what fails.

NO FABRICATED NUMBERS, EVER
02 · HOW IT STAYS CURRENT REV 2026.08

A map this wide falls out of date the moment you stop reading. Nobody can staff that. So we train our own models to maintain it, on evidence mined from the map itself. They read everything that arrives, every day, and they get sharper as the map grows.

PROV-1-EXTRACT
Reads a 300 page filing and returns typed edges, each pinned to a span. Several models read it separately. A verifier then throws out anything the text does not actually say.
967K documents · 476K claims returned
+76% recall over
any single model
PROV-1-EMBED
Retrieval that speaks semiconductor. General embeddings treat a fab, a process node and a part number as the same kind of noise. We trained this one to tell them apart.
pairs mined from our own graph
+11% over frontier
commercial retrieval
PROV-1-JUDGE
Decides whether a quote really carries the claim attached to it. It runs on every write, so a thin citation never reaches the map in the first place.
judged claim and quote pairs
0.87 on held out
claims
PROV-1-RESOLVE
Collapses the same firm across jurisdictions, transliterations and shell layers. It matches on the neighborhood around a name as well as the name itself.
133K entities · the full alias surface
0 silent merges

One family, built on evidence nobody else holds. We measure every figure here on a frozen eval set we hold back, and we run it again before any model ships. A regression cannot slip through quietly.

03 · HOW IT WORKS

Every number traces back to paper

General models invent supply chains because nobody can check what sits underneath. We build the part you can check. A corpus held on cursors. Claims pinned to spans. Entities that carry their aliases. An answer path that fails closed when the record cannot support the question. We started with semiconductors because the chain is deep, global and unusually well documented. The same machinery travels to other industries, and that comes later.

133,107
ENTITIES RESOLVED
1,188,498
SOURCED EDGES
94,325
DOCUMENTS IN THE GRAPH

We pull from a standing corpus of 1,979,194 documents. It sits on durable cursors, so we can read all of it again every time the models improve.

I
Acquisition
Dozens of live sources sit on durable cursors. US and Asian regulatory filings, monthly revenue disclosures, patent grants, regional trade press, transcripts, prediction markets. We normalize each one the hour it lands and never fetch it twice.
II
Extraction
The panel turns documents into typed claims against a fixed ontology of supply, production, input, ownership and exposure. Each claim points to a span you can reopen.
III
Resolution
Names collapse into entities across jurisdictions and transliterations. Every merge keeps its receipt. When the model is unsure, the pair waits for a human.
IV
Fusion
We fuse evidence at query time and keep every source separate on the way in. Hundreds of thousands of evidence nodes resolve into a posterior for each edge. We cluster wire reprints so one story counts once, and support decays on its own as it ages.
V
Agent interface
An MCP server sits beside the typed REST surface, so you can implement the graph into your own agents as a tool. Ask a question. Walk a supply chain. Pull a metric history. Drill through to the document. It answers with numbers and their sources, and it abstains when the record cannot carry the question.
VI
Back to training
Every judgment the pipeline makes gets kept as training data. Entailment checks, materiality calls, extractions. That is what the next generation of our models trains on, and a better model rereads documents we already hold and finds what the last one missed.
04 · FROM THE LAB DRAFTS · NOT YET PUBLISHED
AUG 2026 · PAPER
The controls that increased China revenue
Thousands of extracted China share claims match filed geographic disclosures to about a point, blind. Then they reverse the consensus read on export controls. A share is not a revenue, and the stockpile year is not a baseline.
DRAFT V3 · IN REVIEW
AUG 2026 · PAPER
The packaging pivot that wasn't
We measured a quarter of a million patents by each firm's own filing mix. TSMC never pivoted. Intel is the only real reallocation. Both obvious ways to run this study return a confident, plausible, wrong answer.
DRAFT · IN REVIEW
JUL 2026 · TRAINING
What a frozen benchmark is for
Four changes shipped together looked like a regression and nearly cost us the recipe. So we ran one lever per arm against a frozen benchmark. Two of them were real gains buried under two mistakes. Split apart and stacked again, they gave us our best retrieval model yet.
ABLATION · ONE LEVER PER ARM