Downstream

seedownstream.com · downstream.sh

Markets price whether a bill passes.
Nobody prices what happens next.

The consequence layer for U.S. law — an open, point-in-time graph of statute and regulation joined to what actually happened next.

Founding memo v0.9 · August 2026 · Confidential

The gap

The market moves on passage day. The economics resolve years later.

Statute
enacted

Priced

Agency
rulemaking

Unpriced

Notice &
comment

Unpriced

APA
litigation

Unpriced

Money
obligated

Unpriced

Enforcement
& behavior

Unpriced

Every link is separately observable in public data. Only the first one is priced.

Post-Loper Bright, rule survival is a genuinely open probabilistic question for the first time since 1984.

The insight

The unit of analysis is the provision — classified by mechanism.

Everyone else treats the bill as the atom. That is why the category is search boxes and vibes.

Delegation“the Secretary shall prescribe…”Days to final rule; P(never finalized)

Authorization“authorized to be appropriated”Authorized → obligated conversion rate

Private right of action“may bring a civil action”Forward filing volume, forum concentration

“What happens if this passes?” is unanswerable. “Given an HHS delegation with a 180-day deadline, what is the distribution of days-to-final-rule?” is a hazard model.

Architecture

Two layers. The open one makes the paid one believable.

Layer 1

The Statutory Graph

Deterministic · Verifiable · Open source · Free forever

Bill § → Code § → CFR part → docket → Treasury account
Every edge citable to a primary source
No model risk. Cannot hallucinate.

Layer 2

The Transition Model

Probabilistic · Calibrated · Proprietary · The business

Hazard rates fit on 30 years of outcomes
Every forecast traces to a Layer 1 graph path
Never a naked point estimate, never a recommendation

Layer 1 is already shipping, and is what an OSS community will genuinely maintain — 50 states, 50 formats, constant drift. Layer 2 needs capital, evals, and reputation. That asymmetry is the strategy.

Why now

Four things changed in twenty-four months.

1

Chevron fell

Rule survival is an open probabilistic question for the first time since 1984. Demand created; nothing built to meet it.

2

Inference got cheap

Per-clause extraction across every bill in a Congress was absurd in 2022. It is now a rounding error.

3

The corpus is tiny

~15k documents a Congress. Four orders of magnitude smaller than litigation.

4

Federal data is disappearing

Guidance is edited and deleted silently, with no version history. This is the urgent one.

The archive point is the urgent one: you cannot retroactively snapshot a page that was quietly deleted.

The hard part

Ground truth is what kills every competitor in this category.

“What happens if this passes” resolves in 3–7 years. No feedback loop, no falsifiability, no reason to trust the output. That is why the space is full of confident nonsense.

How we get it

Freeze the corpus at date T — bitemporal, enforced
Score against what actually happened

Baselines we must beat

The mechanism base rate — the honest null
A frontier model on bill text alone — the critical one

Nothing gets graded until the gates clear. A wrong “silent mandate” claim is fatal for a scorekeeper in a way no amount of later accuracy repairs.

The moat

The data is public. Time and track record are not.

Point-in-time series

eCFR gives you today. Historical reconstruction is lossy and in places impossible. In 24 months this cannot be purchased, only waited for.

~$200/mo to start

The linkage layer

Bill § → Code § → CFR part
→ RIN → docket → case
→ Treasury account

Open-sourced on purpose

The Ledger

Every claim and forecast, timestamped, with its eventual resolution. Append-only, externally anchored, public.

Compounds; cannot be back-filled

We sell the ability to ask a new question. We give away, permanently, whether the old answers were right.

Honest bear case: a frontier model plus our open corpus does 70% of this. The last 30% — calibration, point-in-time discipline, resolved history — is where the value sits. That is a real bet.

Go to market

Funds pay for the timing edge so Congress doesn't pay at all.

BuyerWTPFrictionNote

Policy-risk desks at macro & event-driven fundsVery highLowBuy on a call. No procurement. The wedge.

Corporate government affairsHighMedium“What does this do to us in 3 years” is the job.

AmLaw regulatory practicesHighMediumClient deliverables, and a channel.

Insurers, reinsurers, ratingsHighHighRegulatory risk pricing. Slow, sticky.

Congressional & state staffFree—Mission value, credibility, best annotators.

Journalists, advocacy, academicsFree—Distribution, QA, validation, citations.

The paying side funds the free side, and we say so out loud. That is the triple bottom line structurally rather than rhetorically.

What's running

Two of the ten are live, and the archive has not missed a day since it switched on.

D1 · live

The daily issue

Every document the Federal Register published that day, and the dated obligations it created. New every business day.

A page per day · 68 agency pages

D3 · live

The MCP server

The same record, addressable from any assistant. Five tools, no key, nothing to install.

mcp.downstream.sh · free tier forever

The spine · running

The archive

Fetch, hash, store, and parse only from the stored copy. Every day sealed and chained to the day before.

Cannot be back-filled later

And nothing is published yet: every Consequence Card is unverified, and the publish step refuses while that is true. That is the design, not a backlog item.

What we're building

Ten deliverables. Each has to stand on its own, and force a piece of the pipeline into existence.

THE RECORD

“What did the government do today?”

D1The daily issuelive

D2Diff of the day

D3The MCP serverlive

THE JOINS

Deadlines and dollars

D4The catalyst calendar

D5Cost-estimate explainers

D6Did the money move?

THE GRADED CLAIMS

The map in miniature, and the first forecasts

D7Per-agency mandate scorecards

D8One executable formula

D9The Ledger, v0

D10Digests and personalisation

Two are live; the other eight are specified and not built. You cannot grade an agency while being sloppier than the agency. · True as of 13 August 2026.

Team

Fourth company. Fourth raise. First one where the dataset is the product.

Lucas Dickey

Founder. Three prior companies, three prior raises.

Which companies, which outcomes, and what in them predicts this — one line each.

Why this is buildable solo

Everything in the two previous slides was specified, built, tested and shipped by one person, in public, in weeks.

Adjacency, disclosed

A complementary legal-tech dataset in the immediate family. Papered as an arm's-length commercial licence from day one, and raised before anyone has to ask.

Hiring plan and the first two roles: to fill in.

The upper bound

There is no other domain of comparable consequence where the actors get no feedback on whether it worked.

Doctors have outcomes data. Engineers have failure analysis. Traders have P&L. Legislators have elections — an extremely noisy proxy for whether the law did what it was meant to do.

The comp set is wrong

Not FiscalNote and policy intelligence — a $2–5B category. Verisk, Moody's, MSCI: the authoritative dataset plus the model on top.

The endgame isn't a product

If our provision IDs become how people refer to statutory structure, we're the layer competitors are built on.

This is the mission

Building the feedback loop for lawmaking. It happens to be exactly what the Layer 1 dataset already is.

Downstream · seedownstream.com · downstream.sh

The ask

The scarce resource is the two months, not the proof.

Raising

Amount, instrument, and how much is already committed.

What it buys

Two uninterrupted months on the backtest and the first graded forecasts — the calibration curve that changes both the price and who is willing to lead.

What you can check first

Every page, every hash and the MCP endpoint are public today. The misses will be public on the same terms.

Downstream · seedownstream.com · lucas@lucasdickey.com