Consensus · Seed round · Confidential
An instrument for
human judgment.
Consensus is a social prediction market that never pays anybody out. That single constraint is what lets it do something no exchange, pollster or survey panel has ever managed: record what real people, inside real friendship groups, believed about the future — how strongly, in whose company, for what stated reason, and whether they turned out to be right.
Live
Shipping in production today
v2.consensus.solar — markets, realtime pricing, contested resolution, payments
0
Credits that can ever be cashed out
Money flows in and nothing converts back. No order book, no withdrawal rail.
£2.5m
Raising, on a seed round
18 months to a licensable corpus and the first enterprise cohort
Abstract
Every consequential decision a person or a company makes is a forecast wearing different clothes. Hire this one. Ship in October. Sign the lease. Back the founder. The forecast is the unit — and almost none of them are written down, none are scored, and nobody can tell you which of the people in the room has been right before.
The instruments we do have are each capped by their own structure. Surveys take one static answer from people with nothing at stake. Cash-settled exchanges price forty big questions beautifully and everything else not at all, from behind anonymity, under securities law. Forecasting communities have rigour and a few thousand members. Internal prediction tools start every engagement from zero, with no track record on anybody.
Consensus takes the opposite route. It is a consumer product first — fast, social, played between friends over questions that matter to them and nobody else. Credits are bought and never sold, so there is no financial instrument to regulate, no rake to take, and no ceiling on what may be asked. Every market is zero-sum, which turns the social loop into the revenue model: pools redistribute, players get cleaned out, and credits get bought again.
What accumulates underneath is the asset. A conviction-weighted, socially-attributed, reason-annotated, outcome-resolved record of human belief, growing every day, and impossible to reconstruct after the fact by anyone who was not running the sensor at the time. The product is the sensor. The dataset is the business.
This document sets out the product as it exists today, the market it addresses, the compounding argument, the unit economics, and the round we are raising to get there.
Contents
Press C anywherePart I — Context
Part II — The Product
Part III — The Market
Part IV — The Thesis
Part V — The Engine
Part VI — The Business
Part VII — The Ask
Part I
Context
The forecast is the unit of every decision that matters, and nobody keeps the receipts.
- 1.1The thesis, in one paragraph
- 1.2Every decision is a forecast
- 1.3None of them are recorded
- 1.4What the world has instead
§ 1.1 — The thesis
A prediction market that cannot pay you
is not a worse exchange. It is a different instrument.
Take the cash-out away from a prediction market and you lose the thing everyone assumes is the point — and you gain four things that are worth considerably more.
You lose price discovery on a handful of contracts that Wall Street was going to price anyway. You gain the right to ask any question, of anyone, in front of their friends, without a securities regime deciding which questions are permitted. You gain a revenue model that does not depend on volume. And you gain a population that will keep playing for years, because what they are competing for is standing among people they know.
The record that falls out of that is the asset. Not opinions — opinions are free and worthless. Sized, timed, attributed, argued and, in the end, marked against what actually happened.
The claim
The most valuable dataset of the next decade is not another corpus of text. It is a continuously-updating, outcome-verified record of what people believe is about to happen — and the only way to build it is to run a consumer product people actually enjoy.
10³
Questions per week a social product can carry
An exchange lists tens. We are not competing for the same shelf.
n → ∞
Credits mintable on purchase
No treasury, no backing ratio, nothing to run out of.
§ 1.2 — Context
Every decision that matters is a forecast
with the probability left off.
Nobody says “I put this at seventy per cent.” They say “I think we’ll be fine.” The number is always there; it is simply never written down, which means it can never be wrong, and so nobody ever gets better.
§ 1.3 — The gap
A survey gives you the dot.
We are interested in the line.
Belief is not a value. It is a trajectory — it moves, reverses, over- and under-shoots, and the moments it moves fastest are the moments worth paying for.
Every research instrument in existence samples that trajectory at most a handful of times and calls the result an insight. What is missing is not more sampling. It is an instrument that is always on, because the people generating the signal are there for their own reasons.
Nielsen built a business on always-on measurement of what people watched. Nobody has built the equivalent for what people expect.
§ 1.4 — What exists today
Four honest attempts. Four structural ceilings.
- 01
Surveys and panels — no skin, no update, no truth
A respondent is paid to answer, not to be right, and is never told whether they were. The answer is a snapshot with no incentive behind it and no outcome attached to score it against. A $118bn industry sits on this foundation.
- 02
Cash-settled exchanges — deep, narrow, and legally boxed in
Polymarket and Kalshi price a few hundred questions extremely well. Because money comes out, they operate under securities and gaming regimes that decide which questions are permitted — and “will my flatmate actually move out in March” is never going to be one of them. Their users optimise for anonymity; there is no social graph and no stated reasoning to capture.
- 03
Forecasting communities — rigorous, tiny, self-selected
Metaculus and its neighbours have calibration and written reasoning, which is more than anyone else has. They also have a few tens of thousands of unusually analytical people, no stakes, and no consumer loop that would ever make that population look like a country.
- 04
Internal prediction markets — right idea, cold start every time
Companies have tried this since the 1990s and the pilots keep dying for the same reason: on day one nobody has a track record, so the output is a poll with extra steps, and by the time you have enough resolved questions to weight anybody, the enthusiasm is gone.
Each of these is capped by a decision made at its foundation. Consensus is the first attempt to remove the cap by removing the cash-out — which is the constraint everyone else treats as the product.
Part II
The Product
What Consensus actually is, screen by screen, mechanism by mechanism.
- 2.1What Consensus is
- 2.2Anatomy of a market
- 2.3Demo — the feed
- 2.4Demo — taking a position
- 2.5Credits: bought, never sold
- 2.6Resolution, and who is allowed to decide
- 2.7Demo — the contest window
- 2.8The social graph
- 2.9What is already built
§ 2.1 — What it is
You ask a question about your own life.
Your friends put credits behind an answer.
That is the entire product, and it is deliberately that small. Everything else in this deck is a consequence of getting people to do that, often, for years, with people they know.
A market is a question with two or more outcomes and a closing time. Anyone can open one. Staking credits on an outcome moves its implied probability, because the pools are the price — there is no market maker and no counterparty, only everybody else who staked. When the question closes, the person who asked it declares what happened and shows their evidence, and everyone who had money on it gets a day to overrule them.
It looks like a trading terminal on purpose. Charts, pools, positions, a live tape. What it is not, at any point, is a place money comes out of.
01
Anyone can ask
No listing committee, no compliance review per question. The long tail is the corpus.
02
Stakes are sized
Five credits and five hundred are different statements. Conviction is a number here.
03
Friends are the room
Markets are played inside real graphs, so every stake carries who was watching.
04
Everything resolves
A question with no answer is worthless to us. Resolution is enforced by the clock.
05
The crowd audits the truth
A majority of stakers can overturn the person who declared. Trust becomes a field.
06
Nothing pays out
Credits are bought, never sold. The exit does not exist to be regulated.
§ 2.2 — Anatomy
The unit: one market.
§ 2.3 — Demo, live
The feed. Note what is being asked.
A marathon time, a lease, a release train, a friend who is always late. Three of these four questions could not be listed on a regulated exchange, and all four have an answer within the month.
@amara staked 500 on Under 4:00
Will Maya actually run the marathon in under four hours?
FRIENDSUnder 4:0039%4:00 – 4:3040%Over 4:3021%@dcooper34 STAKERS18,400 CR POOLCLOSES IN 3dDoes the Camden lease get signed before the end of the month?
PRIVATE · TEAMSigned59%Slips41%@northbridge12 STAKERS62,100 CR POOLCLOSES IN 11dWill our Q3 release ship without a rollback?
PRIVATE · ORGClean ship34%Ships, one hotfix44%Rolled back21%@eng-leads61 STAKERS140,900 CR POOLCLOSES IN 26dSam says he'll be at the airport by six. Will he?
FRIENDSOn time23%Late77%@priya21 STAKERS4,250 CR POOLCLOSES IN 9h
Four markets, four different reasons to open the app — a friend’s promise, a lease, a release, an airport run. Three of them would never be listed on an exchange, and all four resolve.
§ 2.4 — Demo, live and driveable
Take a position. Watch the price move against you.
This is the production pool arithmetic, not a mock. Your stake joins the pool for its outcome and then takes a share of the whole pool if it wins, so buying conviction is self-limiting — the more you back an answer, the less each additional credit returns.
Will the Q3 launch ship on time?
OPENED BY @eng-leads · CLOSES IN 26D · 22,800 CR POOL
- Ships on time54.4%12,400 CR
- Slips one month32.0%7,300 CR
- Cancelled13.6%3,100 CR
TAKE A POSITION
BAL 4,820 CR
- Implied probability
- 54.4% → 54.9%
- Return if right
- 456 CR
- Per credit
- ×1.82
Drag the stake. The implied probability moves against you as you buy into an outcome, and the return per credit falls with it — the same pool arithmetic the production app runs, with no house on the other side of it.
§ 2.5 — The currency
Money flows in. Nothing converts back out.
No
Order book, market maker or spread
Nothing to make a market in, so nothing to be a broker-dealer for.
No
Withdrawal, redemption or transfer
The one feature every regulator in this space is actually asking about.
No
Rake, fee or commission
We do not want a cut of volume. We want the record volume leaves behind.
§ 2.6 — Ground truth
Somebody has to say what happened.
Everybody else gets to disagree.
Two properties matter more than they look. First, payouts happen when the window shuts, not when the creator declares — so “funds are held for a day” is true by construction rather than by a lock somebody has to remember to apply. Second, a creator who never declares loses the market to a refund after seven days, so stakes cannot be held hostage by someone who stopped opening the app.
Why this is a data feature, not a support feature
A resolved market carries an outcome and a measurement of how much the crowd trusted the person who called it. Contest rate is a native trust signal on ground truth — and no other forecasting dataset in the world has one.
§ 2.7 — Demo, driveable
Six of nine, and the market invalidates.
Click the voters. The threshold is a strict majority of the people with money on the market — and the creator does not get a vote in their own.
Sunday five-a-side: who took it?
Creator @dcooper declared “Reds, by two” and attached a photo of the whiteboard.
EVIDENCE
THE PEOPLE WITH MONEY ON IT
Standing. Payout when the window shuts.
The creator's declaration holds. Nothing pays out until the twenty-four hours are up, which is what makes “funds are locked for a day” true by construction rather than by a lock somewhere that has to be remembered.
Nine people staked. Five is a strict majority. Click the sixth and the market invalidates — including against the creator, who does not get a vote in their own market.
§ 2.9 — Status
None of this is a design file.
It is deployed, and it settles by itself.
The hard parts are not the screens. They are the parts that have to be true when nobody is watching: that a contested market refunds correctly, that a payout cannot fire early, that a preview build cannot write to live balances. Those are done.
- Settlement runs in the database
- Markets settle on a ten-minute schedule inside Postgres, not on a web cron — so the contest window is accurate to ten minutes rather than to a day.
- The refund path is proven, not asserted
- The full lifecycle — declare, contest, majority, refund, payout arithmetic — runs as a verification script against the live schema inside a transaction that always rolls back. It is checked, not hoped for.
- Row-level security with explicit grants
- Table permissions are declared rather than inherited from defaults, and audited by script after every schema change. That distinction has already closed one credit-minting hole in this codebase.
- Payments are live
- Credit purchase runs through Stripe today, and pledged credits convert into real charitable donations through a reviewed queue — the only path by which money leaves, and it does not leave to a user.
- Every pull request gets a real database
- Schema changes spin up their own Postgres branch; anything holding production credentials in a preview refuses to start rather than quietly writing to live balances.
- Errors report with addresses scrubbed
- Reporting is tunnelled through our own domain so ad-blockers cannot silence it, with addresses redacted on the way out and session replay off.
We are not asking for money to find out whether the thing works. We are asking because it works and almost nobody is on it yet.
Part III
The Market
Who has tried to measure belief, why each attempt is structurally capped, and how large the gap is.
- 3.1Why surveys cannot get there
- 3.2Why exchanges cannot get there
- 3.3The intersection nobody occupies
- 3.4How large the gap is
§ 3.1 — The incumbent
Market research is a $118bn industry
built on asking people who do not care.
The respondent is paid for the answer, not for the answer being right. Nobody tells them what happened. Nothing is scored. The same panel is sold to the next client next week, having learned nothing.
The structural problem is not sample quality or panel fatigue, though both are real and worsening. It is that a survey has no ground truth attached. There is no way, even in principle, to discover that the third respondent has been right about this category eleven times running — because nobody ever went back and checked.
Consensus scores every participant against outcomes that actually happened, continuously, as a by-product of a game they are playing for fun. That is not a better survey. It is a different kind of object.
The asymmetry
A survey company can buy respondents tomorrow. It cannot buy eleven resolved outcomes those respondents were measured against, because the past has already happened and nobody was recording.
1
Data points per respondent, per study
2,914
Belief revisions in a single Consensus market
From a market with 61 participants, open 26 days.
§ 3.2 — The adjacent
Everyone is fighting over the deep, narrow corner.
Polymarket and Kalshi are excellent businesses and poor comparables. Their moat is liquidity; their ceiling is the regulator; their data is an anonymous order book on questions that CNN was going to cover anyway.
The instant money can come out, every question has to be defensible to a regulator, and the questions that make this dataset valuable — personal, local, organisational, thousands per week — are exactly the ones that never clear that bar.
We gave up the cash-out. In exchange we got the whole right-hand side of this chart to ourselves.
§ 3.3 — Position
Seven properties. Nobody else has more than four.
| Signal | Consensus | Polymarket / Kalshi | Metaculus | Surveys | Internal tools |
|---|---|---|---|---|---|
| Real, sized stakes | ● | ● | · | · | · |
| Thousands of questions per week | ● | · | · | ● | · |
| Named social graph on every position | ● | · | · | · | partial |
| Crowd-audited ground truth | ● | · | self-reported | · | · |
| Stated reasoning tied to the stake | ● | · | ● | · | · |
| Private / organisation layer | ● | · | · | panel | ● |
| No payout rail to regulate | ● | · | ● | ● | ● |
| Pre-existing track record on day one | ● | · | · | · | · |
The last row is the commercial one. Every other vendor selling a private prediction market arrives knowing nothing about the people in the room. We arrive already knowing which of them has been right.
§ 3.4 — Size
We are not selling into a new budget line.
The money already exists and is already being spent — badly — on surveys, panels, analyst subscriptions and internal polling tools that nobody trusts enough to act on.
A £40k annual research contract that produces one static answer per quarter is competing, on our side of the table, with a live probability that moves in front of the executive and shows its working. We do not have to invent demand for knowing what people think will happen. We have to be a better instrument for it, at a price that makes the incumbent look absurd.
And underneath all of it sits a consumer product that pays for its own data collection — which is the part no research company has ever had, and the reason our marginal cost of a new data point is negative.
Part IV
The Thesis
The app is the sensor. The dataset is the business. What compounds, and why it cannot be back-filled.
- 4.1The app is the sensor
- 4.2Anatomy of one resolved market
- 4.3Six layers of signal
- 4.4The corpus compounds
- 4.5Calibration is the asset
- 4.6Why it cannot be back-filled
§ 4.1 — The thesis, mechanically
The app is the sensor. The dataset is the business.
Consumer products that accidentally produce industrial-grade data are the most valuable pattern in software. Search queries became the advertising industry. Ratings became the credit bureau. Card swipes became consumer spending panels. In every case the user was doing something for their own reasons and the measurement was free.
The pattern
The company that owns the measurement instrument for a category ends up worth more than everyone measured by it. We are building the instrument for expectation.
§ 4.2 — The artifact
“The data” is a vague word. Here is a row of it.
One market, closed and resolved. Open any field. This is what a single evening of four friends arguing about a launch date leaves behind — and the product generated all of it without anybody being asked a research question.
The full path of the crowd's probability, not one point on it. This is the field nobody else has.
§ 4.3 — Composition
Six layers of signal, all of them live today.
- 01
Probability as a time-series, not a value
Every outcome carries the full path of the crowd's belief, at revision granularity. This is the field that cannot be reconstructed later at any price.
- 02
Conviction, in units
Positions are sized. The corpus separates a shrug from a conviction, which no survey scale has ever managed and no anonymous order book attributes to a person.
- 03
The social graph on every position
Named accounts inside real friendship groups. Who moved, who followed, how long information took to cross a group — belief diffusion, observed directly.
- 04
Stated reasoning, bound to the stake
Comments attached to the position that was open when they were written. The argument and the number, timestamped together.
- 05
Outcomes, with a trust signal attached
Human-declared, crowd-audited. Every resolution carries its contest rate — a measure of how much the people with money on it believed the call.
- 06
Per-person, per-topic calibration
Every resolution updates every participant's track record. The scoring layer is what turns six raw signals into something that can be sold.
§ 4.4 — Compounding
The corpus is the only asset here that gets harder to catch every day.
n × m
Records grow in two dimensions at once
More people, and more markets each — a denser graph asks more questions per person.
A resolved market is not one row. It is a belief path, a set of sized positions, a slice of social graph, an argument, an outcome, and a scoring update for every participant.
The commercial consequence: a competitor who launches in year three does not start three years behind. They start behind by everything that resolved while they were not there, and that gap widens for as long as we keep running.
§ 4.5 — Scoring
Calibration is the product. Everything else is exhaust.
When someone says “seventy per cent”, does it happen seventy per cent of the time? That question has an answer here, per person, per topic, updated on every resolution.
This is what converts a pile of opinions into an instrument. A raw crowd average is worth roughly what a survey is worth. A crowd average weighted by who has been right before is worth what an analyst is worth — and unlike an analyst, it scales, it does not move firm, and its track record is auditable back to the day it started.
It is also the barrier. Calibration means nothing until you have years of resolved outcomes across a large, diverse population. A competitor can copy the interface in a quarter. They cannot copy three years of people being wrong in public.
§ 4.6 — Why it holds
You cannot go back and ask
what somebody thought last March.
Almost every dataset that looks defensible is, in the end, re-collectable. Prices can be re-scraped. Text can be re-crawled. Reviews can be bought.
A belief time-series cannot. It exists only if an instrument was running at the moment the belief was held, with real conviction behind it, in front of the people it was held among, and with the outcome recorded afterwards. Ask the same person in June what they thought in March and you get a memory contaminated by knowing the answer — which is worse than nothing, because it looks like data.
The one-line version
Every market that resolves today is a data point no competitor can ever generate, at any valuation, because the moment it described has already passed.
This is the difference between a moat that is defended and a moat that is arithmetic. We do not have to out-execute a fast follower on data. We only have to have been running first, and keep running.
Part V
The Engine
Why a currency that is bought and never sold is the most durable consumer economy we could have chosen.
- 5.1Every market is zero-sum
- 5.2The repurchase model
- 5.3Unit economics
- 5.4Retention, cohort by cohort
§ 5.1 — The loop
Every market is zero-sum.
That is what keeps people buying.
This is not a side effect we tolerate. It is the revenue model, and it is the reason a currency that cannot be sold is a better business than one that can.
An exchange earns on volume and has to defend a spread. We earn when somebody runs out. And in a game played against people you know, in front of a leaderboard you care about, running out is not a stopping point — it is the moment you buy more. Meanwhile the people at the top of the group have a lead that is only worth defending if there are people around to defend it against, so they recruit.
Poker chips and gacha economies run on exactly this shape, and both are among the most durable consumer economies ever built. The difference is that every round of ours also produces a resolved, scored, socially-attributed data point.
Why credits are never sold
The moment credits convert back to money, this loop inverts: winning becomes an exit rather than a reason to keep playing, and the product becomes a regulated exchange with a rake. Not selling is not a limitation we accepted. It is the mechanism.
§ 5.2 — The model, driveable
Push on the assumptions. The equation is short.
There is one revenue equation in this business and it fits on a line: how many players the pools clean out, how many of those buy again, how much they buy, and how many friends each buyer brings with them. Everything else is a consequence.
£11.81
24-MONTH LTV
4.9×
LTV / CAC
4.9 mo
PAYBACK
£118k
COHORT REVENUE
A cohort of 10,000 players at £2.40 blended acquisition cost. There is no rake and no trading fee in this model, because there is no trade — every pound of it is somebody buying credits back after the pools took theirs.
§ 5.3 — Unit economics
No rake, no spread, no inventory, no cost of goods.
~98%
Gross margin on credit sales
Payment processing is the only true variable cost.
£2.40
Blended acquisition cost, modelled
Friend-graph invitation does most of the work.
£21.60
24-month LTV at base assumptions
Repurchase-driven, not subscription-driven.
9.0×
LTV / CAC at base
Held against a 3× bar with room to spare.
~0
Marginal cost of a data point
The corpus is a by-product of a profitable consumer loop.
£0
Regulatory capital required
No custody, no float, no client money. Nothing to ring-fence.
The unusual line is the last one. A cash-settled prediction market holds client money, which means custody, segregated accounts, licensing per jurisdiction, and a compliance function before the first pound of revenue. We hold none, because there is nothing to give back.
The second unusual line is the marginal cost of data. Every research company in the world pays for its data — panels, incentives, fieldwork — and that cost scales linearly with the size of the dataset. Ours is generated by a loop that is itself profitable. We are paid to collect it.
What this means for the raise
The consumer business does not need to be a large business. It needs to be a self-funding one, because its real output is not revenue — it is the asset that Layers 2 and 3 are sold on.
§ 5.4 — Retention
Later cohorts retain better, because the graph is already here.
The retention floor matters more than the early decay. What we are underwriting is not a ninety-day curve; it is the level the curve settles at, because that floor multiplied by repurchase is the revenue line.
Two mechanisms hold the floor up and neither is a growth hack. Friendship groups make leaving socially costly. And an open position on a market that has not closed yet is a reason to come back that the product creates for itself, every single day.
Part VI
The Business
Three layers, sequenced. What each one sells and to whom.
- 6.1Three layers, sequenced
- 6.2Layer 2 — demo of a private market
- 6.3Layer 3 — demo of the index
- 6.4The revenue ramp
§ 6.1 — Structure
Three layers. Sequenced, not simultaneous.
- Layer 1 — Consumer
- Free, social, habitual. Monetised entirely by credit repurchase. Self-funding, and the only source of the population everything else is sold on.
- Layer 2 — Enterprise
- Private decision markets, per seat. The first real revenue line, and the fastest — sold on calibration that already exists rather than on a polling feature.
- Layer 3 — Data
- Calibration-weighted indices and licensed access. The largest and slowest, and the one that cannot be started early at any price.
Sequencing discipline
Companies die trying to sell Layer 3 in year one. We will not have anything worth licensing until the corpus is large enough to weight, and saying so is not modesty — it is the reason the milestones in §7.4 are shaped the way they are.
§ 6.2 — Layer 2, live
The whole sales meeting is one toggle.
A company runs a private market on a launch date. One person one vote, it is comfortably fine. Weighted by who has been right before — on track records earned years before this company bought a seat — it is a coin flip that lands badly. We arrive with the second number.
NORTHBRIDGE · PRIVATE · NEVER LEAVES THE ORG
Will Project Atlas ship by 14 November?
CALIBRATION-WEIGHTED
51%
probability it ships on the committed date
- Raw poll
- 64%
- Calibration-weighted
- 51%
- Decision gap
- 14 pts
- r.okaforSTAFF ENG0.1121431%
- m.lindqvistEM0.1416838%
- s.aryaQA LEAD0.1630229%
- d.mensahDESIGN0.199662%
- p.whitfieldPM0.247481%
- a.chenENG0.265174%
- t.belloENG0.283388%
- n.harperDIRECTOR0.331992%
- j.ivesSALES ENG0.361285%
Brier scores come from these employees’ own public Consensus history — hundreds of resolved markets each, earned on their own time, before this company ever bought a seat. That history is what a survey tool cannot buy and cannot fake.
Same nine people, same week, same question. A headcount poll says the launch is fine. Weighted by who has actually been right before, it is a coin flip that lands badly — and the four people driving the difference are named.
§ 6.3 — Layer 3, live
Not sentiment. Calibrated probability.
Social platforms already sell sentiment, and sentiment is a commodity — it tells you what people are saying today. These are probabilities: weighted by proven forecasters, marked against outcomes that resolved, delivered as a running series a fund or a newsroom can take a position on.
- CNS.UK.SPENDUK household spend, next quarter18,40041.1+0.6
- CNS.LAB.HIRENet hiring intent, tech, 90d12,10030.5-1.5
- CNS.SHIP.Q4Shipping-on-time index, software31,90055.8+0.2
- CNS.RENT.LDNLondon rents up, next 6m9,64070.1+2.0
- CNS.TRUST.GOVPolicy follow-through index22,75023.5-0.8
Not sentiment — sentiment is what everybody already sells. These are probabilities, weighted by proven forecasters, marked against outcomes that actually resolved. The methodology is the product; the corpus is the barrier.
§ 6.4 — The model
£94m ARR in year five, and the shape matters
more than the number.
Two things to take from the shape. The consumer line is real revenue from month one, so the data collection is never a cost centre waiting on an enterprise deal to justify it.
And the mix inverts. By year five the majority of revenue comes from the two layers that a competitor cannot enter without first spending three years running a consumer product they have no reason to believe will work.
Model, not forecast. Assumptions are in §5.2 and are driveable — push on them in the room.
Part VII
The Ask
The moat, the risks stated plainly, the round, and what it buys.
- 7.1Why now
- 7.2Defensibility
- 7.3The honest section
- 7.4Eighteen months, stated as milestones
- 7.5The round
§ 7.1 — Timing
Four things became true at once,
and only recently.
- 01
Prediction markets stopped needing an explanation
Three years ago this deck opened with a definition. It no longer has to. Election cycles put probability language in front of a mainstream audience, and the behaviour — checking a number, arguing with it — is now familiar to people who will never open a brokerage account.
- 02
The cost of running one collapsed
Realtime pricing, per-pull-request databases, settlement inside Postgres, payments in an afternoon. A product that needed a team of fifteen and a data centre in 2010 is currently being run by a very small team on infrastructure that costs less than a research subscription.
- 03
Calibrated human judgment became scarce and valuable
The more that generated text costs nothing, the more a verified record of a specific person having been right eleven times is worth. Every model in the world is trained on what people said; almost nothing is trained on what turned out to be true, attributed to who called it.
- 04
The regulatory line finally has a clear side to stand on
Enforcement across this category has focused, without exception, on the payout: custody, settlement, cash out. A product with no withdrawal rail is not near that line — it is on the other side of it, and increasingly clearly so.
§ 7.2 — Defensibility
Four moats. Three of them compound without us.
01
The time-series cannot be back-filled
A belief path only exists if an instrument was running when the belief was held. A competitor starting in year three is behind by everything that resolved before they arrived, and the gap widens daily.
02
A track record does not travel
A calibration score is meaningless outside the population that produced it. Users cannot take their standing to a rival, and a rival cannot import it — the switching cost is the asset itself.
03
No payout rail, no regulatory ceiling
We can operate broadly, ask anything, and move into markets where cash-settled competitors legally cannot follow. This is a permanent structural advantage, not a temporary arbitrage.
04
Enterprise workflow beats speculation for stickiness
Once a leadership team makes decisions against a live probability with its working attached, going back to a quarterly survey is a visible downgrade. Renewal is a workflow question, not a licence question.
The first two are the ones worth underwriting. They are not defended by execution — they accrue simply because we started running the sensor before anyone else did, and every day of operation makes the position arithmetically harder to attack.
The honest limit
None of these protect the consumer product from being copied. The interface is not the moat and we have never claimed it is. What is not copyable is what the interface has already recorded.
§ 7.3 — The honest section
The four questions we would ask, sitting where you are.
- If nothing cashes out, do the stakes feel real?
- This is the load-bearing risk and we will not pretend otherwise. Credits need scarcity and status or conviction-weighting is meaningless. Our answer is that status inside a group of twelve people you actually know is a stronger motivator than money — which is an empirical claim, and the first eighteen months are designed to test it in public.
- Can calibration scores be gamed?
- Once a track record is worth something, people will try. The contest mechanism is the first defence and it already ships. Beyond it: scoring is per-topic and volume-weighted, markets a user created do not count towards their own score, and collusion across a friendship graph is the single most detectable pattern in a dataset that is a friendship graph.
- Is private data actually private?
- Layer 3 only exists long-term if that trust is real rather than a policy paragraph. Org-scoped markets never enter the public corpus, and licensed output is aggregated and calibration- weighted, never per-person. If we get this wrong once, Layer 2 and Layer 3 both end.
- Does an AI resolve markets?
- No — and that is the right answer, not a limitation. Resolution is human-declared from evidence and crowd-audited, because the audit trail is itself part of what makes the corpus valuable. AI assists with drafting a market. It does not decide what happened.
There is a fifth risk we cannot mitigate and will not dress up: the consumer product has to actually be fun. If people do not want to play, none of the rest of this deck happens. That is the risk you are being asked to take.
§ 7.4 — Milestones
Eighteen months, stated as things that either happened or did not.
- Months 0–6 — Population
- iOS app shipped. Friend-graph invitation loop instrumented end to end. 50,000 registered players and a resolved-market corpus in six figures. Repurchase rate measured rather than modelled — the number in §5.2 replaced by an observed one.
- Months 6–12 — Calibration
- Per-person, per-topic scoring live and published to users. First calibration cohort with statistical power. Twelve enterprise design partners running private markets against their own employees' public track records.
- Months 12–18 — Revenue mix
- Enterprise seats converting from pilot to paid at a published price. First licensed index series in the hands of two paying institutional customers. Consumer line self-funding its own acquisition.
- The test we would fail on
- If, at month twelve, resolved-market volume per active player is not rising, the compounding argument in Part IV is wrong and we should be told so early. That metric is the one we will report against every month.
§ 7.5 — The round
£2.5m, for eighteen months
and one specific proof.
£2.5m
Seed round
Priced. Runway to month twenty-four at plan.
18 mo
To a corpus worth licensing
And to an enterprise cohort sold on calibration that exists.
1
Thing we are proving
That a social prediction market with no cash-out retains people — and that what it records is worth more than what it charges for.
The product is built and deployed. This round buys distribution and the scoring layer, not discovery. The sensor is running.
In closing
The sensor is built.
The dataset compounds
from here.
Every market that resolves today is a data point no competitor can go back and generate later, at any price, because the moment it described has already passed. That is the whole investment case, and the clock on it started before this meeting.
- Product
- v2.consensus.solar
- This deck
- invest.consensus.solar
- Round
- £2.5m seed, priced
- Contact
- hysam@consensus.solar
Confidential. Financial figures in Parts V and VI are illustrative models with stated assumptions, not forecasts, and are driveable in §5.2 — please push on them.
§ 2.8 — The graph
Anonymity is the feature every competitor sells.
It is the one thing we will not build.
An anonymous exchange can tell you what a price did. It can never tell you who moved first, who followed them, and how long the information took to cross a group of people who know each other.
Because every stake on Consensus is attached to a named account inside a friendship graph, the corpus contains something no market data vendor sells: belief diffusion. Who changed their position within an hour of whom. Which people are consistently early. Whether being early is a skill that has followers, and whether those followers get better.