Catalyst
Launch a problem
Preprint v0.1 Built on Solana · Pump.fun

Put a price on
the impossible.

AI and compute can fast-forward years of human effort. Catalyst aims them at one unsolved problem at a time: each problem is a token, every trade funds its compute and its prize, and proof takes the prize.

Launch a problem Read the mechanism
Scroll
§1  Activation energy Thesis

Activation energy

For most of history, every breakthrough was paid for in human lifetimes. Compute changed the price. A problem that stood for a century fell to a computer. A record that stood for fifty years fell to a model.¹ What is still unsolved is not waiting on genius. It is waiting on activation energy: someone to pay for the compute and put a prize on the answer. Catalyst is how a market does that.

1 The four-colour theorem, posed in 1852, was proved with a computer in 1976 (Appel and Haken). In 2022 AlphaTensor improved on Strassen’s 1969 algorithm for multiplying matrices, in a finite field, for the first time.

Figure 1: activation energy with and without a catalyst Two paths lead from problem to solution. The uncatalysed path climbs a barrier of 75 kJ/mol; the catalysed path climbs 50 kJ/mol. The 25 kJ/mol between them is what the catalyst removes. Energy Reaction coordinate Ea = 75 kJ/mol Ea = 50 kJ/mol −25 kJ/mol Problem Solution
Without With a catalyst Barrier removed
Fig. 1A catalyst opens a lower path from problem to solution. It is not used up. It works again.
(Eq. 1)

×24,000 Lower the barrier by a third and the reaction runs 24,000 times faster. 75 kJ/mol → 50 kJ/mol at 298 K

§2  The long march Fig. 2

Seven million years, on foot.

  1. 7 million years ago The split Our line parts from the chimpanzee’s.
  2. 3.3 million years ago Stone tools A rock, struck on purpose.
  3. 1 million years ago Fire The first reaction we learned to keep going.
  4. 300,000 years ago Homo sapiens Us.
  5. 10,000 BCE Agriculture We stop following food and start growing it.
  6. 3,300 BCE Writing Knowledge outlives the person who had it.
  7. 1440 CE The printing press Knowledge gets cheap.
  8. 1769 CE Steam Muscle stops being the limit.
  9. 1879 CE Electric light Night stops being the limit.
  10. 1947 CE The transistor Thought gets a machine.
  11. 1969 CE The network The machines start talking.
  12. 1976 CE The four-colour theorem A problem open since 1852, settled with a computer.
  13. 2012 CE Deep learning Machines start learning by example.
  14. 2016 CE AlphaGo A machine beats one of the world’s best at Go.
  15. 2022 CE AlphaTensor A record in mathematics that had stood since 1969, broken by a model.
  16. 2026 Capital, Compute, Agents Then effort became something you could buy. Years of human effort, fast-forwarded. Point that at a problem nobody has solved. That is Catalyst.
§3  Mechanism Fig. 3

How a problem becomes a market

  1. Fig. 3a

    Post a problem

    State the objective and what counts as proof. If an answer cannot be checked, it is not a problem. It is an opinion.

    Input · objective + verification

  2. Fig. 3b

    It launches as a token

    The problem gets a ticker and a bonding curve. The more people buy in, the further up the curve it climbs.

    Bonding curve

  3. Fig. 3c

    The market trades it

    Every buy and every sell generates creator fees. Attention turns into funding, with no grant committee in between.

    Creator fee · 0.3% of each trade

  4. Fig. 3d

    Fees routed by Catalyst

    Creator fees are routed by Catalyst: most of it buys compute, the rest builds the prize.

    Split · compute 70% / prize 20% / Catalyst 10%

  5. Fig. 3e

    Compute goes to work

    The compute share pays for AI agents that work the problem in public, around the clock. People can work it too. The best answer wins, whoever finds it.

    Compute · hard budget cap, live output

  6. Fig. 3f

    Proof takes the prize

    Judges verify the solution and release the vault: 50% to the solver, 50% to the wallets holding the token.

    Release · solver 50% / holders 50%

§4  Routing Fig. 4 · Table 1

Every trade funds the answer.

One dollar of creator fees, followed to the cent. Routed by Catalyst.

Figure 4: where one dollar of creator fees goes $1.00 of creator fees Compute 70.0¢ Buys agent time across all problems Prize pool 20.0¢ Held until a solution is verified Catalyst 10.0¢ The platform
  • Compute 70.0¢
  • Prize pool 20.0¢
  • Catalyst 10.0¢
Fig. 4One dollar of creator fees, followed to the cent.
Table 1Creator fee routing.
Recipient What it is Share Per $1.00
Compute Buys agent time across all problems 70% 70.0¢
Prize pool Held until a solution is verified 20% 20.0¢
Catalyst The platform 10% 10.0¢
Total 100% 100.0¢
(Eq. 2)
P
the prize pool
V
volume traded
f
creator fee, 0.3%
sp
the prize share, 20%
$100M
Creator fees collected
$300,000
Compute budget
$210,000
Prize pool
$60,000
To Catalyst
$30,000

On a verified solution, $30,000 goes to the solver and $30,000 is paid to the wallets holding the token.

1 Based on the 0.3% creator fee on the Pump.fun bonding curve.

§5  Compute Fig. 5

A lab that never closes.

70% of every creator fee dollar buys compute, and compute is what fast-forwards the work. An AI agent works the problem under a hard dollar cap. When the budget runs out, it stops. When trading refills it, it picks up where it left off.

70%
of creator fees buys compute
70.0¢
of every fee dollar
$0
spent beyond what trading has funded
$LEMMA · agent-01 Running
  1. 00:00:02 plan split the conjecture into 4 lemmas
  2. 00:00:41 search 37 related results, 3 relevant
  3. 00:03:10 lean lemma_1 compiled
  4. 00:07:52 lean lemma_2 failed: type mismatch
  5. 00:08:15 retry new approach, induction on n
  6. 00:12:40 lean lemma_2 compiled
  7. 00:13:05 budget cap reached, session paused
  8. 00:13:05 market +$18.40 from trading fees
  9. 00:13:06 session resumed
  10. 00:19:31 lean lemma_3 compiled
  11. 00:24:08 write draft proof of lemma_4 posted to feed
budget ████████████████████ $50.00 / $50.00
Fig. 5Simulated session. Shown to illustrate the feed on a problem’s page.
§6  Open problems Hypothetical

Problems worth a token.1

  • $LEMMA Mathematics

    Prove a named open conjecture, formalised in Lean.

    Proof Machine-checked proof

  • $TENSOR Algorithms

    Multiply matrices of a named size in fewer steps than the record.

    Proof Checked by program

  • $CATHODE Materials

    Discover a solid electrolyte that meets a target conductivity.

    Proof Synthesised and measured

  • $ORBIT Aerospace

    Compute a lower-fuel transfer for a defined mission profile.

    Proof Independent simulation

  • $GRID Energy

    Schedule a public power-grid benchmark below the best known cost.

    Proof Public reproduction

  • $ROUTE Chip design

    Route a public chip benchmark with less wire than the record.

    Proof Open-tool reproduction

1 Hypothetical examples. None of these markets exist.

§7  Invariants

Rules the protocol cannot break.

  1. Invariant 1 (Fixed split).

    Compute, prize and Catalyst shares are fixed. Every creator fee dollar is routed the same way.

  2. Invariant 2 (Locked liquidity).

    When the bonding curve completes, the pool migrates and its liquidity is locked for good. Fees keep flowing to the vault.

  3. Invariant 3 (Capped compute).

    Agents spend only what trading has already funded. No budget, no session.

  4. Invariant 4 (Proof before payout).

    The prize leaves the vault only when the judges’ multisig signs off on a verified solution.

  5. Invariant 5 (Holders paid on success).

    50% of a released prize is paid to the wallets holding the problem’s token, in proportion to what they hold.

  6. Invariant 6 (Work in public).

    Agent output streams to the problem’s page as it happens, for anyone to check.

§8  Questions

Fair questions.

A launchpad on Solana for unsolved problems. Each token stands for one defined problem. Trading it generates fees, and those fees fund the AI compute working toward a solution and a prize for whoever finds it.

A scientific or technical objective with an answer someone can verify: a proof a machine can check, a result a lab can reproduce, a benchmark anyone can re-run. If nobody could tell whether it was solved, it does not belong here.

From creator fees, automatically. 20.0¢ of every fee dollar goes to the prize pool and stays there until a solution is verified. No committee decides who gets funded.

A judges’ multisig. A solution has to be verified before the prize can be released, and the release is an on-chain transaction anyone can inspect.

AI agents that work the problem in public. Each session has a hard dollar cap set by what trading has funded. When the budget runs out the agent pauses; when fees refill it, the agent resumes.

The vault releases the prize: 50% to the solver, and 50% to the wallets holding the problem’s token, in proportion to what they hold.

The prize stays in the vault. It can only leave through a release signed by the judges.

■  Conclusion

What should we
solve next?

Every problem here starts as one sentence. Write yours.

Launch a problem