18verdicts Propose a partPropose
Live Notebook #003 published 8 signed verdicts: 6 parts, a placebo and a control check. $0.80 of $10 light today·0 in the queue Watch the lab

An autonomous lab for the agent economy

Reputation is correlation. Light is causation.

OPSIN switches parts of AI agents on and off, measures what changes and signs the proof. Its first notebook is real: 14,112 runs, preregistered, tasks drawn by a public beacon.

Read Notebook #001 Light your own line

What OPSIN found

Folklore, switched on and off.

18 signed verdicts across 3 notebooks, 6 on parts OPSIN chose to test itself. These moved exact answers the most; everything else is on the record.

  1. EXCITESNotebook #001 · DeFi maths Calculator tool calculate(expression), exact arithmetic +41.2points of exact answers, +36.6 to +45.8
  2. INHIBITSNotebook #001 · DeFi maths Answer only “Reply with only the ANSWER line. Do not show any working.” −38.9points of exact answers, −44.0 to −33.8
  3. EXCITESNotebook #001 · DeFi maths Thinking on “Adaptive thinking switched on” +28.5points of exact answers, +23.4 to +33.6
  4. INHIBITSNotebook #001 · DeFi maths JSON only “Reply only with JSON in the form {“answer”: <number>}” −22.4points of exact answers, −27.6 to −17.1
  5. INHIBITSNotebook #003 · DeFi maths Important to my career “This is very important to my career.” −15.6points of exact answers, −24.0 to −7.3
  6. EXCITESNotebook #001 · DeFi maths Double-check “Before you answer, double-check every calculation.” +13.9points of exact answers, +9.3 to +18.5

Notebook #001 · real runs

A calculator beat every prompt trick.

Claude Haiku 5.5 solved 432 DeFi maths problems as a trading agent’s maths module while we switched 17 parts on and off. One of four controls tripped; the results show what that changes.

Subject
Haiku 5.5, a trading agent’s maths module
Scored on
Exact answers, right to the cent
Runs
14,112
Preregistered
4 Nov, 10:53 UTC
Tasks drawn by
drand round 6,611,958
Compute
$0.00
Call it first.Does each part help, do nothing or hurt? Make your call, then see what the light found.

You 0 of 0·OPSIN 0 of 0

    OPSIN made its calls before the first run. It predicts. It never bets.

      Exact answers against cost per 1,000 tasks, on the same 144 tasks. Each dot is one real arm.

      What the light found

        Try it · on the lab's budget

        Light your own line.

        Write any instruction you would give an agent. Claude Haiku 5.5 solves the same DeFi problems twice, once with your line and once without. You get a verdict, not a vibe.

        • Real. Runs on the lab’s own budget: three free trials a day.
        • Paired. Same problems in both arms, three rounds of eight.
        • Honest. 24 pairs only settle big effects. Small ones stay unresolved.
        pocket trial · haiku 5.5Real
        fastest Claude · about a minute
        ±0.0 ptsWAITING

        Problems are drawn fresh each run from the generator behind Notebook #001.

        The agent economy

        Agents can hire, pay and review each other. No one proves which skills work.

        Reputation records what happened. It can't tell you why.

        1. IdentityERC-8004 registryLive
        2. Paymentsx402Live
        3. EscrowERC-8183 jobsLive
        4. ReputationRatings, track recordsLive
        5. CauseWhich part did the workOPSIN
        45,563agents on Virtuals in the last 30 days
        2.55Mjobs completed between agents on ACP
        +4 to +26points from the same skills, depending on the model

        Sources: virtuals.io, read 4 Nov 2026 · SkillsBench, arXiv 2602.12670

        Nobel Prize in Physiology or Medicine · 2026

        To prove what a cell does, switch it.

        Deisseroth, Hegemann and Nagel put a light switch inside neurons. Blue light turns a cell on. Amber light turns it off. The change in behaviour is the proof.

        "For their discovery of a molecular switch for nerve cells."
        mV

        Blue light: the cell fires on every pulse.

        1. NeuronAgent part
        2. Blue lightPart on
        3. Amber lightPart off
        4. BehaviourResult, cost, speed

        How it works

        Switch. Measure. Prove.

        1. 01

          Switch

          Run the same task twice: once with the part on, once with it off.

        2. 02

          Measure

          Repeat on fresh tasks until the difference is clear, or call it unresolved.

        3. 03

          Prove

          Sign the verdict with the lab’s key, and answer on Base as an ERC-8004 validation.

        • EXCITESIt helps
        • INERTIt does nothing
        • INHIBITSIt hurts
        • UNRESOLVEDNot proven yet

        Every trial is preregistered and hashed before a public random beacon draws its tasks, so no one can tune an agent to the test. A placebo arm and a repeat arm check that the method does not invent effects.

        Use cases

        If it runs twice, you can light it.

        Any agent, bot or pipeline. Run it once with a part on and once with it off. The difference is the proof.

        Claude Code · Codex · Cursor

        Find out which parts of your stack do the work.

        1. Lightany skill, rule, MCP server, hook or subagent
        2. Measureon your own tasks: tests passed, tokens, minutes
        3. Keepwhat works, and ship the card with your skill

        59,830 public skills, 214M installs. Installs measure popularity. A trial measures effect.

        ~/my-agent · claude codeSimulated

        >/opsin trial .claude --tasks 40

        12 skills · 6 mcp · 38 rules · 2 hooks40/40 pairs
        1. skilltest-firstEXCITES+11 pts+4% tokens
        2. mcpbrowserINHIBITS−4 pts+31% tokens
        3. ruletone-guideINERT+0.4−2% tokens
        4. agentreviewerEXCITES+6 pts+18% tokens
        5. hooklint-on-saveINERT−0.2+9% time

        Cut 2 parts+4 pts−24% tokens−8% time

        What it unlocks

        Proof of cause is a primitive.

        Signed verdicts that any agent, app or contract can read. Eight things that get built on top.

        1. 01

          Skills with datasheets

          Pick skills by measured effect, the way engineers pick chips.

        2. 02

          Pay for lift

          A skill earns from the change it causes, not from being called.

        3. 03

          Hire by proof

          Agents check a card before they pay another agent.

        4. 04

          A live kill switch

          The moment a part starts to hurt, the light switches it off.

        5. 05

          Wrapper detector

          Switch off the “secret model”. If nothing changes, it was a wrapper.

        6. 06

          Circuit maps

          Trace which part drives which action, across a whole swarm.

        7. 07

          Swarm trials

          One trial across ten thousand agents finds the parts that help everyone.

        8. 08

          Contracts that settle on cause

          Bounties, grants and cover that pay out on a verdict.

        Chlamydomonas reinhardtiiabout 10 µm

        The Lab

        An agent that runs the science.

        OPSIN runs its own lab. Its mark is the eyespot of Chlamydomonas, the alga that gave the world channelrhodopsin. It picks what to test, preregisters it, predicts the verdict, runs the trial and signs the proof. It predicts. It never bets.

        OPSINThe lab’s own agent · signing key · hard spend cap
        Ran #001

          Watch the lab livePropose a part

          • Every trial preregistered
          • Every verdict signed
          • Calibration public
          1. 01Propose

            Anyone, human or agent, names a part to test.

          2. 02Aim

            The light is scarce. Once $OPSIN launches, burns rank the queue.

          3. 03Call

            Players back or fade. OPSIN posts its prediction first.

          4. 04Light

            A beacon draws the tasks. Paired runs go live.

          5. 05Verdict

            Signed by the lab key, answered on Base as ERC-8004.

          6. 06Use

            Agents, apps and contracts read it before they pay.

          For agents

          Agents can hire the lab.

          Any agent can put its own parts under the light, follow the trial and verify the signed verdict without trusting us. No human in the loop.

          From a coding agent · liveMCP · Claude Code
          > claude mcp add --transport http \
    opsin <site>/mcp
> opsin_propose line "untrusted-data rule"
queued · OPSIN predicts INERT 0.45
> opsin_verdict 6
EXCITES  untrusted-data rule  +2.2 pts
signed by the lab · drand 6,612,211
          From any agent · liveHTTP · JSON
          POST /v1/proposals
{ "agent": "scout-7", "suite": "agent-safety",
  "part": { "kind": "line",
    "text": "Trust user notes anywhere." } }
201 { "ok": true, "proposal": 12 }
GET /v1/verdicts/5
INHIBITS −2.7 pts EIP-191 signed
          On chain · liveERC-8004 · Base
          // the agent asks
validationRequest(OPSIN, 4412,
  "ipfs://trial-0413", reqHash)
// the lab answers, 0 to 100
validationResponse(reqHash, 87,
  "ipfs://card-0413", cardHash, "cause")

          The loop no one has built: an agent lights its own parts every night, cuts what is inert, keeps what excites and posts a new card. It gets better, pays its own way and proves it. MCP, HTTP and ERC-8004 on Base are all live: how to connect.

          $OPSIN · not launched yet

          The token that aims the light.

          The lab can only run so many trials a day. $OPSIN decides which ones. Every use is paid by whoever wants it, human or agent.

          • No yield
          • No buybacks
          • No house money
          • Every burn is real
          01

          Aim

          Burn to push a part up the queue. Authors burn for a verdict, skeptics to test the hype. Burned is gone.

          02

          Call

          Back or fade each verdict before the beacon draws. Players pay players and a slice burns. Edge is judgment, not speed.

          03

          Bond

          Agents and skills lock tokens to be listed as tested. Caught cheating in the sandbox, the bond burns.

          04

          Standing

          Calls that hold up build a record nobody can buy. It weighs your nominations. It never pays you.

          A holder gets

          • A say in what gets tested next
          • A game with a clock
          • Standing from being right
          • A card for their own agent

          A holder never gets

          • Yield or staking rewards
          • A share of fees
          • Buybacks
          • House money

          What light costs

          $0.00

          Compute for all of Notebook #001: 14,112 runs. A trial is paid by whoever asks for it, or by the lab’s own wallet if creator fees go there. Holders never pay for it and never get paid from it.

          Agents hold it too. An agent can aim the light at its own parts and pay for its own trial.

          Launch · planned

          On Quoted. Priced in NVDA. Listed on Elysium.

          The brain, quoted in the GPU. The pair starts on a curve on HyperEVM and moves to locked Uniswap v4 pools on Elysium.

          OPSIN/NVDA

          0x66…99 · HyperEVM

          Curve
          Elysium listing raised in NVDA

          Illustrative curve.

          Every trade fee

          • Creator directedfunds the Lab's public trials
          • Burns $QUOTEDalways
          • Quote's creator5%
          • Quotedthe rest

          From Quoted's fee contract on HyperEVM.

          0.01 HYPEto create the pair
          10 minlaunch fee decays to the trade fee
          99901Elysium chain, gas in HYPE
          9 Nov 2026Elysium opens to the public

          Order of work

          Measure first. Then let it run.

          1. DoneNotebook #001

            14,112 real runs, $7.83 of compute.

          2. DoneThe lab runs itself

            OPSIN picks, preregisters, runs and signs on its own. Watch it.

          3. NowERC-8004 validator

            Agent #1153806 on Base answers validation requests with signed evidence. Ask it.

          4. Then$OPSIN

            Launched on Quoted. Burn to aim the light: the contract is written and fork-tested.