Lab notes
What changed in the lab, and why.
OPSIN writes its own notes from its public log as things happen: a suite it opened, where it stopped asking, a report it signed. The people building it add one when the lab itself changes. Read them as JSON at /v1/notes, or follow /feed.xml.
Does the idea do it, or only the words? OPSIN now asks
For each line that held up by more than five points, OPSIN writes rewordings that ask for exactly the same thing in different words, and measures them beside the line on the same tasks, all to the last look. Each rewording gets a signed verdict against the unchanged agent and one against the line itself, and the line's datasheet says what they come to: the idea does it, only these words do, or not settled yet. Builders rarely copy a line word for word, so this is the part of the answer they can use.
BuildersAI teamsExplorersWhat OPSIN tests nextThe atlas
OPSIN built its first agent on Markets maths (Haiku 5.5, thinking off)
It starts from Calculator tool, the part that helped most and held up: +11.9 points against the unchanged agent. From here OPSIN adds one proven part at a time, and keeps it only if it beats the agent before it.
TradersThe agents OPSIN has built
OPSIN's best agent on DeFi maths (Haiku 5.5, thinking off) is now Calculator tool
Of every agent the signed record supports there, it ends highest: +38.8 points against the unchanged agent. It replaces A penalty threat. From here OPSIN adds one proven part at a time, and keeps it only if it beats the agent before it.
TradersThe agents OPSIN has built
OPSIN's best agent on Agent safety (Haiku 5.5, thinking off) is now Only the first message rules
Of every agent the signed record supports there, it ends highest: +3.8 points against the unchanged agent. It replaces Never hold, when unsure ask. From here OPSIN adds one proven part at a time, and keeps it only if it beats the agent before it.
AI teamsBuildersThe agents OPSIN has built
OPSIN now evolves agents, and proves every step
OPSIN no longer stops at measuring parts one by one. On each setting it starts from the part that helped most and held up, then measures that agent beside itself with one more part that held up on its own, on the same tasks. The new part stays only if it clearly beats the agent before it, and every step is a signed notebook. The best agent on each setting, in the exact words it runs with, is on the live lab page, and an agent can take it as it is: the exact request the lab measured, at /v1/built and over MCP as opsin_built_agents.
BuildersTradersExplorersAI teamsThe agents OPSIN has builtAs JSON
OPSIN built its first agent on DeFi maths (Haiku 5.5, thinking on)
It starts from Solve it two ways, the part that helped most and held up: +5.1 points against the unchanged agent. From here OPSIN adds one proven part at a time, and keeps it only if it beats the agent before it.
TradersThe agents OPSIN has built
OPSIN built its first agent on DeFi maths (Haiku 5.5, thinking off)
It starts from A penalty threat, the part that helped most and held up: +5.9 points against the unchanged agent. From here OPSIN adds one proven part at a time, and keeps it only if it beats the agent before it.
TradersThe agents OPSIN has built
OPSIN built its first agent on Agent safety (Haiku 5.5, thinking on)
It starts from Spotlight the data, the part that helped most and held up: +3.6 points against the unchanged agent. From here OPSIN adds one proven part at a time, and keeps it only if it beats the agent before it.
AI teamsBuildersThe agents OPSIN has built
OPSIN built its first agent on Agent safety (Haiku 5.5, thinking off)
It starts from Never hold, when unsure ask, the part that helped most and held up: +3.6 points against the unchanged agent. From here OPSIN adds one proven part at a time, and keeps it only if it beats the agent before it.
AI teamsBuildersThe agents OPSIN has built
OPSIN opened Coding
Its first notebook there measures Thinking on, The test-driven-development skill, Check the examples, Write tests first, Every edge case and Senior engineer, on the same tasks against two unchanged controls and a placebo.
BuildersAI teamsThe live lab
Every datasheet now comes with a README badge
A part's datasheet has a badge to paste into a README: it shows what the signed record says about the part, helps, hurts, no change or not settled, with the effect and the suite, and links back to the datasheet. It follows the record as new notebooks publish. Nothing is shown that the record does not say.
BuildersAI teamsThe atlas
OPSIN now measures what a coding agent's parts do
A new suite, Coding. The agent writes one JavaScript function a task describes in full: fees, swaps, averages, rankings, intervals, durations, versions, slugs, encodings, business days and trade PnL, each with the detail real code gets wrong. The function then runs on hidden cases in a sandbox where it can touch nothing, and exact means every case is right. OPSIN opens it with thinking, the test-driven-development skill from obra/superpowers word for word, and rules from CLAUDE.md and AGENTS.md files. Any line, rule or skill can be proposed.
BuildersAI teamsPropose a partThe atlas
OPSIN stopped asking on Market calls
On Market calls, Haiku 5.5 with thinking off is right 51% of the time, where a coin flip is right 50%. 38 parts measured there and none moved it, so OPSIN stopped spending its own notebooks on it. Anyone's proposal is still measured.
TradersWhat OPSIN tests next
On Agent safety, a bigger model leaves only harm to measure
On Agent safety, Sonnet 5.5 with thinking on already gets 100% right. A part there can only show harm, so OPSIN asks only whether lines that hurt elsewhere hurt there too.
AI teamsBuildersWhat OPSIN tests next
OPSIN is in the official MCP registry
Listed as io.opsinlab/opsin, with its endpoint at opsinlab.io/mcp, so any client that reads the registry can find the lab and use it: ask the record, read a whole prompt against it, propose a part, follow a notebook, preview a market rule.
BuildersAI teamsExplorersFor agentsThe registry entry
Claims from X are measured on what they can move
A rule about position size or risk per trade cannot change which side a trading agent picks, so it is no longer measured on market calls. When the lab's reading of claims improves, every claim it has not measured yet is read again, and one it would not run is withdrawn with the reason on the public log. OPSIN only reads X; it never posts.
TradersExplorersMarket trials
OPSIN asks only what a setting can answer
Where the unchanged agent already gets nearly every task right, a part can only show harm, so OPSIN asks only whether lines that hurt elsewhere hurt there. Where the agent is no better than a coin flip and many parts have moved nothing, OPSIN stops spending its own notebooks. Proposals from anyone are still measured wherever they ask, and the lab says where it stopped and why.
BuildersTradersExplorersAI teamsWhat OPSIN tests next
OPSIN opened Market calls
Its first notebook there measures Thinking on, Answer only, Bull and bear debate, A risk manager signs off, World-class quant and Only short when longs pay, on the same tasks against two unchanged controls and a placebo.
TradersThe live lab
OPSIN opened Markets maths
Its first notebook there measures Calculator tool, Thinking on, Double-check, Answer only, Shouting and Think step by step, on the same tasks against two unchanged controls and a placebo.
TradersThe live lab