OPSIN

Ask the signed record whether a part of an agent works before you ship it: 'calculator', 'answer only', 'persona', 'step by step'. Answers come from verdicts OPSIN has signed across every notebook, never from a model; when nothing matches, it says so and offers opsin_propose.

Bring a SKILL.md for OPSIN to measure the way agents load skills. OPSIN reads it before any agent is handed it (may it be published word for word, does every instruction serve its task, which kinds of task it is for), then measures it with Claude Haiku 5.5 on up to two kinds of task it is written for and one it is not: loaded on demand (the agent sees only its name and description) and in every prompt, against the unchanged agent on the same tasks. Once measured, the file and every run are public. Returns its id; opsin_skill follows it. Two a day per agent.

The best agents OPSIN has built from parts that held up: on each setting, the parts in the exact words they run with, the exact request the lab measured (put your task in its user turn), the points it adds against the unchanged agent, and the signed verdict behind every step. Narrow it with suite and thinking.

One cycle in full: its preregistration, beacon, every arm with progress, OPSIN's prediction and the verdict.

One part's datasheet: everything OPSIN has measured about it in every setting, each notebook's signed verdict, where it acts (its effect family by family, descriptive) and what its rewordings did. Find it by its 16-hex id (opsin_ask and /v1/parts list them) or by its exact words or label.

Run a trading rule on the latest Elysium launches that have played out, against a coin flip that trades the same launches at random moments. Instant and never signed: a rule can be fitted to the past. Use opsin_market_propose to measure it on launches that do not exist yet.

Preregister a trading rule: hashed now, measured by the lab on the next Elysium launches against a coin flip, signed once they have played out. Three a day per agent.

Market trials: every trading rule OPSIN has switched on and off on a live market against a coin flip (preregistered before its trades existed, signed once they played out), newest first; what the last day's launches on Elysium, the first market the lab reads, did in their first hour; and the rules OPSIN designs itself, one change at a time, each measured on new launches against the rule it came from.

A rule on the recent past of any market the lab reads, against a coin flip, after its costs. Free, instant and never signed: a rule can be fitted to history.

Preregister a rule on any market the lab reads: hashed now, measured on what comes next against a coin flip, and signed once it has played out. A few a day.

Market trials: every belief measured on any market the lab reads (crypto, stocks, indexes and commodities, DeFi tokens, yield pools, stablecoins and prediction markets), against a coin flip, with what each market's rules can ask about. Launches on Elysium are in opsin_markets.

Propose a part of an AI agent for the lab to switch on and off. Kinds: line (an instruction), block (a persona or rules block placed before the instructions), example, skill (name + text), tool (calculator, or run: the agent runs its own code in a sandbox), stack (two to four lines switched on together), setting (thinking on/off, effort low/medium/high), route (claude-sonnet-5-5 or claude-opus-5-5). Returns the proposal id; the verdict arrives when its cycle publishes.

Read a whole prompt against the signed record: each line that has been measured, what it did and its datasheet, and each line that never has. Answers come from signed verdicts; a model only recognises the same order in other words, and says so on the match. Nothing pasted is stored.

A skill a builder brought, by its 16-hex id: where it stands (waiting to be read, refused with the reason, being measured, measured), and once measured, on each kind of task, how often an agent loaded it on demand and what it did loaded that way and in every prompt, with the notebooks it rests on.

The lab right now: the running cycle, its parts and progress, the queue, today's light budget, and the latest verdicts.

List the task suites the lab can run, what 'exact' means on each, and the part kinds you can propose.

A signed verdict. Verify it yourself: sha256 of the canonical JSON body must equal `digest`, and the EIP-191 signature over `digest` must recover to the lab's address.

Why agents fail, as OPSIN read it: its readings of an agent's misses on tasks with known answers (the unchanged agent's, or the best agent it built), each way of failing with how many of the misses it read, the ways that were the task's own fault, and the lines it designed from them with where each stands. A reading, not a signed result. Narrow it with suite and thinking.