A compact experiment report
Lead with one finding and its scope. Put the result above the mechanism; readers should learn what happened before opening the implementation notes.
| Report element | What it must answer | Existing component |
|---|---|---|
| Finding | Did the preregistered decision rule pass? | ResearchCallout |
| Outcome | How large is the effect, and how uncertain? | ResearchChart for win rates |
| Mechanism | What single decision changed? | StrategyFigure |
| Diagnosis | What changed besides winning? | ResearchPlot: distributions or score heatmaps |
| Example | What did one preselected game look like? | GameReplay |
| Audit | Can someone trace every number and failure? | AgentContext and linked run ledger |
The liquidity experiment is the first worked example. Its score grid compares both policies within each game, and its exchange distribution asks whether behavior changed. Neither diagnostic substitutes for the primary win comparison.
What the example still needs
| Next component | Why it would help |
|---|---|
| Effect-size interval plot with a zero reference | Compare the actual policy contrast, beyond two separate win-rate intervals. |
| Public-safe decision trace | Show the target, alternatives and value terms that caused a move; aggregate counts cannot establish that causal chain. |
| Protocol health strip | Surface readiness, rate limits, retries, failures and unplayed games without burying them in logs. |
| Cohort comparison and population matrix | Keep development separate from fresh evaluation, then show which full opponent lineups support a result. |
Agent notes
New visual implementations belong in design and need Storybook states, keyboard checks and mobile QA. Web should only validate data and bind components. Do not hide custom SVG or arbitrary scripts inside reports.
The current arena has independent server-generated games, not paired seeds. Native decision latency is not a harness measurement. Verified local archives are not remote backups. Those are research infrastructure gaps as well as presentation constraints.