Review of Addendum 1
Mode: refactor. Surface: interactive game and analytical tools. Locale: Czech/English. Source: docs/spec/01-addendum-simulation.md, existing code, tests, and a bounded independent code/reader review. Existing pirate identity, common rules, hot-seat play, and theme registration were preserved.
Findings and disposition
| Severity | Finding / evidence | Action | Confidence |
|---|---|---|---|
| HIGH | One global RNG made extra reveals shift all subsequent draws. | Named, keyed chance sources; pairing/independence regressions. | High |
| HIGH | Exported statistics lacked a complete reusable experiment specification. | Versioned experiment model, saved configuration, CLI/UI reruns and paired comparisons. | High |
| HIGH | Raw legal-action counts could be mistaken for meaningful choices. | Retain their literal label; add a separate policy-conditioned rollout estimate with formula, sample size and caveat. | High |
| HIGH | New agent-context draft exposed a seed that could be inverted to predict hidden outcomes. | Remove all source seeds from public contexts; keep policy randomness private; regression test. | High |
| HIGH | Invalid experiment controls could execute the previous valid JSON. | Disable Run on failed preparation or JSON validation. | High |
| HIGH | Imported subsets of bots could be replaced when only repetitions changed. | Preserve the imported subset until agent selection is explicitly changed. | High |
| MEDIUM | Retroactive bot-assisted trajectories were labeled “human”. | Use unattributed; explicitly label controlled experiment/runner traces. |
High |
| MEDIUM | Small means hid distribution and pairing uncertainty. | Median/percentiles, sample sizes, paired deltas and approximate intervals. | High |
| MEDIUM | Mixed Czech singular/plural instructions and English decimal presentation. | Consistent plural instructions, Czech dashes and locale-aware numeric results. | High |
| MEDIUM | Replacing action/character DOM nodes discarded keyboard focus. | Restore focus to the selected character or next available action. | High |
| KEEP | Theme-neutral mechanics and pirate names explain the same game. | Preserve both; add no engine fork or provider dependency. | High |
The claims ledger for the UI is deliberately narrow: numbers are actual simulation outputs; confidence intervals describe their calculation; the planner is not claimed to be optimal; estimates do not establish fun, age suitability, causal luck contribution, or a dominant strategy.
Dimensions
- Identity grounding: strong — the game and laboratory serve distinct player/designer tasks.
- Copy specificity: strong — controls name concrete actions, recorded artifacts, and comparison settings.
- Structural intentionality: stronger after refactoring — related variants share comparison tables rather than separate undifferentiated cards.
- Visual intentionality: unknown for the new screens — no rendered screenshot was available.
- Voice fidelity: mixed — existing pirate voice retained; technical controls follow the established direct Czech instructions, without independent author samples.
- Factual grounding: pass — generated results are labeled with versions, n, and their statistical limitations.
- Quality floor: partial verification — DOM interaction and engine tests pass; rendered contrast, responsive layout, focus appearance, and real-browser accessibility remain unverified.
The copy scanner was run before and after. Its 95% warning corresponds to the implemented interval, not an invented claim; many lexical hits were JavaScript identifiers. Typography checks were applied to extracted prose because applying them directly to JavaScript produced unsafe quote/code edits. No unsafe source rewrite was accepted.
Validation and follow-up
The suite passed 51 checks after this refactor, including legacy replay compatibility, seed isolation, hidden-information boundaries, trajectory corruption, historical branches, paired experiments, provider-neutral agents, UI import/export and stale-spec regressions. A 200-game paired comparison completed through the CLI.
The preferred VS Code browser could not be controlled: Computer Use again returned “Computer Use permissions are not granted”. DOM tests do not substitute for desktop/mobile visual review. No deployment or external AI calls were performed.
A stable identity profile could be extracted from the current theme definitions later; it is not necessary to block this functional refactor. Advanced analytics and automated optimization remain explicitly deferred in docs/experiments.md.