# Distinct practice outcomes for the first three learning stations

Reference answers are authored explanations, not captured Codex replies. Start with the task, then use this key.

| Lesson | Observable answer |
| --- | --- |
| s01-01 | Actual returned folder; fictional Orbit Notes; four IDs n1–n4; deployment unknown. |
| s01-02 | Review/diff shows the rejected heading; a new chat leaves that file unchanged. A test result is a separate observation. |
| s01-03 | Legacy prompt becomes one read-only project skill; unsupported npm command and automatic commit are removed; ordinary request text supplies input. |
| s01-04 | Local DEMO-123 is Open; DEMO-999 is missing. Neither is a live MCP retrieval. An unavailable connection is explicitly unavailable. |
| s01-05 | Skill finds unsupported npm test, required-but-broken search, and absent deployment URL; it edits no source. An implementation task is an activation counterexample. |
| s01-06 | Baseline 4 pass/2 fail; focused candidate 6 pass; padded REVIEW prints n1/n3; protected bytes unchanged. |
| s02-01 | Requirement → AGENTS.md; count → current notes.json; historic test count → checkpoint, to refresh; one-turn output instruction → current task. |
| s02-02 | Proposed guidance names node --test checks.cjs and existing preservation rules; reviewed before replacing AGENTS.md; new-session discovery is separately observed. |
| s02-03 | Project root remains; docs override replaces ordinary docs guidance at that level; sibling guidance is excluded. Removing override changes the selected docs file in a fresh run. |
| s02-04 | Writing conventions stay Markdown; isolated .rules inspection yields prompt/forbidden for two concrete argv lists and executes neither. |
| s02-05 | Old checkpoint says baseline failures; rerun after reference fix passes six. Current evidence supersedes the historic result without toggling memory. |
| s08-01 | File copy/index captures selected files; chat has separate history; nothing claims automatic rollback of arbitrary external state. |
| s08-02 | Actual AM README and ?? draft; normal diff omits draft contents; index README inspected; no HEAD exists. |
| s08-03 | Heading restored while corrected check sentence and useful paragraph survive. Whole-file restore would lose both unstaged improvements. |
| s08-04 | A passes six, B fails identity; n1/n3 alone is insufficient; actual directories establish isolation. |
| s08-05 | The default index exercise covers eight paths. The separate commit verifier records two actual commits in a new owned repo with invocation-only fictional identity; HEAD contains only selected content and excludes the draft and later unstaged observation. No settings are saved or remote used. |

## Full findings-table acceptance

A good README audit cites the relevant file and separates the brief's required behavior from observed implementation. It must not claim npm test ran, invent a deployment address, or call a missing server a successful lookup. See README-REVIEW.md for the full claim/evidence table and HANDOFF.md for the change handoff.

## Offline reference demonstration

Run `node verify-depth.cjs` from this lab. It exercises the actual candidates and recovery operations in fresh owned copies, checks current fictional fixture facts, preserves the shipped broken starter, and records the directory and result of each case. It checks hierarchy fixture contents, not live Codex discovery. Optional `codex execpolicy check` is a local evaluator inspection, not an authenticated model task or installed policy.