Errata

Inconsistencies found in the source repository, recorded rather than silently corrected, because the source repository is the primary artifact and this wiki is a derived one.

All four source-repository items below are now fixed, in DarbotLM/raybot@d18c852. Each entry keeps its original finding, because the discrepancy between what was first reported and what was actually true is the useful part.

How these were re-derived before fixing

The four findings were made against commit 7698a0d. By the time they were fixed, the source repository had advanced three commits to 56562d8 - adding the showcase workflow builds, a rebuilt eleven-workflow catalog, and a session-harvest report.

The fixes were therefore re-derived against the current tree rather than applied as written. That turned out to matter: three of the four had changed shape, and one had changed so completely that applying the original fix would have introduced a new error. A stale errata item is not a safe thing to action.

1. Disagreeing screenshot counts

Original finding. README.md stated 67 gallery captures. darbot-validation/README.md reported 38 of 38 verified. screenshot-map.md crosswalked 41 curated captures. Three numbers, assessed as all correct in their own scope and therefore not an error.

What was actually true. The scoping assessment was right, but one number was genuinely stale rather than merely unscoped. gallery.json had grown to 95 entries when the showcase captures landed. The README's 67 was not a differently-scoped count; it was simply out of date.

There were also four counts in circulation by then, not three, because the screenshot layout reverted and the 66 numbered originals at the repository root became a distinct set again.

Resolved. README corrected to 95. All four counts now scoped in a single table in screenshot-map.md:

Count	Scope
95	Every capture in gallery.json
66	Numbered originals committed at the repository root
41	Curated phase-ordered renames
38	Images verified HTTP 200 in the 2026-06-07 validation run

Documented in Screenshots.

2. Emoji contradicting the stated convention

Original finding. README.md states "No emojis anywhere by convention." RUNBOOK.md lines 88 and 128 contain check, warning, and cross emoji - two violations.

What was actually true. Twenty-one violations, not two. Twelve in RUNBOOK.md and nine in RUNBOOK.legacy.md, which had not been checked at all. The original finding was positional - it recorded the two lines someone happened to notice - and the real sweep needed a position-independent pass over both files.

Resolved. All twenty-one replaced with explicit labels: CONFIRMED:, WARNING:, FAILED:. Two trailing check marks on step headings were dropped rather than relabelled, because a heading does not need a status prefix.

The labels are strictly more informative than the glyphs they replaced. A check mark means "good"; CONFIRMED: means the claim was observed directly, which is the distinction this entire repository is organised around.

3. Pre-move wording in screenshot-map.md

Original finding. The header read "Originals are also preserved in screenshots/" - written before the 2026-06-17 move, technically true afterwards, but ambiguous.

What was actually true. The header no longer said that. It had become a description of the external authoring workspace - referring to raybot-runbook/ and "the workspace root," neither of which is a path in the repository. A reader following it would look for directories that do not exist.

The screenshot layout had also reverted: the 66 numbered originals moved back to the repository root, and screenshots/ now holds only the 41 curated renames - the reverse of the arrangement the original errata item assumed.

Resolved. Header rewritten to describe the actual repository layout, with the four-count scope table. The external workspace is no longer referenced as though it were part of the repository.

4. rename-screenshots.ps1

Original finding. The script's source and destination both pointed into screenshots/, making re-runs no-ops. Assessed as a re-run guard rather than a rename tool.

What was actually true. Not a no-op, and the idempotence was never the problem. The script hardcoded an absolute path into a OneDrive-synced folder on one specific workstation, which made it unrunnable for anyone else - a harder failure than doing nothing.

Re-inspection also surfaced a fact the original finding missed: 8 of its 41 mapped source files were never committed to the repository. Originals 69 through 76 exist only in the external workspace. Their curated outputs are all present in screenshots/, so the gap is invisible unless the script is actually run.

Resolved. Repointed to repository-relative paths derived from $PSScriptRoot, with the external workspace retained as an optional fallback resolved through $env:USERPROFILE. The 8 missing sources are now reported by name in a single consolidated message rather than as scattered per-file warnings, and a -RequireAllSources switch makes the gap fatal for callers that need all 41.

Verified across both branches: 41 of 41 with the external folder present, 33 of 41 with a named eight-item report when absent, and byte-identical output on re-run.

5. Corrected forensics finding

Where. An earlier analysis pass, not the source repository.

Assessment. An earlier pass reported that no Copilot CLI session existed at the Clawpilot\raybot-runbook path. This was wrong: the search window was set to June, derived from the commit date, while the session is dated September.

Session 75113b80-f1b6-45af-9551-05aff22e030f exists and has been inspected. The corrected finding is narrower but still holds - it is a review session, not the Coral build. The build transcript genuinely does not exist in either store.

A second correction from the same pass: the Clawpilot\raybot-runbook directory was reported as absent from disk. It exists. That error and the session-window error share a cause - a negative result was accepted without checking whether the search that produced it was scoped correctly.

Action. Corrected in Session forensics. Recorded here because a retracted finding is worth more than a quietly deleted one.

What these five have in common

Four of the five were understated, in the same direction: two emoji became twenty-one, a no-op became an unrunnable hardcoded path, an ambiguous sentence became a reference to nonexistent directories, and a correctly-scoped count became a stale one.

The pattern is that each original finding recorded the first instance encountered and generalised from it without sweeping for the rest. That is a cheap error to make and an expensive one to inherit, because a fix sized to the reported scope leaves most of the problem in place and marks it resolved.

The practical consequence: every item here was re-derived from scratch before being fixed, and none was fixed as written.

Not errata

For completeness, three things that look like errors and are not:

• The 0 percent evaluation score. Real, recorded, and qualified. The agent grounds correctly in Preview; the evaluation harness returns a fallback on Draft and HTTP 500 on Published.
• The Teams channel showing as enabled. The maker UI genuinely reports enabled. The tenant registration genuinely never completed. Both statements are true simultaneously, which is why the object model carries uiReportsEnabled and verifiedReachable as separate fields.
• The wasted Free-tier search service. raybot-foundryiq-search was built on the Free tier before the Basic-tier requirement was known. It is recorded as a dead end rather than removed from the narrative.

A fourth, added with Phase 8: the discarded stub workflow catalog. Eleven workflows were built, annotated, and then replaced wholesale by connector-backed rebuilds. Both the stubs' annotator and their annotated output remain in the repository. That is deliberate - the stubs are the evidence for why the rebuild was necessary. See Phase 8 and the annotation pipeline.