Visitar URL original
Guide §6.0/§5.5: "(Recommended)" marks the Default; "(course default)" on every row's course option; model choice moves to the confirmed checklist by Carr1005 · Pull Request #57 · https-deeplearning-ai/spec-build-lab · GitHub
Skip to content

Guide §6.0/§5.5: "(Recommended)" marks the Default; "(course default)" on every row's course option; model choice moves to the confirmed checklist - #57

Merged
Carr1005 merged 8 commits into
mainfrom
guide-gate-labels
Oct 19, 2026
Merged

Carr1005 merged 8 commits into
mainfrom
guide-gate-labels

Conversation

@Carr1005

Copy link
Copy Markdown
Collaborator

One subject: the gate's option labels (owner decisions R1 (a), R2, R3, U7, 2026-10-19).

What changes

Before After
"(Recommended)" the build agent's own judgment at the gate (since 9fab72c, 2026-07-16) the row's Default — the generator's generation-time choice of the most efficient build of the course's pattern, i.e. what the recommended baseline builds. Tagged in the Default's own cell; the gate says once that it does not reflect the person's setup
"(course default)" in the spec required in the cell only on substitution rows on the course's option in every row where the course made a choice; shipped-code vs shown-only (#48 U7): the shipped one carries it, the shown-only one says so; the scope row carries none
Merged label "(Recommended - course default)" unchanged
Model currency via the "(Recommended)" flag a step-5 checklist line per model role the chosen options use; if no current model is known and the person names none, the course's model id, marked "may need changing"
Checklist = course choice / ≠ course choice adds no course choice (scope row); the build starts only after the person confirms the checklist

Touched: guide Options field, labels paragraph, gate steps 3 and 5, keyed-tool rule's last sentence, §14 (four lines); lab overlay ADD 1 and runtime file (model choice now points at step 5 — also gives the recurring "model roles" overlay gap a home in §0). Overlay re-pinned each commit; its four quoted passages verified with grep -F.

Why

Since July, "(Recommended)" could differ between builds of the same spec and from what the baseline built. The owner's original F12 proposal (#18, 2026-09-07) was the generator's efficient pick; the build-agent override came from an assistant amendment. A current-guide regeneration also dropped "(course default)" on 10 of 14 rows, which the old rule allowed.

Checked

  • 3 clean-room regenerations (Opus 5.5, guide 774a12a): agent-memory agnostic (15 rows), loop-engineering lab (6), loop-engineering agnostic (11) — every row's Default carries "(Recommended)" or the merged label; the course's option carries "(course default)" on every row it chose; scope row untagged; §0 copied verbatim; checker 0.
  • 3 gate probes on the lab spec (an agent walks the gate on paper, builds nothing): labels shown exactly as the cells write them; the probes found four gate-wording gaps (note placement, a mark for the scope row, one model line per used role, explicit confirmation) and one model-name gap — all fixed and re-probed. Final probe: baseline path, person names no model → every role ends with a named model; nothing undecided at build start.

Follow-ups (not in this PR)

  • Part 2 rule 4 / Part 5 evidence from the probes: "no grader" retires more tests than its option lists; the forced "no learning loop" option's text still describes a grader; the gate still asks rows made moot by an earlier answer.
  • The keyed-tool rule's note ("omit the event-driven layer … a row option the course never presented") conflicts with Guide §5.5: a stepping stone the course shows working is an option; structural subjects stay rows; one subject, one row #52's stepping stones; one generation read it as forbidding manual-only runs.
  • Minor wording: "that provider" (OpenAI vs the lab proxy); which model-id form goes into env vars; "Question N of M" counting in the lab runtime; recording python3 --version before any command runs.

Existing specs on main pin their guide hash and are unchanged until regenerated. Merge as a merge commit, not squash.

🤖 Generated with Claude Code

Carr1005 and others added 8 commits October 19, 2026 15:57
…fault)" on every row's course option; model choice moves to the checklist

One subject — the gate's option labels (owner decisions R1 (a), R2, R3, U7):
- "(Recommended)" now marks the row's Default: the generator's
  generation-time choice of the most efficient build of the course's
  pattern, i.e. the option the recommended baseline takes. The build agent
  no longer moves it on gate-time context; the gate says once that it does
  not reflect the person's setup. Merged label unchanged.
- The option the course actually used carries "(course default)" in its
  own cell on every row (was: only on substitution rows). Shipped-code vs
  shown-only realizations: the shipped one carries it, the shown-only one
  says so (#48 U7). The scope-boundary row carries none.
- Model currency moves from the "(Recommended)" flag to a new step-5
  checklist line: the build agent names a current model per role from the
  Default provider. Lab overlay ADD 1 and the runtime file updated to point
  at it (this also gives the lab's recurring "model roles" overlay gap a
  home inside §0).

Why: since 9fab72c (2026-07-16) "(Recommended)" was each build agent's own
judgment, so one spec could recommend different options in different
builds and a recommended option could differ from what the baseline built;
the owner's original F12 proposal (2026-09-07) was the generator's
efficient pick. A current-guide regeneration also dropped "(course
default)" on 10 of 14 rows, which the old rule allowed.

Touched: Options field, labels paragraph, gate steps 3 and 5, keyed-tool
rule's last sentence, §14 (three lines); overlay ADD 1; runtime file.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…uoted passages unchanged, verified with grep -F)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…d with "(course default)" when they coincide), so both labels are written in the spec as the gate shows them

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…uoted passages unchanged, verified with grep -F)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…one model line per used role, and explicit checklist confirmation

From a gate probe on a generated lab spec (labels rendered exactly as
written; four wording gaps):
- the "(Recommended) does not reflect your setup" note goes before the
  first Ledger row, not the baseline-or-customize question;
- the checklist gets a third mark, `no course choice`, for the scope
  boundary, where the course made no choice;
- the model line is one line per model role the chosen options use, naming
  a model the agent knows the provider serves, or asking the person when it
  cannot tell;
- the build begins only after the person confirms the checklist (the model
  line already relied on that confirmation, but no step asked for it).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…uoted passages unchanged, verified with grep -F)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…names none, use the course's model id from the spec and say it may need changing; a missing model name never blocks confirming the checklist

From the second gate probe: the agent could not name current models (its
knowledge predates the build, and the lab forbids listing endpoints), asked
the person, and the person confirmed without naming one — leaving the model
undecided after the gate.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…uoted passages unchanged, verified with grep -F)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant