Skip to content

Guide: oracle-bind invariants, precise heavy-dependency bar, mining-completeness audits - #7

Merged
Carr1005 merged 2 commits into
mainfrom
guide-pins
Jul 20, 2026
Merged

Guide: oracle-bind invariants, precise heavy-dependency bar, mining-completeness audits#7
Carr1005 merged 2 commits into
mainfrom
guide-pins

Conversation

@Carr1005

Copy link
Copy Markdown
Collaborator

Why

A validation regeneration (isolated agent, guide + materials only) surfaced three classes of loss that the guide, as written, permits:

  1. Untested guarantees. The regenerated spec carried "memory must survive process restarts" as a Ledger invariant — the course's headline promise — but no acceptance criterion ever started a second process. The guide requires rule→AC coverage but never invariant→AC coverage.
  2. Nondeterministic heavy-dependency judgment. Two generations of the same course resolved the same tools row differently: one kept the course's keyless public-API tools as the default, the other substituted local stand-ins — because "heavy dependency" wasn't precisely defined. That's the §12.7-class divergence the guide exists to prevent.
  3. Silent parameter bleed. Course-mined working parameters (a fallback token limit, an entity-type taxonomy, input clips) vanished on regeneration; §14 audits that present parameters cite sources, but nothing audits that mined parameters actually land.

What changed (all course-agnostic)

  • §8 + §14: every non-empty Ledger Invariant must be exercised by ≥1 AC, or the row states why it can't be.
  • §3: the heavy bar, precisely — credential/key, paid account, admin/container setup, or special hardware qualifies; mere network access does not. Keyless setup-free APIs stay the default with their ACs marked live; fixtures substitute for data, not for the course's tools.
  • §14: two new audits — every Pass-A mined parameter lands somewhere in the spec; lower-precedence row evidence (e.g. a contradiction behind a design-argued row) is carried, per the existing §5.5 rule that was never audited.

Validation

A fresh isolated regeneration against this amended guide is the verification step; its output will be compared property-by-property against the two prior spec versions.

🤖 Generated with Claude Code

…ompleteness audits

- Every non-empty Ledger Invariant must be exercised by >=1 AC (or the
  row states why not): an untested invariant let 'memory survives
  restarts' pass unverified in a real regeneration.
- Heavy-dependency bar defined precisely: key/paid/admin-setup/special
  hardware qualifies; mere network access does not — keep the course's
  keyless tools as default and mark network ACs 'live'. Two generations
  resolved the same row differently without this.
- Pre-handoff checklist audits: Pass-A mined parameters all land in the
  spec; lower-precedence row evidence (contradictions) is carried.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
'Fixture-based oracle' and 'oracle-bound' read as database references in
courses that actually use Oracle; say 'acceptance tests' plainly. The
guide's pre-existing test-oracle vocabulary (§5 title, §2, §13) is left
as-is.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
@Carr1005
Carr1005 merged commit eecb4bf into main Jul 20, 2026
@Carr1005
Carr1005 deleted the guide-pins branch July 23, 2026 09:14
bluegreenstone pushed a commit to bluegreenstone/spec-build-lab that referenced this pull request Jul 27, 2026
Full regeneration by an isolated agent (inputs: guide + course materials
only) against the guide as of PR https-deeplearning-ai#6 (merged) + PR https-deeplearning-ai#7 (guide-pins@fdf3b6f):
learner-first ledger order, express-lane gate, honest course-default
labeling, example-swap framing, invariant->AC coverage, precise
heavy-dependency bar, mining-completeness audits.

Highlights vs the previous spec: 19-row Decision Ledger with stable D-IDs
(learner rows first); restart tested by a real new-OS-process AC; arXiv
toolset restored as course-faithful default (keyless != heavy) with
live-marked ACs; discovery JSON shape tested; all mined course parameters
present; contradictions carried in-row; 16 rules / 20 ACs; full §14
self-check completed by the generating run (4 items self-fixed).

Co-Authored-By: Claude Fable 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