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

[Carr1005]

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