session 2026-07-03: doctrine phase complete (Wave 0 + full spec-revision rulings + Loeb substrate reframe); PENDING-46; Phase 2 v2.0-draft next

claude/memory/skill-harvest-register.md

@@ -288,3 +288,13 @@ The single place proposed skills live so they don't evaporate between sessions.
 | **Symmetria §3 flag: re-derived-instructions-instead-of-citing-the-governed-doc** | Symmetria §3 flag | Wrote agents hand-made instructions (and invented fields) instead of pointing them at `conversion-runbook.yaml`/the spec — the direct cause of the graduation divergence. Kin to `reinvented-governed-tooling-without-checking-it-exists`, at the *instruction* layer. Antidote: cite the governed doc; if it's silent, that's a runbook gap to close, not a prompt to improvise into. | Symmetria §3 | **PROPOSED** |
 
 *(Four. The load-bearing two: BUILD `/graduate-chamber-source` now [it's the throughput multiplier for the whole Making batch], and record the graduation-spec+validator+gate as a named governed instrument [the steward's own idea, proven]. The two §3 flags are the session's real returns — both steward-caught, both class-fixed in the spec/gate. Not manufactured.)*
+
+### Harvest 2026-07-03 (doctrine phase — Wave 0, full spec-revision rulings, Loeb substrate reframe)
+
+| Element | Kind | One-line | Where it lands | Status |
+|---|---|---|---|---|
+| **Symmetria §3 flag: claim-from-derived-artifact-when-the-source-is-checkable** | Symmetria §3 flag | **The load-bearing harvest — jurist-elevated to STANDING PRACTICE.** I trusted a derived artifact / a regex-over-derived-text over the AUTHORITATIVE SOURCE **three times in one session**: the break-count (frontmatter fences + bullet-lists counted as thematic breaks), the heading-count (title-as-structure, then line-numbers-as-divisions, Plutarch-chapters missed), and "anchors absent" (searched Stephanus form `153a` when the DSL uses `St. II p. 153`). The steward twice sent me to the substrate (`loeb.dsl`) and was right both times; the jurist ruled it a governance rule: *any corpus-population claim feeding spec doctrine must be sourced from the source directly, or explicitly flagged extract-derived+provisional, before it does normative work.* Same discipline gap-2/gap-3 enforce on the corpus, turned on how the corpus is AUDITED. Antidote: verify against the authoritative source, never the derived form, when the source is checkable — and validate a matcher against a known positive AND negative before trusting its count (kin to census-through-a-pattern, generalized to source-vs-derived). | Symmetria §3 + `reference-verification-ladder.md` (a named "source-not-derived for doctrine-feeding claims" entry) + a feedback memory | **PROPOSED (strongly earned — 3× + jurist-ratified as standing practice)** |
+| **`/spec-amendment`** (the RFC-supersession amendment process) | create (later) | The now-RATIFIED chamber amendment process as a codified discipline: normative spec change = a superseding version (Obsoletes/Updates, never in-place edit) + change-class test ("does this change what any gate accepts?" → FIX vs PROPOSAL) + semver-MAJOR re-verifies all consumers + PENDING/REVIEWED record. Recurs every spec amendment. **Defer** until v2.0 is drafted (the process's first real exercise IS Phase 2 — build the skill from that empirical run, don't pre-author). | new skill | PROPOSED (build-after-first-use, i.e. after Phase 2) |
+| **per-claim citation verification (don't let a sub-agent's blanket "verified" ride)** | feedback memory or verification-ladder note | Caught by the jurist: I cited arXiv 2605.24229 for a claim it didn't support, having let a sub-agent's aggregate "4/4 papers verified" stand without per-citation checking. Verify each citation against the SPECIFIC claim it carries; never let an agent's blanket verification substitute for per-claim checking. | feedback memory / ladder | PROPOSED |
+
+*(Three, all earned. The first is the session's real return — a drift I hit three times and the jurist turned into a standing rule; it should become a Symmetria flag AND a ladder entry AND a feedback memory. The second is a genuine future skill but correctly deferred to after its first real use. The third is a concrete citation-integrity guardrail from a real miss. None manufactured.)*