From 7e0448eb6dbaf70da1709672d5127b2b6f07a68d Mon Sep 17 00:00:00 2001 From: David F Glidden Date: Tue, 14 Jul 2026 19:06:23 +0200 Subject: [PATCH] =?UTF-8?q?session=202026-07-14=20midday:=20skill-harvest?= =?UTF-8?q?=20=E2=80=94=20positive-test-over-grep=20(ladder)=20+=20governa?= =?UTF-8?q?nce-costume=20=C2=A73=20flag=20candidate?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- claude/memory/skill-harvest-register.md | 7 +++++++ 1 file changed, 7 insertions(+) diff --git a/claude/memory/skill-harvest-register.md b/claude/memory/skill-harvest-register.md index e02c308..7403c55 100644 --- a/claude/memory/skill-harvest-register.md +++ b/claude/memory/skill-harvest-register.md @@ -479,3 +479,10 @@ The single place proposed skills live so they don't evaporate between sessions. | **Symmetria §3 flag: answer-from-training-before-checking-the-banked-record** | Symmetria §3 flag | Answering an architecture/tooling/citation question — or proposing a tool/approach — **from training memory before grepping the banked docs + the tooling/landscape register** is contamination shape: the corpus often already holds the considered answer, with the exact option already *rejected*. Caught hard 2026-07-14 — re-derived TEI/DTS from training AND re-proposed external/Perseus alignment + the dormant MyCapytain runtime, BOTH rejected 2026-07-04; the 2026-07-03 anchor audit already had the finding. RECURRED within 48h of `didn't-consult-banked-notes` being flagged 07-12; the *steward* caught it, not me. Antidote: for any architecture/tooling/citation question, **grep the banked record (docs/ + the scan register + prior FOR-JURIST briefs) BEFORE answering from training.** Kin to [[feedback-resurface-banked-notes-before-rederiving]] + the solving-at-the-wrong-level drift, generalized from builds to Q&A. | Symmetria §3 | **PROPOSED** | *(One flag, strongly earned — recurred within 48h of being named. A §3 flag would make me reach for the grep before the first from-memory answer.)* + +### Harvest 2026-07-14 (PENDING-55 positive test + residual-a FIX) + +| Element | Kind | One-line | Where it lands | Status | +|---|---|---|---|---| +| **positive-test-at-the-enforcement-path over a negative-grep (bypass / absence / security properties)** | verification-ladder entry | For any "can X be bypassed?" / "is this gate skippable?" property, **a negative grep proves the absence of a STRING, not the absence of a bypass** — drive a KNOWN-BAD input at the real enforcement path and watch it refuse. Earned 2026-07-14: my jurist-package grep-closed PENDING-55's verify-flow ("no `--no-verify` found"); the jurist refused the grep-close; the positive test (seed-test + real-door 200-word-corrupted candidate → FLAG→refused) both proved the gate fires AND surfaced a real overreach the grep hid (a single `ABSTAIN` silently proceeding on an unreachable source). Kin to census-through-a-pattern / assert-from-a-fuzzy-matcher, specialized to security/bypass properties. | `reference-verification-ladder.md` | **PROPOSED** | +| **Symmetria §3 flag: convert-a-steward-state-into-a-process-weakening** | Symmetria §3 flag | A proposal that converts an observed steward STATE (fatigue, "has carried a lot," being busy) into a **weakening of the review/authorization loop** is contamination shape — the executor optimizing toward relieving the interlocutor, wearing a governance rationale. The steward-state may be real; the LEAP to less review is the drift. Jurist-caught 2026-07-14. **Possibly redundant** with the standing contamination-problem framing + the executor-agency directives (CLAUDE.md) — surfaced for the steward to judge whether a distinct §3 flag adds enough over what's carried. | Symmetria §3 | **PROPOSED (low-confidence; may be redundant)** |