session 2026-07-13 evening: matcher key-selection fix (chamber 4b34447, MATCH-SUSPECT→0) + PENDING-58 (2b sidecar-wiring amendment, Loeb-first) + session memory/ledger/KG/skill-harvest

2026-07-13 19:05:17 +02:00
parent 32dfd7a14c
commit de1e2053ff
7 changed files with 108 additions and 1 deletions
@@ -463,3 +463,11 @@ The single place proposed skills live so they don't evaporate between sessions.
 | **confirm-a-named-cause-by-swap-in (don't assert from identification)** | verification-ladder instrument | When you NAME the true reference / config / cause behind an anomaly, don't assert the fix from the identification — **swap the candidate in and confirm the anomaly collapses** (ideally to zero). Today: naming augustine's true key ('Confessions' full) and philo's ('On Abraham') wasn't trusted; re-running the comparison against each drove fab+lost to exactly 0/0 — the checkable version of "this is the wrong match." The specific form of [[feedback-checkable-claim-surfaces-bugs]] for root-cause claims. | `reference-verification-ladder.md` | **PROPOSED** |

 *(Two, both earned in use, both recurring-shaped and distinct from the morning's checkable-claim §3 flag. The first is the session's load-bearing lesson — magnitude ≠ cause in a diagnostic map. Watch-list, not proposed: "counted-the-bundled-representation-as-items" [the ③ table row = 3 Aeschylus books counted as 1] — logged as a KG drift-pattern; may graduate to a census-through-a-pattern refinement if it recurs.)*
+
+### Harvest 2026-07-13 (evening — the matcher key-selection fix)
+
+| Element | Kind | One-line | Where it lands | Status |
+|---|---|---|---|---|
+| **method-class-vs-calibration** | verification-ladder entry (primary) OR Symmetria §3 flag | **When a metric/gate fails to separate two cases, ask whether it is mis-CALIBRATED or structurally BLIND to the distinction — no amount of retuning a set metric reaches an order-only difference, because the blindness is in the method CLASS, not its threshold.** Antidote: reach for a different method *category* and order tiers strongest-most-specific-signal FIRST (exact ordered-string before fuzzy set metrics). Earned today: Jaccard=1.0 AND raw-overlap both TIED on Suetonius 'Grammarians' vs 'Rhetoricians' (identical token sets — both words in "Grammarians and Rhetoricians"); only the ordered STRING separates them, and it had to lead. The steward elevated it: "the problem was in the *category* of method, not its calibration… no amount of retuning a set metric would ever have separated them." Generalizes past this matcher — any check that ties/fails where a distinction is real. | `reference-verification-ladder.md` (a diagnostic instrument), with a possible Symmetria §3 shadow ("retune-a-structurally-blind-metric" = reaching for the calibration knob when the method class can't make the distinction) | **PROPOSED** |
+
+*(One candidate, load-bearing and steward-named. Primary home is the verification-ladder — it's a positive diagnostic technique, not only a contamination shape. The §3-flag form is offered in case the steward reads it as the drift-shadow [reaching for the convenient tuning knob] rather than the instrument. His ruling on which home, or both.)*