session 2026-07-13: keystone landed+wired + Loeb corpus-health map (PENDING-57 CLOSED); +2 feedback memories (checkable-claim-surfaces-bugs, shorter-concentrated-sessions); MEMORY.md compacted 20.3→17.0KB

2026-07-13 15:42:13 +02:00
parent 48a2aec25d
commit 9c6279aca0
8 changed files with 174 additions and 9 deletions
@@ -0,0 +1,16 @@
+---
+name: feedback-checkable-claim-surfaces-bugs
+description: "Insisting on a CHECKABLE claim (a number, a fact, a verification) instead of accepting a soft classification repeatedly surfaces a real BUG, not just the figure asked for — the demand for verifiability is itself a defect-detector."
+metadata: 
+  node_type: memory
+  type: feedback
+  originSessionId: 734c43aa-f3a0-4cd1-a39d-27146420ee4a
+---
+
+Demanding a checkable claim — a reportable number, a substrate-verified fact, an injection test — where a soft classification would have sufficed does not merely produce the number: it repeatedly EXPOSES a real bug that the soft label was hiding. The act of forcing the claim to be verifiable is itself a defect-detector.
+
+**Why:** a soft classification ("this is reordering", "this is apparatus", "magnitude unresolved") can be *true-shaped and wrong* — it papers over a defect because nothing forces it to reconcile against the substrate. A checkable claim can't: it either reconciles or it breaks, and where it breaks is a bug. This is the studium engine's own thesis (free-the-reasoner / tighten-the-verifier) turned onto my own working method.
+
+**How to apply:** when tempted to ship a soft label, ask what CHECKABLE claim it's standing in for, and produce that instead — the number with its cause said plainly, the fact verified against code/substrate, the discriminating test. If the checkable version can't be produced cleanly, that inability is the finding. Never force a figure onto a comparison that can't bear it (that's noise wearing the costume of signal — worse than no number); name the actual problem instead ([[MATCH-SUSPECT]] over a fake deficit).
+
+**Evidence (steward-named 2026-07-13, "not coincidence at this point — it's what the discipline was for"):** five times in one day, pushing for a checkable claim surfaced a real bug rather than just the asked-for output — nagarjuna, the nested-block family, the loss-half injection test, the Aeschylus reordering (position-vs-multiset reconciliation), and the matcher key-selection bug (exposed only by restoring the REORDER? magnitude the map had wrongly withheld). Kin to [[feedback-completion-is-a-tripwire]] and [[feedback-trust-prior-pass-frame]]; the generative principle beneath the verification ladder.