From c31bc82621c7c9577db57e9089c062ce6f22f7bd Mon Sep 17 00:00:00 2001 From: Shay Date: Thu, 4 Jun 2026 02:49:00 -0700 Subject: [PATCH] =?UTF-8?q?docs:=20third=20leg=20=E2=80=94=20candidate-gra?= =?UTF-8?q?ph=20builds=200=20admissible=20candidates=20on=20held-out?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit All three GSM8K paths exhausted on the held-out 500: - candidate-graph (sound filter): 0 admissible candidates built (498/500 build nothing) - composer (open): 87 wrong - narrow reader (open): 61 wrong The sound path is wrong=0 only because it constructs no reading to commit. No sound lift exists in the current substrate; confirmed empirically, three ways. --- .../real-gsm8k-capability-measurement-2026-06-04.md | 7 +++++++ 1 file changed, 7 insertions(+) diff --git a/docs/analysis/real-gsm8k-capability-measurement-2026-06-04.md b/docs/analysis/real-gsm8k-capability-measurement-2026-06-04.md index 7b38bde4..20cfbd9c 100644 --- a/docs/analysis/real-gsm8k-capability-measurement-2026-06-04.md +++ b/docs/analysis/real-gsm8k-capability-measurement-2026-06-04.md @@ -65,6 +65,13 @@ cue word" are almost never 2-operand problems — they are multi-step problems w shallow reading is wrong **every single time**. Shallow committing **cannot** be sound on real GSM8K. This is measured, not asserted. +**Third leg — the sound path has no headroom either.** The candidate-graph (the +roundtrip-admissible + disagreement filter that *is* `wrong=0` on held-out) was checked for +disambiguation headroom: on the 500 it builds **0 admissible candidates** (498/500 enumerate +zero branches at all). It is `wrong=0` not because it disambiguates well but because **it +constructs no reading to commit.** There is nothing to soundly tiebreak. All three paths are +exhausted: the sound one builds nothing; the two that build, confabulate. + ## 3c. The architectural impasse (the load-bearing finding) CORE has two GSM8K paths, and the held-out data shows **neither can soundly lift capability**: