Kai: release 0.4.32 serves F32 on the CPU; its evidence and canary

kai/CLAIMS.md

@@ -229,3 +229,12 @@ The two listings are request shapes with states written for this paper; they
 carry no measured value. The count labels `"0"` to `"100"` and `"more"` are as in
 section 3. Commit `f814147` is the implementation repository's head when this
 paper was written.
+
+## Release 0.4.32 (added after the roll)
+
+| claim | source |
+|---|---|
+| no argmax flip over 955 answers, largest difference 1.8e-5, arm64 vs x86-64 | decision `results/release/v0.4.32/rows.json` (`argmax_flips`, `max_dp`) |
+| 400 typed policy rows, 0 policy-outcome flips | decision `results/release/v0.4.32/parity.json` (`outcome_flips`) |
+| mid-roll: new replica within 1.6e-6, old replica .015 to .043 off | decision `results/release/v0.4.32/canary-production-midroll.jsonl` against `canary-reference-midroll.jsonl` |
+| after the roll: 16 of 16 within 1.5e-6 | decision `results/release/v0.4.32/canary-production.jsonl` against `canary-reference.jsonl` |

kai/sections/05-runtime.tex

@@ -73,10 +73,15 @@ approximates them.
 
 The decision, recorded on 2026-10-21, is that F32 is the production CPU path,
 with no flag and no pinning to one CPU family. It costs 1.68 times the median
-latency on x86-64 and 710\,MiB more per replica. At the time of writing
-production still runs the last qualified image (Q8\_0, release 0.4.31). The
-F32 build rolls one replica at a time, after its canary through the public API
-matches the F32 references, with 0.4.31 as the rollback.
+latency on x86-64 and 710\,MiB more per replica. Release 0.4.32 carries it and
+serves since 2026-10-22. Before the roll, its binary read the canary pair and the
+951 retained rows on both families with no argmax flip (largest difference
+$1.8\times10^{-5}$) and 400 typed policy rows with no policy-outcome flip. It
+rolled one replica at a time with 0.4.31 as the rollback. Mid-roll, the new
+replica's answers through the public API matched the F32 reference to
+$1.6\times10^{-6}$, while the 0.4.31 replica beside it read 0.015 to 0.043 off on
+the same requests. After the roll, 16 of 16 answers matched to
+$1.5\times10^{-6}$.
 
 One side effect surfaced at once. In F32 on x86-64, the CPU's matrix kernel for
 a small batch depends on the batch's shape, so a sequence encoded beside others

kai/sections/09-limits.tex

@@ -56,6 +56,6 @@ phrasings, not text written by someone who wants the model to fail. The next
 suite is to be sealed by an evaluator who built neither the candidate nor its
 data (Section~\ref{sec:future}). Each reader was trained with one seed.
 
-\paragraph{Serving.} The F32 CPU path is decided and not yet rolled. A long
+\paragraph{Serving.} F32 costs about 1.7 times the median latency on x86-64. A long
 state on a CPU replica can exceed the gateway's time limit. A batch-variance
 failure on CUDA in BF16 is open, outside the serving path.