Kai: stage a6 on the frozen harness, what it wins and loses and why, the score-head probe and its fix, the sibling rows for a5 and a6, and what a6 leaves open.

Claude-Session: https://claude.ai/code/session_018Gt5TvrE2rxXuZikWSauv3

kai/sections/01-introduction.tex

@@ -47,9 +47,11 @@ designed, not built.
     floor, license classes and a contamination barrier
     (Section~\ref{sec:training}).
   \item \textbf{What went wrong}: stage A's encoder stopped relating two texts;
-    the measurements, the ablations that locate the cause, and the fix; and how
+    the measurements, the ablations that locate the cause, and the fix; how
     much of the typed-decisions test set Laya's checkpoint had seen in near
-    form (Section~\ref{sec:findings}).
+    form, and that it did not help it; and a score head that lost thirty
+    points to a softmax over level texts on the same encoder and data
+    (Section~\ref{sec:findings}).
   \item \textbf{Policy precedence} as a lattice join, with a proof that
     enforcing a model only adds friction, and \textbf{programs and packages}
     that replay, diff, refresh and say what flips a decision, with SysML~v2

kai/sections/05-findings.tex

@@ -112,8 +112,9 @@ states without a sibling and the 165 with one apart
 (Table~\ref{tab:siblings}). Laya's typed checkpoint scores alike on both
 halves, so its lead is not memorized from the siblings; Jev, which never saw
 them, is 8 points better without them; and Kai falls on exactly the states
-whose siblings the barrier removed from its training. The lead is the head and
-the training, and this section is a caveat about Kai's data, not Laya's score.
+whose siblings the barrier removed from its training. Stage a5 read the split
+whole and recovered the with-sibling half; the lead is the head and the
+training, and this section is a caveat about Kai's data, not Laya's score.
 
 \begin{table}[h]
   \centering
@@ -123,6 +124,8 @@ the training, and this section is a caveat about Kai's data, not Laya's score.
     typed decisions & all 400 & without a sibling (235) & with a sibling (165) \\
     \midrule
     Kai (stage a4) & 0.523 & 0.605 & 0.406 \\
+    Kai (stage a5) & 0.705 & 0.697 & 0.716 \\
+    Kai (stage a6) & 0.750 & 0.734 & 0.773 \\
     Laya typed & 0.766 & 0.773 & 0.756 \\
     Jev & 0.736 & 0.770 & 0.686 \\
     \bottomrule
@@ -133,6 +136,24 @@ the training, and this section is a caveat about Kai's data, not Laya's score.
   \label{tab:siblings}
 \end{table}
 
+\subsection{The score head}
+\label{sec:score}
+
+With typed decisions at 0.705 in stage a5 against Laya's 0.766, one probe put
+the gap in the head rather than the encoder or the data: Kai's head trained on
+Laya's own typed encoder (\texttt{kai-1-agent}) and typed decisions' train
+split alone, six epochs, scored 0.605 on the harness, choice 0.608, \op{noul}
+0.833, score 0.431, where Laya's checkpoint scores 0.733, 0.857 and 0.723
+\meas. The cumulative-link score head of Section~\ref{sec:model}, one location
+against ordered cutpoints, lost thirty points on score questions to Laya's
+formulation, a softmax over the level texts. Kai now answers a score question as
+a choice over its levels when the checkpoint says so (\texttt{ordinal: false}),
+trained with the ranked probability score in place of the Brier so that mass
+far from the gold level costs more than mass beside it; the same probe with
+that head reached 0.742 on the typed validation split at the point where the
+ordinal head sat at 0.571, and stage a6 raised typed decisions' score-question
+accuracy on the harness from 0.603 to 0.711 (Section~\ref{sec:accuracy}).
+
 \subsection{Other defects, each fixed}
 \label{sec:other}
 

kai/sections/08-evaluation.tex

@@ -3,15 +3,16 @@
 
 Every table in this section is written by \texttt{bench tex} in
 \texttt{hanzoai/decision} from the kept results of the runs it names, into
-\texttt{tables/}; no number in a table is typed by hand. The capability and speed tables
-are stubs until the scored checkpoint is timed; each says so. Each table's last row
+\texttt{tables/}; no number in a table is typed by hand. Each table's last row
 names the Kai revision, the device and the harness and benchmark commits it
 reports.
 
 \paragraph{Backends.}
-\textbf{Kai}: stage a5, one epoch on from the final stage-A checkpoint
-(Section~\ref{sec:collapse}) with typed decisions' train split read whole
-(Section~\ref{sec:overlap}), on the native runtime. \textbf{Laya}: Laya 0.3.20's published weights served natively
+\textbf{Kai}: stage a6, two stages on from the final stage-A checkpoint
+(Section~\ref{sec:collapse}): a5 read the suites a4 trailed with more rows and
+typed decisions' train split whole (Section~\ref{sec:overlap}); a6 answers
+score questions as a choice over their level texts (Section~\ref{sec:score})
+and repeats the suites still behind; one epoch each, on the native runtime. \textbf{Laya}: Laya 0.3.20's published weights served natively
 by the Rust port of its joint head (parity in Section~\ref{sec:checkpoints}),
 each case on the checkpoint Laya's router picks, except typed decisions, which
 are answered by Laya's typed-decisions checkpoint: the router sends them to the
@@ -61,6 +62,21 @@ tex}; the last row names the revisions.}
 \label{tab:harness}
 \end{table}
 
+\paragraph{What the table says.}
+Stage a6 leads on seven suites: emotion 0.935, Banking77 0.922, phishing
+0.993, toxicity 0.807, RAG relevance 0.652, model routing 1.000 and MASSIVE
+0.910 across 51 languages against Jev's 0.890 and Laya's 0.382. It ties email
+spam (0.998 to 0.998). It trails on four: AG News 0.940 to Laya's 0.950,
+jailbreak 0.920 to Jev's 0.940, typed decisions 0.750 to Laya's typed checkpoint
+at 0.766 (above Jev's 0.736), and support triage 0.435 to Laya's 0.502.
+Between a5 and a6 the score-question accuracy of typed decisions rose from
+0.603 to 0.711 (Section~\ref{sec:score}), choice and \op{noul} within two
+points of Laya's; support triage fell from 0.490 to 0.435 while its validation
+rose from 0.672 to 0.718: five passes over a train split whose labels are noisy
+overfit its conventions, and the next stage repeats it twice and adds rows
+written by our own model instead. Every figure here is
+\texttt{hanzoai/benchmarks} \texttt{decision/results/kai-a6} \meas.
+
 \paragraph{The gate.}
 A checkpoint replaces a baseline in a deployment only when \texttt{bench gate}
 accepts it: on up to 100 drawn cases per harness and extension suite and their
@@ -69,8 +85,12 @@ gain more than 0.03 ECE, whatever the macro average, and for Laya the native
 runtime's parity must hold. The first stage-A checkpoint was rejected: macro
 accuracy 0.639 over the 62 suites against Laya's 0.438, parity held, and seven
 suites regressed, RAG relevance, typed decisions, Banking77 and jailbreak among
-them \meas{} (\texttt{hanzoai/benchmarks} \texttt{9aa6a46}). The final
-checkpoint's verdict is reported with its tables.
+them \meas{} (\texttt{hanzoai/benchmarks} \texttt{9aa6a46}). Stage a6 is
+rejected against Laya on one suite, support triage, and against Jev on
+jailbreak, typed decisions, toxicity (its ECE on the drawn cases, at 0.807
+accuracy to Jev's 0.662 over the suite) and eight of the 51 MASSIVE languages
+\meas; so it does not replace either baseline in a deployment, and this paper
+claims no such thing.
 
 \subsection{Capabilities the baselines lack by construction}
 \label{sec:capabilities}
@@ -95,7 +115,9 @@ kept runs are at \texttt{4e57fb2}.
     readiness and review), with gold computed from a latent world by a policy
     the state spells out, so every gold answer keeps every rule. The baselines
     answer each question independently; Kai runs the refinement of
-    Section~\ref{sec:diffusion}. Per-variable accuracy, the share of states with
+    Section~\ref{sec:diffusion} over a decoder stage a6 has not trained, and
+    loses this suite to Jev (0.306 per variable to 0.837, and a rule broken in
+    every state before projection): the decoder is what stage C trains. Per-variable accuracy, the share of states with
     all five right, and the share whose answers break a rule, as answered and
     after the same exact projection applied to every backend's marginals.
   \item[Invariance.] 200 Banking77 questions of 16 options asked in five

kai/sections/09-limits.tex

@@ -34,6 +34,15 @@ explaining it. Local SGD at a gentle outer step may still move the encoder
 further than one process does at the same learning rates, and only the harness
 and the context similarity say whether it did.
 
+\paragraph{What stage a6 leaves open.}
+Support triage fell under repetition while its validation rose, so the
+harness, not validation, must pick a mixture. The joint-coherence suite is lost
+outright until stage C trains the decoder. The serving modes of the design, a
+decision over a cached latent in microseconds and an incremental encoder over
+evidence chunks, are designed and partly built (the cache of question, state
+and option latents is exact and tested), not measured; every latency in this
+paper is a full encode per call.
+
 \paragraph{Known failure modes of Laya.}
 Its authors report that boolean-word labels can be followed instead of their
 descriptions, negated requests answered as unnegated, \op{noul} answers driven

kai/sections/99-conclusion.tex

@@ -20,9 +20,11 @@ set had same-wording siblings in its training data, and measuring the halves
 apart showed the siblings did not help it. Both are written down here with the
 runs that show them.
 
-Kai's case does not rest on winning every benchmark: Laya leads where it was
-trained and trails Jev off its mixture, and Kai enters this paper only through
-the tables its benchmark writes. It rests on what a hosted model cannot offer:
+Kai's case does not rest on winning every benchmark: after stage a6 Laya leads
+on support triage and AG News, where its mixture holds the sources, Jev on
+jailbreak and by a point and a half on typed decisions through its typed
+checkpoint, and Kai on the other eight, MASSIVE among them; Kai enters this
+paper only through the tables its benchmark writes. It rests on what a hosted model cannot offer:
 a pinned revision, deployment where the data live, no per-call fee, a runtime
 linked into the system it governs, a package per run that says what was
 decided, from what, and what would change it, and, in Enso, less generation per