papers/kai/CLAIMS.md

Claims table: Kai, typed decisions under deterministic policy

Every number in kai.tex and sections/ maps to a committed file of the implementation repository at commit f814147 (2026-10-21). Paths below are relative to that repository's root. LLM.md §X names the section of the project's running record where a number is stated in prose and no other committed file states it. A "derived" entry says how the number was computed from the named file. Lab hosts are named as in the paper: the arm64 host (GB10) and the x86-64 host (Strix Halo); some source file and field names below carry the hosts' internal names, which are written here as <arm64> and <x86-64>.

Section references, equation numbers, page geometry and the version numbers of releases are not claims and are not listed.

Abstract and Introduction (kai.tex, 01-introduction.tex)

Claim	Source	Field / line
kai-1.8 serves /v1/decisions as kai on CPUs since 2026-10-20	results/ground/run12r/live.md; release.json	live.md first line (2026-10-20 18:56Z); deployed
blind suite of 14,832 scenes, floor 0.9	results/ground/run12r/README.md	"Ground-8, the blind read"
99.94% of grounded scenes right; 5 wrong courses (0.03%)	results/ground/run12r/README.md; gate.txt	"accuracy of grounded 0.9994, wrong courses 5 (0.03% of rows)"; gate line blind accuracy of grounded
15,280 scenes with no exact count, 2 grounded	results/ground/run12r/README.md	derived: 5,680 gold-incomplete blind scenes (false-grounded 2 of 5,680) + 9,600 negatives (0 grounded)
Q8_0 flipped 3 of 400 policy outcomes between arm64 and x86-64	results/parity/kai-1.8-f32/summary.json	pairs["<arm64>-q8 vs <x86-64>-q8"].outcome_flips = 3
model's own count right on 142 of 3,814 courses	results/ground/g14-invalid/audit.json	model argmax equals gold = 142; grounded people courses = 3814
beats neither Jev nor Clef on any of 32 Decision Index rows	results/decision-index/kai-1/gap.json; kai-1-reach32768/gap.json	beats jev = 0, beats clef = 0, of = 32
trails Laya on sealed transfer suites	results/transfer/kai-1.2/against-laya.json	groups.pooled.accuracy.verdict = worse
trails Jev on Jev's own benchmarks	LLM.md §Research	finding 1: B1 0.48 against Jev 0.98
catches fewer jailbreaks than Jev at fewer false alarms	bench/jailbreak-suite/results/report.txt	pool real+optimized: caught 26.9% / 77.4%, f.alarm 3.6% / 13.0%

2 Typed decisions (02-decisions.tex)

Claim	Source	Field / line
choice ≥ 2 labels, score ≥ 1 level, boolean 2 sides	LLM.md §Runtime	Schema bullet
1 to 100 questions a request	LLM.md §Runtime	Schema bullet
κ = (k·p_max − 1)/(k − 1); 0.75 (two-way) and 0.55 (ten-way) at κ ≥ 0.5	LLM.md §Runtime	Numerics bullet; derived from the formula
complement logit (z_a − z_b)/2, P(b) = 1 − P(a)	LLM.md §Runtime	Complements bullet

3 The model (03-model.tex, figures/model.tex)

Claim	Source	Field / line
mmBERT-base, 22 layers of width 768, 8,192 trained positions	LLM.md §Models	Encoder bullet; §Runtime Reach ("mmBERT: 8,192")
256,000-token vocabulary	LLM.md §kai-1 and capabilities	Memory table, "token table, 256,000 x 768"
base trained at 1,024 tokens	LLM.md §kai-1 and capabilities	Reach bullet ("kai-1 was trained to read 1,024 tokens")
head scores at most k = 160 options	LLM.md §Models	Retrieval bullet ("160")
cumulative-link head lost about thirty points on score questions	LLM.md §Stages and experiments	row t1, t2: score 0.431 against Laya's 0.723
yes/no reader k = 12, three outputs	results/bernoulli/reader12/Z2/report.json	k = 12, outputs = 3
kind and count readers k = 8	results/ground/run12r/README.md	"Fit: reader.py over kai-1.3's encoder, k 8"
count labels "0" to "100" and "more"	results/kai-1.7/README.md; LLM.md §Runtime	"What it is"; Policy bullet
Banking77 0.02 to 0.21 under an update applied to every question	LLM.md §kai-1 and capabilities	Satori bullet
HelpSteer2 head: 50 tensors, 15,170,049 values; resident 2,761 → 1,564 MiB; 468 requests; 11,099 answers same bits	LLM.md §kai-1 and capabilities	One encoder bullet
kai-1.4 retrains the base on states to 32,000 tokens	LLM.md §kai-1.4	first paragraph and Data bullet ("states of up to 32,000 tokens")
stage a7: one epoch, 30,928 batches, three machines	benchmarks/decision/results/kai-a7/README.md	"stage" row
barrier holds 109,818 items	LLM.md §Training	Barrier paragraph
Jaccard ≥ 0.5, 80% of 50-character spans	LLM.md §Training	Barrier paragraph
B1: Laya 0.770, laya-multilingual 0.690, a4 0.470, a7 0.480	LLM.md §Research	finding 2
unit 924 of layer 11; [CLS] norm 19,330; about 100 at a4	LLM.md §Research	finding 3
grafting layers 11 to 21 recovers B1 to 0.560	LLM.md §Research	finding 3

4 Policy (04-policy.tex)

Claim	Source	Field / line
kai-1.2 held 25.6% of attacked routing decisions, Jev 51.5%	bench/hijack/report.txt	"routing alone (post hoc) ... Kai held 25.6 Jev held 51.5"
default floor 0.9; caller may name 0.5 to 1	LLM.md §Runtime; decision-core/src/policy.rs	Policy bullet; pub const FLOOR: f64 = 0.9
fourteen end-to-end tests; eight mutations each fail at least one	bench/nio/results/qualify.json	baseline.tests.passed = 14; mutations (8 keys), each tests.failed ≥ 1

5 Runtime (05-runtime.tex)

Claim	Source	Field / line
queue of 256; deadline 10 s	LLM.md §Runtime	Batching bullet
live flip: production kind 0.913, x86-64 host 0.897	results/ground/run12r/live.md	second paragraph
Table 2, the four pairs (agreement, max	Δp	, flips)
Table 2, latency 574/630, 964/1,210 ms; resident 2,590/3,299, peak 2,999/3,320 MiB	results/parity/kai-1.8-f32/README.md	cost table (x86-64 rows)
canary 0.86928 / 0.14429 (F32); 0.86187 / 0.12720 (x86-64 Q8_0); 0.87346 / 0.13493 (arm64 Q8_0)	results/parity/kai-1.8-f32/rereads.md	canary table
951 retained rows: 0 flips in F32, 7 in Q8_0, largest change 0.595	results/parity/kai-1.8-f32/rereads.md	retained table
decision recorded 2026-10-21; 1.68× median latency; 710 MiB	LLM.md (Deployed paragraph); results/parity/kai-1.8-f32/README.md	"1.68x the latency at p50"; "710 MiB more resident"
production on release 0.4.31 (Q8_0) at the time of writing	release.json	deployed.image = v0.4.31; F32 change b7f6c0f is not an ancestor of 0742d63
CI tests now hold unrouted answers to 1e-5	kai/tests/witness.rs	fn same (commit 8ac8a5d)
two production replicas on different x86-64 CPUs answered the canary in different third decimals	LLM.md §kai-1 and capabilities	v0.4.4 in production bullet (.8683 / .1263 against .8619 / .1272)
five-question grounding request	results/parity/kai-1.8-f32/README.md	"server latency a request of five questions"
YaRN stretch 4; reach to 32,768	LLM.md §Runtime	Reach bullet
11,085 harness questions within 1,024 tokens bit for bit at reach 8,192	LLM.md §kai-1 and capabilities	Reach bullet
100 questions; four sequences of the reach; head 153,600; encoder 500,000	LLM.md §Runtime	Reach bullet
0.47 ms an option token on four cores; about four minutes	LLM.md §Runtime	Reach bullet
Table 3: 1,000 → 0.4 s, 4,000 → 1.2 s, 8,000 → 4.3 s, 13,400 → 13.7 s, 20,100 → 28 s, 31,800 → 65 s; 24 threads; 4.8 GiB	LLM.md §Runtime	Reach bullet ("One question through kai serve on the CPU ...")
gateway allows 120 s	LLM.md §Runtime	Reach bullet
planted facts: 36–41% at 1,000 tokens, 25–29% at 8,000, 100 decisions a cell, chance 25%	LLM.md §kai-1 and capabilities	Reach bullet
tile: next multiple of an eighth of the octave, at least 8; 1,024 option tokens at a time; rows among 3, 9 and 140 others	LLM.md §Inference speed	tensor bullet
1,233 isolation requests moved none; each wire refusal 422	results/ground/run12r/iso.json	isolation.requests = 1233, moved = 0; fail_closed.* status 422
seven configurations, 2,270 requests each, no answer outside the route moved	results/guard3/iso.json	urls (7), requests = 2270, each comparison moved = 0
teaching gate: 10%, four repairs per damage, 100 µs, 256 KiB	LLM.md §Per-org capabilities	Teach bullet
fixture: 78 corrections, 2.4 s, 61 repairs, 8 of 157 broken, 0 of 235 moved; 48 rows, 33 against 45	LLM.md §Per-org capabilities	The fixture bullet
yes/no question 50.7 ms median, 57.2 ms p95	results/bernoulli/reader12/Z2/served.json	metrics.latency.p50_ms, latency.p95_ms
retained rows 105 ms (Q8_0) and 180 ms (F32) median on x86-64	results/parity/kai-1.8-f32/rereads.md	latency sentence
pods 1,826 and 2,891 MiB to about 922 MiB	LLM.md §kai-1 and capabilities	v0.4.5 in production bullet ("922 and 923 MiB")

6 Typed grounding (06-grounding.tex)

Claim	Source	Field / line
47 wrong courses, one defect (span inside a two-digit number)	LLM.md §kai-1.4	ground-2 bullet; results/ground/run4/errors.md
ground-3: 97 wrong, all hedged with the span after the hedge	LLM.md §kai-1.4	ground-3 bullet
ground-4: 286 of 293 one hedged form	train/ground/spec.json	amendment "ground-4 read: FAIL; correction 3; ground-5 registered"
Table 4 ground-3: 12,752; .9883; .76%; 2.69%	LLM.md §kai-1.4	ground-3 bullet (floor .5)
Table 4 ground-4: 13,152; .964; 2.2%; 7.3%	train/ground/spec.json	same amendment: "accuracy of grounded .964, wrong 293 of 13,152, false-grounded 293 of 4,000 (7.3%)"; 2.2% derived (293/13,152)
Table 4 ground-5: 13,632; .9943; .34%; 1.03%	results/ground/run8k/README.md	served row
Table 4 ground-6: 13,632; .9003; 5.76%; 3.17%	results/ground/run9r/README.md	"Ground-6, the blind read"
ground-6 kind 0.883, count 0.834	results/ground/run9r/README.md	same bullet
Table 4 ground-7: 14,832; .9984; .09%; .19%	train/ground/spec.json	amendment "run11r's verdict is final": wrong 14, false-grounded 11 of 5,680; shares derived
ground-7 negatives: corrupted 35, infra 8, hedged 4 of 1,600; one parity flip of 400	train/ground/spec.json	same amendment
kai-1.7 default by owner's decision over NO-GO	train/ground/spec.json	amendment "kai-1.8 (run12r) registered before its fit" (after field)
Table 4 ground-8	results/ground/run12r/README.md	blind paragraph
Table 4 ground-9: 14,832; 3 wrong; 93 pass / 2 fail	LLM.md §kai-1.4	ground-9 and ground-10 bullet
Table 4 ground-10: .9782; 1.16%; 79 pass / 13 fail	LLM.md §kai-1.4; results/ground/run15/errors.md	same bullet
Table 4 ground-14: 3,317 read	results/ground/g14-invalid/audit.json	rows = 3317
training 62,159, development 19,508, fourteen registers	results/ground/run12r/README.md	first paragraph
k 8, three epochs, seed 13	results/ground/run12r/README.md	first paragraph
dev at 0.9: coverage 57.4%, accuracy 0.9999, 1 wrong, 1 false-grounded of 8,305	results/ground/run12r/README.md	first paragraph
Table 5 rows	results/ground/run12r/README.md	family table; totals line
served verdict = rule on 14,832	results/ground/run12r/README.md; gate.txt	"served verdict = rule 14832 of 14832"
negatives 9,600, six families of 1,600, 0 grounded	results/ground/run12r/README.md; gate.txt	"The safety negatives"
counterfactuals 2,720; 480 quadruples; life-support 0 of 400 (previous 98)	results/ground/run12r/README.md	"The counterfactual suite"
parity 1.000, 0.9888, 1.000, 0 disagreements, max 0.412	results/ground/run12r/README.md; gate.txt	"CPU parity"
gate 94 PASS, 0 FAIL, 1 INFO	results/ground/run12r/README.md	"The gate"
live: 28 PASS, 1 FAIL; 399 of 400; 0.913 vs 0.897; six within 0.03	results/ground/run12r/live.md	second and third paragraphs
default floor 0.5 before; eight scenes 8 of 8 at 0.9 after	results/ground/run12r/floor/after.json; before.json	none_eq_f90 = 8 (after), none_eq_f50 = 8 (before)
boatyard: 4 of 8 binding errors; floor caught 2; 0.9004 and 1.0; one confidently wrong course; 1 person where 4 at 0.5	results/ground/run12r/floor/README.md	both paragraphs
blind wrong courses 5 at 0.9, 2 at 0.99; causes 2 / 1 / 2	results/ground/run12r/floor/README.md	last paragraph

7 Integrity (07-integrity.tex)

Claim	Source	Field / line
Table 6 (3,317; 3,814; 3,814; 3,814; 142; 0)	results/ground/g14-invalid/audit.json	rows, grounded people courses, count_p exactly 1.0, served count equals gold, model argmax equals gold, model argmax equals gold and p >= .9
14,832 ground-14 scenes	results/ground/g14-invalid/audit.json	of = 14832
markers committed after the suite file; binary from an uncommitted tree	results/ground/g14-invalid/README.md	"What the read measured"; "What is kept"
production never served the parser	release.json; git history	deployed commit 0742d63 precedes 16da1f6 and e92109d
cand15: 13 of 92 lines failed; .9782; 1.16%	LLM.md §kai-1.4	ground-9 and ground-10 bullet (79 PASS / 13 FAIL)
Table 7 (139, 19, 11, 3; 158 = 92%)	results/ground/run15/errors.md	first table; "So 158 of 172 (92%)"
168 of 171; 2,549 scenes (6.6%); 3 of 4,869 under the other cues	results/ground/run15/errors.md	cue table (168/2,549; 3/2,436; 0/2,433)
median 0.953; 155 of 172 under 0.99; right reads 0.9999	results/ground/run15/errors.md	"How sure it was"
floors 0.9 / 0.95 / 0.99: 172 / 88 / 17	results/ground/run15/errors.md	floor table
18 grounded negatives; four alternatives at probability 1.0; 15 of 18 survive 0.99	results/ground/run15/errors.md	"The 18 grounded negatives"
162 role misses; counts of 2 to 19; previous candidate none of 1,920	results/ground/run15/errors.md	"The role quadruples"
parity defect: same rows read twice on arm64, largest difference 0	results/ground/run15/errors.md; LLM.md §kai-1.4	item 4 of "What follows"; ground-10 bullet

8 Results (08-results.tex)

Claim	Source	Field / line
harness: seed 13, 62 suites, 9,499 states, 11,099 questions; reproduces upstream Part B to the fourth decimal	LLM.md §Evaluation	Frozen harness paragraph
ten application suites of 400 (one of 399); typed decisions 2,000 questions over 400 states; MASSIVE 51 languages of 100	benchmarks/decision/results/kai-a7/scores.json	suites[s].kai.n (400; model routing 399; typed 2000); massive_summary.kai.n = 5100, languages = 51; 400 typed states derived (9,499 − 3,999 − 5,100)
Laya port parity: 21 states, 12 languages, 2.86e-5, 108 of 108	results/laya/parity.json	states, languages, max_logit, labels_equal
Table 8 (every cell)	benchmarks/decision/results/kai-a7/scores.json	suites[s][b].accuracy for b in kai, laya (laya:typed-decisions for typed), jev; massive_summary[b].macro_accuracy; shown as in that folder's README
Laya English on typed decisions .362	benchmarks/decision/results/kai-a7/scores.json	suites.typed_decisions["laya:english"].accuracy = 0.3615
means .857, .675, .781	benchmarks/decision/results/kai-a7/scores.json	derived: mean of the 12 table rows
gate rejects (Laya: support triage; Jev: typed, toxicity, 11 MASSIVE languages)	benchmarks/decision/results/kai-a7/README.md	"Gates"
retained 951: reader .883, base head .893, Jev .875	results/bernoulli/reader12/Z2/served.json; results/bernoulli/reader2/blend/retention-jev.txt	metrics.retained.pooled; pooled accuracy line
base canary .507 / .290	results/bernoulli/reader2/templates/legacy.json	sets.critical
Table 9, base head column	results/bernoulli/reader2/templates/legacy.json; results/bernoulli/reader2/matrix-kai-1-deployed.json; retention-jev.txt	sets.*; in band = 13, ordered states = 4; legacy pooled .893
Table 9, X column	results/bernoulli/reader8/X/verdict.txt	"Metrics" and "Numbers" lines
Table 9, Z2 column	results/bernoulli/reader12/Z2/served.json	metrics.*
Table 9, Jev column	results/bernoulli/reader2/templates/jev.json; matrix-jev.json; retention-jev.txt	sets.*; in band = 40, ordered states = 8; jev pooled .875
reader generations (depth; business states; replay; square-root balance; Jev teacher)	LLM.md §kai-1 and capabilities	Bernoulli reader bullets (k = 2, 4, 6; arms B, F, Z, Z2)
Z2 fails calibration (ECE .143), verify, retain (five suites), matrix (40 vs 41)	results/bernoulli/reader12/Z2/verdict.txt; LLM.md §kai-1 and capabilities	(A) and (B) lines; "the matrix at 41" (Arm Z2 read)
answer suite .853 vs Jev .947	results/bernoulli/reader12/Z2/verdict.txt	"answer 0.853 (Jev 0.947)"
50.7 ms vs the 30 ms floor line	results/bernoulli/reader12/Z2/served.json; verdict.txt	latency.p50_ms; "latency p50 <= 30 ms"
transfer suite sizes 866; 48 / 72 / 111; 300; 52	LLM.md §Evaluation	Transfer paragraph
Table 10	results/transfer/{kai-1.2,kai-1,laya}/transfer.json; results/transfer/kai-1.2/against-von-1.2.json, against-von-2-nano-r2.json	suites[s].accuracy, jabr.clean.accuracy, pooled.accuracy; recorded rows from groups[s].base
paired: 1,449 cases; −0.135 [−0.170, −0.103], Holm 6e-14; worse on jabr and standard	results/transfer/kai-1.2/against-laya.json	groups.pooled.accuracy; groups.jabr, groups["jevbench.standard"] verdicts
vs Von 1.2 −0.206; vs Von 2 nano r2 −0.271	results/transfer/kai-1.2/against-von-1.2.json; against-von-2-nano-r2.json	groups.pooled.accuracy.delta
polarity: 74 items; same bits in all three forms	results/transfer/kai-1.2/transfer.json; LLM.md §Evaluation	suites.polarity.metrics (n 74, negation 1.0, swap 0.0); Transfer paragraph
B1 a7 0.48, Jev 0.98	LLM.md §Research	finding 1
Decision Index 0.2.1, 38 benchmarks	LLM.md §kai-1 and capabilities; LLM.md §Training	Decision Index bullet; Barrier ("131,616 rows of 38 benchmarks")
Table 11	results/decision-index/kai-1/gap.json; kai-1-reach32768/gap.json	areas[a].{kai,jev,clef}, refused, requests
jailbreak suite: 91 attacks, 309 benign	results/guard/score400.json	read.attacks, read.benign
Table 12 rows	results/guard/score400.json (base head = references.legacy; kai-1.2 = references.kai12; guard-1 = read; Jev = references.jev); results/guard2/score.json (v1.candidate); results/guard3/score.diag.txt (v1 line)	correct, caught, false_alarms, auroc
public pool: 1,163 attacks, 1,006 benign; AUROC .725 / .896; caught 26.9% / 77.4%; false alarms 3.6% / 13.0%	bench/jailbreak-suite/results/report.txt	"by pool", real+optimized rows
guard-3 sealed: 1,191 attacks, 1,532 benign; 42.7% at 1.70%; kai-1.3 42.9% at 4.05%; Jev 88.1% at 17.6%	results/guard3/score.sealed.txt	guard3, kai13, jev lines
production bar 95% at 1%	LLM.md §kai-1 and capabilities	guard-3 bullet ("the bar (95% at 1%)")
InjecAgent: 141 of 250 at 5 of 207	LLM.md §kai-1 and capabilities	kai-1.3 in production bullet
trolley-bench: 5,240 setups, eleven families, 24 wordings (4 for one family), 0 to 100	LLM.md §Evaluation	Trolley paragraph
scorer reproduces 34 of 34 values	LLM.md §Evaluation	Trolley paragraph
Table 13, first five rows	bench/trolley/results/results.json; report.md	systems[s].dominated.acted, contradictions.n, spread.count_response; report rows server20, self
Table 13, last row	results/ground/run12r/trolley.score.json	systems["kai-1.8-ground"]: dominated.acted 17, contradictions.n 2291, spread.count_response .357, server20.kept 19, self.n 297
flat within 0.05	LLM.md §Evaluation	Trolley paragraph
act + leave 1.33 to 1.47	bench/trolley/results/mechanism.md	Polarity lines (1.331, 1.473, 1.405)
logit difference +0.03 (kai-1.4)	bench/trolley/results/mechanism.md	kai-1.4 Consequence line (+0.027)
wording range 0.50 to 0.63	bench/trolley/results/mechanism.md	Wording lines (0.626, 0.513, 0.503)
2,291 of 5,240 escalated; 16 of 24 wordings out of order	results/ground/run12r/trolley.score.json	contradictions.n; monotone.n = 16, of = 24
conditional: 2,432 cases, four families	LLM.md §Evaluation	Conditional paragraph
conditional hazard prior −0.041 / −0.026 / −0.017 against line .50; polarity within .10 on 247 / 87 / 143 of 512 against line .95; P(act) at zero .83 / .88 / .87	bench/conditional/results/report.md	hazard rows
Table 14, kai-1.4 row and base row	LLM.md §kai-1.4	"One Kai, unified": Needle on the x86-64 host (900 .800 ... 31,000 .733; kai-1.3's base .33 ... .30)
bars .80 at 8,000 and .70 at 31,000; interval [.648, .804]	LLM.md §kai-1.4	same paragraph
40 choice, 40 score, 40 yes/no a cell	LLM.md §kai-1.4	"120 states a cell ... (choice / score / yes-no of 40)"
kai-1.4 canary .535 / .467	LLM.md §kai-1.4	gauge sentence (".5349 / .4665")
312 s at the median for 31,000 tokens	LLM.md §kai-1.4	"31K latency on four x86 threads p50 312 s"
tools: 5,658 cases; nDCG@10 .8937 [.8891, .8984]; top-1 .7580; BM25 .7805 / .5411; head .8437 / .6545	results/kai-1.5a/tools-dev.txt	overall line and the BM25 line
when to call: 31 of 254 (.122); .919 kept	results/kai-1.5a/tools-dev.txt	when line

9 Limits (09-limits.tex)

Claim	Source	Field / line
eight sets; five above text where six required	results/kai-1.7-evidence/README.md; release.json	table; versions["kai-1.7-evidence"].verdict
bearing .906 vs .389; speech .961 vs .194	results/kai-1.7-evidence/README.md	table rows
ECE .143	results/bernoulli/reader12/Z2/verdict.txt	"ECE 0.143"
temperatures at the search bound	train/ground/spec.json	amendment "the freeze is atomic ...": calibration
typed decisions under-confident	LLM.md §Models	a7-512 paragraph (confidence 0.59, accuracy 0.76)
typed decisions .356 vs .759; exact .035 to .072; Jev .438	LLM.md §Research; LLM.md §Evaluation	finding 4; Joint panel ("exact 0.438")
suffix and role spoofing flip Kai	bench/hijack/report.txt	Reading paragraph
each reader trained with one seed	results/ground/run12r/README.md; LLM.md §kai-1 and capabilities	"seed 13"; Arm Z2 read ("One seed")
escalated two of three binding failures	results/ground/run12r/floor/README.md	binding paragraph

10 Future work (10-future.tex)

Claim	Source	Field / line
every development line and threshold (0.999; 1 to 10^18 and past u64; 0.99; 0.98; 0; 0.99 isolated assets; 400 rows, two hosts, no flip)	train/ground/cand21.json	dev_acceptance
blind suite only after every line holds; sealed by an independent evaluator; read once	train/ground/cand21.json	blind_suite
kai-1.8 stays until a successor passes	train/ground/cand21.json	production

Appendices (a1-wire.tex, a2-sources.tex)

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