Captured 2026-08-31 by POSTing [internal endpoint] on :3334 directly, eight questions
in order, one at a time.
The first capture below was taken against an unconfirmed process. The engine was
then genuinely restarted (boot banner [WordMath] word_calculus: ON, v852+, 40,093
cells, 605,674 synapses) and all eight questions were asked again.
Seven of eight replies were identical. Q1 and Q4 again returned the same
Normalizing Flows paragraph as each other; Q2 and Q3 again refused; Q5, Q7 and Q8
identical. Only Q6 varied, between two different refusals ("Hard to say. My field
doesn't have a strong signal on that yet." -> "I'm drawing a blank there — I'll get
better at that over time."). Q1 again took ~19s cold, everything after it under a
second.
That reproducibility is the important part: this behaviour is a property of the
lattice and the scorer, not an artifact of which build was running. The table
below stands as the baseline for Phases 1-3.
[FieldHealth] EMPTY: keywords is set on 0 of 40093 cells (0.0%). Checked
before drawing any conclusion: keyword_overlap_score_idf takes cell_text: &str
and tokenises it, so the keyword term in the predictive scorer is not reading
that empty field and is not structurally zero. Some other consumer is; that is worth
chasing separately and is not part of this test.
[FieldHealth] UNIFORM: vitality is 1.00 on 100.0% of cells.
recycle_dead_claims() retains vitality > 0.0, so it can never remove anything.
Also out of scope here, also real.
ORIGINAL PROVENANCE WARNING (kept for the record). The engine log
(scratch/oracle-discord-kai.out.log) is frozen at Aug 28 22:01 and no file on
disk contains [Speech/Oracle], a trace that has existed since v9.10.845. So the
kai.exe that answered these eight questions **has not been confirmed to be the
v9.10.870 build**, and Start-KAI.ps1 explicitly does not start a duplicate when a
kai.exe is already running. These answers are still a valid *baseline* - every flag
is off in either binary - but no later phase can be compared against them until the
running process is identified.
Q1 and Q4 returned BYTE-IDENTICAL answers. "What is your name?" and "How are you
doing today?" have no content word in common. Getting the same paragraph about
Normalizing Flows from both is the clearest possible confirmation of the measured
root cause: after stopwords, these queries retain almost nothing, the keyword term is
binary presence, hundreds of cells tie, and the winner is effectively arbitrary. It
is not answering the question - it is returning whatever sits highest in a tie.
Q8 matched on the word "next". "What should we do next?" returned a passage about
finding the *next number* in a sequence. Literal token collision, no notion of what
the question was for.
Q2 and Q3 say "I don't have a solid answer" to the two questions KAI should be
most able to answer - the architecture it *is*, and who made it. The information is
certainly in the lattice; the scorer cannot reach it.
Q7 leaked transcript scaffolding into the reply ("From past conversation
(transcript history): Leo: Taz [Voice]:…").
Q1 took 17.8 seconds; every later question took under a second. Cold path on the
first query.
Baseline recorded before any flag was thrown, per KAI-TEST-wordmath-2026-08-31.md.