Owner's call. E4B is 4.2GB against 6.7GB plus a 0.86GB draft, so with CW2
resident the card holds 5.8GB of 16GB instead of 9.2GB.
Measured against a same-session 12B control on the 96-case fixture: 83.3% full
against 84.4%, 89.6% intent-only against 91.7%, destination 19/33 against
23/33, p50 294ms against 344ms. Destination is the column that moved. E4B names
nothing where the 12B names recall or calendar, which walks the whole chain
rather than answering wrong.
MTP is gone with the 12B and cannot come back. It is a separate gguf of
architecture gemma4-assistant with nextn_predict_layers=4, and the only one on
disk is trained against the 12B's hidden states. Neither target gguf carries
nextn tensors, so neither self-speculates.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013ptwopxyo3Z2kwFckHkLvN
serve.py lived only on workpc, which was fine while systemd launched it and is
not fine now that mavgpud does. Two endpoints and no framework: /health answers
503 until the model is loaded, /transcribe takes raw PCM and returns
{"text","confidence"}.
The unit carries CW2_TOKEN through EnvironmentFile and the child inherits it,
so the token is never a flag value.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013ptwopxyo3Z2kwFckHkLvN
CW2 is a ROCm process, so it registers on the KFD like any contender. Running
it as its own systemd unit made mavgpud yield llama-server to it every few
seconds. The gemma-4-12b arm was down for eight minutes on 2026-08-09 and
routing had silently fallen back to the resident model.
So mavgpud takes an `stt` block and runs the transcriber itself. `foreign` now
excludes every child rather than one pid, which is the fix. Yielding is all or
nothing, because a job that wants the card wants all of it. Idle unloading
stays llama-server's alone: CW2 holds 1.6GB and unloading it would only send
the next voice turn to the homesrv floor.
Maven still talks to the transcriber directly on 8081. There is no proxy,
because with no idle timer there is nothing for one to measure.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013ptwopxyo3Z2kwFckHkLvN
The block is inert until the code in PR #208 lands, and deleting it sends
every utterance back to mavsttd, which is what the box does today.
Port 8081 and not mavgpud's 8080, because whisper.cpp cannot load
CrisperWhisper 2.0 at all and it runs under transformers as its own service.
The token comes from deploy/telegram.env like every other secret here. It is
what stops anything on the LAN posting audio to that port.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013ptwopxyo3Z2kwFckHkLvN
sttSeam is modelSeam for audio and sits at the same place in wireVoice, so
the voice path and the meeting recorder share one transcriber as they
always have.
A box with no workstation.stt block behaves byte-for-byte as it did before
this existed: the floor is handed back untouched and nothing probes. An
empty URL is normalised to no block at all, the way the model block already
works.
Health defaults to the URL's origin rather than the URL itself, because the
transcribe endpoint names a path and appending would ask for
/transcribe/health. A block with no token logs once that anything on the
LAN can post audio to that port.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013ptwopxyo3Z2kwFckHkLvN
Same arrangement as llm.Pair and for the same reason. The microphone is at
workpc, the card there has 16GB, and CrisperWhisper 2.0 turbo scores 10.4%
WER in Russian against 27.5% for the ggml-small.bin homesrv loads. The
workstation is never assumed up: it sleeps, and the card is often held.
Admission is a cached atomic written only by the prober, so no voice turn
ever waits on a machine that may be asleep.
Speech-to-text has only the silent half of the degradation rule. A worse
transcript is still a turn, so there is nothing to name a gap about and
Transcribe always falls back. That is the whole difference from llm.Pair,
which also carries CompleteRemote for callers that must refuse instead. A
remote that dies mid-request corrects the cache and falls back in the same
turn, which is what TestPairFallsBackWhenRemoteFails pins.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013ptwopxyo3Z2kwFckHkLvN
mavsttd is whisper.cpp linked into a Go daemon and reached over a unix
socket. CrisperWhisper 2.0 cannot be reached that way. whisper.cpp derives
its language count from the vocabulary size, and CW2's 51897 tokens shift
seven special token ids, so it never loads at all.
So it runs under transformers on workpc and this is the client. Same
stt.Transcriber interface and one method, a second transport rather than a
second seam. The body is the PCM itself, because a minute of 16kHz mono is
under 2MB raw and the format is fixed by audio.PCM16kMono.
Audio is the most sensitive thing that crosses this seam, so the client
carries a bearer token.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013ptwopxyo3Z2kwFckHkLvN
Turbo in Intended mode scores 10.4% WER on 200 Golos crowd clips, against
27.5% for the ggml-small.bin the box loads today. It also beats its own base
model and CW2 large, which inverts what the card implies about turbo.
The mode choice is not settled by this corpus. Intended and verbatim disagree
on 29 of 200 after normalization, and the disagreement is script rather than
disfluency. Golos crowd carries almost no disfluency to disagree about.
whisper.cpp cannot load CW2: num_languages() derives from n_vocab and CW2's
51897 shifts seven special token ids. So the runtime is workpc under V-486,
with whisper on homesrv as the floor.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013ptwopxyo3Z2kwFckHkLvN
embedder.heads_path is empty by default and deploy/mavend.json sets
it. A missing or broken weights file logs and leaves the heads nil,
because refusing to start over a routing accelerator would trade a
working box for a better one.
TestONNXRoutingHeads is the same cascade TestONNXBaseline scores with
one arm added, so the two are directly comparable. It also checks the
Go tokenizer against the Python one, since the heads were trained
through transformers and are read through a hand-written tokenizer: a
mismatch shows up here as a score below what Python measured on the
same weights, and nowhere else. That is how the reversed word pieces
were found.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013ptwopxyo3Z2kwFckHkLvN
They run before the model because they are two orders of magnitude
faster and score better on both halves of the route. They decline
rather than clarify, so a declined turn carries on to the model and
then the classifier, which is what a box with no weights file does on
every turn. Nil heads are byte-for-byte the cascade that shipped
before this.
Measured on the 96-case fixture, classifier+ONNX either way:
intent 76.0% -> 96.9%
destination 36.4% -> 75.8%
false clarify 0 -> 1
missed clarify 8 -> 1
p50 24.5ms -> 27.9ms
That beats the gemma-4-12b cascade on both halves, 84.4% and 72.7%, at
a twelfth of its 329ms. The four remaining destination misses are all
calendar, which is the stage 0 trade V-660 flagged and the owner has
not called yet.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013ptwopxyo3Z2kwFckHkLvN
The heads trained in V-661 ran nowhere. This loads the exported graph
and reads intent, destination and clarify off one forward pass. It
declines below 0.6 max softmax rather than clarifying, so a declined
turn reaches whatever is behind it.
The slot head is exported and deliberately not read: slots already
come from the stage-2 extractor, and mapping BIO tags back to text
needs character offsets the tokenizer does not keep.
The clarify head decides on its own and decides first. It answers a
different question from the intent head, so a low intent confidence is
no reason to discard it. Reading it only above the intent threshold
cost 6 of the 8 ambiguous cases on the fixture: the word for water
reads as intent act at 0.23 and clarify at 0.98.
0.6 is the knee measured on the intent fixture: every higher value up
to 0.9 drops right answers and keeps the same two wrong ones.
The body is a fine-tuned COPY of the resident embedder and must never
replace it, because memory recall depends on that file scoring what it
scored.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013ptwopxyo3Z2kwFckHkLvN
encodeWord backtracks the Viterbi path from the end of the word and
prepends each piece, which puts them back in reading order. A second
reverse after that loop undid it. So "query: вода" tokenized to
[0 12 1294 41 12489 2] where the reference tokenizer gives
[0 41 1294 12 12489 2], and every multi-piece Russian word reached the
model with its pieces in the wrong order.
Measured on the recall fixture, same 27 cases either way:
recall@1 70.4% -> 77.8%
recall@3 85.2% -> 96.3%
answered after gate 63.0% -> 66.7%
false recall 0/5 -> 1/5
The classifier barely moves, 76.0% to 75.0% on the routing fixture,
because seeds and queries were mangled the same way and cosine survived
it. Recall is where it cost, because a stored passage and a live query
are different lengths and break differently.
The embedder id now names a tokenizer revision. Stored vectors were
written under rev 1 and no longer sit in the same space as a query
embedded now, and the model file's name never moved, so nothing would
have triggered ReembedAll.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013ptwopxyo3Z2kwFckHkLvN
Tail turns 21 to 17, the longest ride 8 turns to 4, and the two worst
replies in the corpus are gone: "спасибо" and "привет" are no longer
answered with "Сейчас 21:25. В какой день?".
MaxRides is not what fired. With the pleasantry counted as an aside the run
of asides is unbroken, so MaxSuspends reached three and ended it. Rides is
the backstop for the shape where an answer really does break the run, and
no turn in this corpus reaches it. Said so rather than crediting the new
bound.
Four rides did not move. They are asides against a question the owner never
answers, which MaxSuspends already bounds at four turns each.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013ptwopxyo3Z2kwFckHkLvN
classifyTurnRole read "спасибо" and "привет" as answers to whatever was
parked, so she re-asked "В какой день?" at a man saying thank you and
spent one of three attempts doing it. That attempt is a bound meant to end
the ride, so the pleasantry both produced the worst reply in the corpus and
paid for the privilege.
They are asides now: answered as themselves, the question resumed on the
tail, no attempt spent, one ride counted.
The set is a new closed lexicon entry, matched as WHOLE utterances. Every
token rule tried was wrong on something. "вечер" answers "это утра или
вечера?" and "нет" answers a confirm, so anything that could fill a slot
stays out. The control words stay out too, because isCancel owns them.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013ptwopxyo3Z2kwFckHkLvN
MaxSuspends did not move the number it was written for. Twenty-six of 140
turns carried a parked clarify tail before it landed and twenty-six after.
Two bounds rearm each other. An aside spends no attempt, so MaxAttempts
never reaches it. A turn reading as a failed answer zeroes Suspends, so
MaxSuspends never reaches the asides. Alternating them restores each bound
with the other's traffic. Measured on 2026-08-08: one question about a
reminder's day rode turns 7 to 13.
PendingQuestion.Rides is the same event counted without the resets. Set
once, incremented only in noteSuspended, carried across the re-park in
askRemainingGap, read by nothing that could lower it. MaxRides is 4, one
looser than MaxSuspends so the tighter statement about a run stays
reachable.
It ends the measured ride one turn early and no more. Most of that ride is
attempts, spent because classifyTurnRole reads "спасибо" and "привет" as
failed answers. Said so in the constant and in the design doc rather than
claiming a fix.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013ptwopxyo3Z2kwFckHkLvN
The third run of the same 140 turns, with the harness fix in. Sixty-eight
turns name a source.
Two findings the wording could not carry. The unfixed homelab turns are
claimed by weather and by feeds, which the destination fixture predicted.
And agenda questions are claimed by the personal boundary and by Praxis,
not by the calendar: 3 of 6, the same 3 of 6 the destination fixture and
every routing-head seed score.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013ptwopxyo3Z2kwFckHkLvN
scripts/usage-run.py read the redirect parameter "src". cmd/mavweb/chat.go
writes it as "s". So Source came back empty on all 140 turns of both
fortnight runs, and every finding in those two docs is read off the reply
wording instead of off the badge.
Re-run confirms the column now arrives: 68 of 140 turns name a source.
The two homelab misses are now direct evidence rather than inference.
"какая скорость у меня сейчас?" is claimed by weather and
"хватает ли места под новые бэкапы?" by feeds.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013ptwopxyo3Z2kwFckHkLvN
V-655 fixed four of the six turns a guessing query source claimed. The
clean win is 'что такое TCP?', which stage 0 names world and search now
answers instead of weather asking for a city. 'что я сохранил про Сочи?'
is no longer read as a capture.
The two that did not move are both homelab questions, which is the enum
and not the walk. SourceRecall, SourceNetwork and SourceAttention overlap
on every question about the box, and the destination fixture already
flagged that cluster.
The parked clarify is unchanged at 26 turns. It is dialogue state and no
query source could have touched it.
Latency is reported and not attributed. The workstation was up for the
re-run and its state during the baseline was never recorded.
Also corrects the baseline's '49 of 140 turns'. That was the sum of
occurrences. It is 41 turns.
CLAUDE.md said naming SourceWorld leaves his notes, his facts and the
personal boundary running first. The first two are true and the third is
not. The boundary is marked guesses: true, so queryWalk drops it whenever
the named destination is not recall.
That is deliberate and tested. It is what stops the boundary answering
'кто такой Линус Торвальдс?' with 'не нашла у тебя такой записи'. But it
means a destination a model wrote can take the boundary off a turn about
him, and the doc claimed the opposite.
Flagged as the owner's call rather than changed. Only the utterance leaves
the box either way.
Clarify is not a value of intent. It is a second question over the same
pooled vector: can Maven act on this at all. The eight want_clarify fixture
cases sat outside every number the heads measured, because a softmax has no
clarify class.
gen_clarify.py makes the class the corpus lacks. Every existing row was
generated FOR an intent, so every one is answerable. The router-prompt
agreement filter cannot work here, because routeGrammar has no clarify value
and a generated line always agrees with itself. A judge replaces it.
The first judge called 24 of 40 answerable rows underspecified. It judged
against a generic assistant, one that asks where about lunch. Restating
Maven's contract took that to 16 of 60, with all eight fixture cases caught.
Three seeds: 7.0 of 8 caught, 2.3 false of 88. The cascade today misses 1 and
produces 2. Confidence separates too, 0.851 right against 0.604 wrong.
BIO tags had no Maven-domain corpus, which was true of found corpora and
false of made ones. A GBNF closed over Maven's five slots plus a
substring check gives 2178 spans out of gemma-4-12b at no second call.
Three heads over one forward pass: intent 92.8%, destination 82.8%, slot
span F1 72.4% over three seeds. The slot head is free.
Epoch selection reads the intent dev slice, so it stops the slot head
about 4 points early. Recorded rather than fixed.
The 2026-08-07 week of usage was typed by hand and cannot be replayed, so
it measured a build and not a change. scripts/usage-run.py drives the same
reach from a turns file, which makes the next run a diff.
Baseline is master at beb093a: 140 turns, p50 1.6s, zero errors. Three
defects to move. A parked reminder clarify contaminates 19 later turns and
survives a day boundary. Query sources that guess claim six turns they
cannot answer, which is the class V-655 removes. And one question was read
as a capture.
Also records the slot head: gemma distils 2178 spans, three heads score
intent 92.8%, destination 82.8%, slot span F1 72.4% over three seeds.
The first 120 floor rows carried one sentence shape, because the generator
varies a topic and ambiguity is not a topic. Rotating six shapes takes the
floor 3/7 to 6/7 and the destination mean 75.8% to 80.8%.
Calendar stays 3/6 at every seed. The possessive agenda rules claim those
cases at stage 0 and name nothing, so no label reaches the head.
Also corrects the floor-case count in three files: five of the seven are
homelab, not six.
Step 3 of the routing-heads plan. Intent and destination share one masked mean
pool on e5-small. Destination scores 26/33 against 12/33 for the classifier
cascade and 24/33 for the cascade with gemma-4-12b, which is the teacher these
labels were distilled from. Recall goes 0/15 to 15/15.
Two heads, not four, and both cuts are label problems rather than GPU time.
Mood describes her own reply state and no dataset maps onto it. BIO slot tags
have no Maven-domain corpus.
The MASSIVE warm-start from step 2 is worth nothing here either. Stock ties it
on intent and leads by a third of a case on destination.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013ptwopxyo3Z2kwFckHkLvN
The training workspace labels with routeSystem and routeGrammar, and it held
its own copies. V-660 changed both. A retyped prompt drifts silently, which is
the problem llm/check_prompt_parity.py exists for on the other side.
Inert unless MAVEN_DUMP_PROMPT names a directory.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013ptwopxyo3Z2kwFckHkLvN
Numbers and the argument in docs/evals/2026-08-08-destination-model-arm.md,
pointer and the short version in CLAUDE.md. The finding worth carrying is
not the 72.7%: it is that stage 0's silence on the possessive agenda rules
used to be free and now costs four destination points, because there is
finally something downstream that would have named the calendar.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013ptwopxyo3Z2kwFckHkLvN
V-659 measured the destination at 12/33 on the classifier cascade and named
the gap: recall 0/15, because nothing anywhere names it. The model could not
help, for a structural reason rather than a capability one. Nothing in
routeSystem mentioned a Source and routeGrammar could not emit one, so there
was no string for it to write. Same shape as the Praxis reach V-517
measured at 0/12.
routeGrammar grows a source rule, closed over router.Sources plus the empty
floor. A grammar cannot emit a destination that does not exist, which is the
guarantee V-546 wants from a softmax and gets here for free. The prompt
lists the twelve in Russian, one line each, and says plainly that "" is a
normal answer to give often: two sources that can both answer means the
chain walks, and guessing is the failure mode this whole field exists to
stop.
The read-back goes through ValidSource and runs on IntentQuery alone. The
grammar already bounds the enum, but it is a request to a server that may be
running another build, and only a query reaches queryWalk.
Measured against gemma-4-12b on the workstation, same fixture, cascade with
a hash fallback: destination 24/33 (72.7%) against the classifier's 12/33,
and intent 81/96 (84.4%) which is where it already was. Recall is the whole
move, 0/15 to 14/15. The model alone scores 26/33.
Four cases the cascade loses and llm-only wins are calendar. The possessive
agenda rules claim them at stage 0 and deliberately name nothing, because
"что у меня в списке покупок" matches the same rule and naming the calendar
would take the list source off the turn. So stage 0's caution now costs four
destination points it did not cost before. That is a real trade and it wants
its own argument, not a quiet edit here.
The resident Qwen3-1.7B is unmeasured: it binds --port 0 inside the
container and no host process can reach it.
llm/check_prompt_parity.py in the training workspace compares its copy of
routeSystem to this one and will fail until that copy gets the same edit.
V-362 covers the catch-up.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013ptwopxyo3Z2kwFckHkLvN
CLAUDE.md said the destination had no fixture and no accuracy number. It
has both now: intent 73/96 and destination 12/33 on the classifier cascade,
with the per-destination split, the floor cases and the grammar drift the
labelling turned up. Anyone adding a grammar now reads that baselineGrammars
mirrors buildRouter and drifts silently when it does not.
docs/evals/2026-08-08-massive-warm-start.md was written on the V-655 branch
and parked in .task/, which git excludes, so it was one `task start` away
from being lost. It is a dated measurement and it belongs under docs/evals
whatever branch produced it. Its "destination has no fixture at all" line is
now a pointer to the file beside it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013ptwopxyo3Z2kwFckHkLvN
Twenty-eight existing query cases get a want_source and five new ones
arrive with theirs. Every label is the destination that SHOULD claim the
turn, which on the five new cases is not the one that did: they were
observed failing on the box on 2026-08-07, so the fixture fails on the day
it is written.
Seven cases assert the SourceUnknown floor, and six of those are homelab
operations. They cluster because SourceRecall, SourceNetwork and
SourceAttention overlap on every question about the box: mavpoll writes its
netdata and uptime-kuma observations into the fact store recall reads.
Naming one destination there takes the other two off a turn that needs
them. That is a finding about the enum, not a gap in the labelling.
The fixture's grammar mirror had drifted. WorldQueryGrammars went into
buildRouter with V-655 and never into baselineGrammars, so the fixture was
scoring a grammar set the daemon does not run — the exact thing the comment
above that function forbids. Adding it moved the destination number 9/33 to
12/33 and moved nothing else.
Measured classifier+onnx: intent 73/96 (76.0%), was 69/91 (75.8%). Four of
the five new cases pass and no existing case moved. Destination 12/33
(36.4%), and the split is the point. World is 5/5, because a stage 0 rule
names it. Calendar is 2/6, because the possessive agenda rules deliberately
do not. Recall is 0/15, because nothing anywhere names it yet.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013ptwopxyo3Z2kwFckHkLvN
The fixture measured the first half of a route and stopped. V-655 split a
routing decision in two, and the second half arrived with no fixture, so
Decision.Source had no accuracy number at all.
want_source is a pointer because the destination has three states and a
bare string has two. Absent is every intent but query, which never reaches
queryWalk. Present and empty is the SourceUnknown contract: name nothing
and let the daemon walk the chain, which is right whenever two destinations
can both answer and the utterance does not choose. Present and named is a
destination the route must produce.
A destination miss does not fail the case. It goes in SourceReason, never
in Reasons, so Accuracy and IntentAccuracy stay the numbers they were and
69/91 still means what it meant. SourceAccuracy is the second number, over
the labelled cases only, because a percentage of the whole fixture would be
a percentage of turns that never ask a query source.
A clarified or mis-routed case still counts in the denominator. It named no
destination and that is a miss, not a case to skip, or the denominator drops
every turn the route already lost.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013ptwopxyo3Z2kwFckHkLvN
querySources splits in two once you look at which sources over-claimed during
the week of 2026-08-07. The clean ones perform a lookup and can come back
empty: fact-by-key, tasks, list, money, calendar, notes. The dirty ones decide
by cosine against frozen seeds and then answer whatever they claimed, because
they have no lookup that could miss. Weather has no local table at all, which
is why "что такое TCP?" became "для какого города?".
So each source now carries its destination and whether it guesses, and
queryWalk takes the guessers that were not named OUT of the chain. It removes
and never reorders, which is the whole safety argument: the table's order is
load-bearing, every comment on it argues a reason between two sources, and
above all it carries "his data first, then the world". Naming SourceWorld does
not send the turn outside. It stops weather claiming a protocol on the way
past. His notes, his facts and the boundary in front of them still run first,
so a wrong destination costs nothing but the guess it prevented.
The skipped sources are recorded as never-asked with the reason, so /trace
shows a narrowed walk rather than a chain that silently shrank.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
claude
kami