The gate loaded and worked on workpc and took mavwaked from 68% of one core
to 335%. onnxruntime sizes its intra-op pool to every core and spins between
runs, which an always-on gate scoring three graphs twelve times a second
provokes for the whole day. One thread per session brings it to 81%, so the
keyword costs about 13% of a core, and each graph still finishes well inside
its 80ms.
The unit now passes the three -wake- flags and the models sit beside
silero_vad.onnx in ~/.local/share/maven/models. The threshold is left at the
binary's default so there is one place to change it.
The first head woke 22 times per hour of continuous Russian speech. Two rounds
of hard negative mining over 40000 unseen Common Voice clips took that to 3.4,
and the second round recovered the recall the first had cost.
The number is crossings per hour, not accuracy per window. A 1.7% false-accept
rate on a gate that scores twelve times a second reads as small and is a wake
every few seconds.
Two things are stated rather than buried: Golos scores 2 wakes in 14 minutes at
every threshold, so a handful of real utterances sit above 0.999 and no
threshold moves them; and no negative in any table is a room recording.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013ptwopxyo3Z2kwFckHkLvN
The 2026-08-09 model swap was measured on routing the same day and E4B lost
four destination cases. Phrasing was not measured, and phrasing is the half the
owner hears.
E4B scores nudges 15/15 and the talk fixture 29/36 at p50 516ms, against the
resident model's 25/36 at p50 2.97s on the same 36 cases. lang, feminine and
address are all 36/36, where the resident model loses three on address. Every
failure is ontopic and none is a parse error.
29/36 is one case off the ceiling. The temperature sweep of 2026-08-05 found
two reply cases that fail at every temperature and named a defect in the reply
path, capping the fixture at 30/36. Both are in E4B's failure list. So the swap
costs nothing on phrasing.
One defect no check catches: in chat E4B writes "Я записала несколько идей!"
when nothing was stored. A claim to have saved something is a claim about state.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013ptwopxyo3Z2kwFckHkLvN
A ZIM title carries a leading capital and the utterance does not: /A/фотосинтез
is a 404 and /A/Фотосинтез is a 200. TitleCandidates tries the spoken form
first, so a title that begins lowercase on purpose keeps its chance.
That takes the measurement from four right to five, and the fifth is the one
that mattered. "столица Франции" returned "Список столиц Олимпийских игр"
and now returns Париж, through a title redirect the ZIM already held. The
2026-08-05 measurement named that case as the one no lexical signal could
reach. Retrieval by title reaches it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013ptwopxyo3Z2kwFckHkLvN
silero-vad replaces the energy threshold when -vad-model points at it.
Everything after the speech decision is the same state machine: the speech
hold, the silence hold, the length cap and the utterance buffer.
The model window is 512 samples and the capture frame is 480, so silero.go
re-chunks across frames. main.go claimed the two matched, which was true of
silero v4.
Stage two, the wake word, is not here. It needs a Russian keyword model that
does not exist yet.
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
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
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
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
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.
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.
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
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
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
Two untracked files left in the tree by the audit session. They are the evidence behind V-654 and several sibling tasks, so they belong on master rather than inside the PR that fixes one of them. Dated eval files under docs/evals/, so they are never edited after the day.
`SetMaxOpenConns(1)` under WAL gives up concurrent reads, and the task
asked whether that costs anything. Measured over a fixed two-second
window, a paced writer against a read loop, three runs per cap:
reads do not queue. Four connections buy 70µs at p50 on a turn that
spends 1.19s in the resident model, and write throughput more than
halves. A 19ms worst case also cannot be the source of the 2.7s router
figure, so that line of enquiry is closed.
What the cap cannot survive is a long-lived transaction. It holds the
only connection, so a second read never completes: two seconds and
`context deadline exceeded`, against 1ms at a cap of four.
`Store.DB` handed out exactly that transaction. It had been there since
the initial commit with no production caller, and its comment described
a loop that never materialised. Its one user was a test helper reading
`delivery_attempts` by raw SQL, which `ListDeliveryAttempts` has covered
since V-390. So the cap stays and the seam goes, and the hazard is gone
by construction rather than by documentation.
`internal/store/conncap_test.go` stays as the standing measurement,
skipped under -short. The comment at the cap and the one in
`internal/ipc/server.go` that leans on it now state the invariant and
cite the numbers.
Measurement: docs/evals/2026-08-07-store-connection-cap.md
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
A repo-wide sweep on 06-08-2026 at 06c1cf2. Three docs, three tasks.
docs/plans/24-no-deadline-on-the-turn-path.md (V-638). Nothing between a
mavweb handler and llama-server can be cancelled, and one hop has a timeout.
Replier takes no context, the ipc client sets no conn deadline and checks ctx
once, and the ipc server dispatches under Background. Four commits, and the
pattern to copy is already in internal/voice/client.go:101.
docs/plans/25-the-two-boot-paths.md (V-639). The passkey-unlock path starts
seven workers outside the WaitGroup that shutdown waits on, shadows that
WaitGroup at main.go:529, and builds a daemonAPI with no nexus and no
getMCPServers. Latent, because db_key_env means the box boots unlocked.
docs/evals/2026-08-06-routing-trajectory.md (V-464). The deterministic path
and the cascade now score the same 69/91, and the cascade has not been
re-measured since V-626 and V-627. Either the model still earns its place or
it is costing 1.17s a turn for nothing. Dated, so it is not edited later.
Committed with --no-verify, on the owner's instruction of 06-08-2026. The
pre-commit hook refuses master and the alternative was three PRs for three
markdown files. Markdown is already exempt from the size cap for the same
reason: docs land as one batch.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Three tangled defects, fixed together because each one hid the others.
DefaultActMatcher matched an exact English prefix and internal/tool.Matcher
delegated straight to it, so no Russian utterance could reach a tool: 55 of the
69 lines in models/seeds/act.txt routed to IntentAct and fell to proposeGap.
Tools now carry spoken aliases from deploy/mavend.json, matched as exact leading
tokens, longest phrase first. Config data, not a stem pattern in code. The
comment claiming "the production matcher is fuzzy" was false and is gone.
Seven lines were exact duplicates inside models/seeds/query.txt, each one a
second identical vector double-weighting its region.
"как дела у сервера" carried both a query and a system label. It leaves
system.txt, because replySystem's stats arm answers "системная статистика пока
не подключена." and always did. The mode inventory records that shape as
act.tool.hoststats rather than a system mode.
Fixture unchanged at 69/91, and it cannot see any of this: no host-stat case and
no Russian act in it. TestActMatcherAliases is the coverage.
docs/evals/2026-08-06-russian-acts-reach-tools.md has the numbers.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0117tgnmbgZpHVV3XSNw8Qua
reminder_verbs held five words and none named an alarm, and ReminderGrammar
did not read the set anyway — it carried the literal напомни|remind me. So no
part of the cascade recognised разбуди, and the three alarm cases in the
fixture went to fact and act at over 0.89.
The lexicon addition alone moved nothing, measured at 66/91. Every consumer
reads the set after a reminder route already exists. Building the grammar's
alternation from the set is what scored: 66/91 to 69/91, three cases gained,
none lost, and each alarm now carries its time slot.
Longest-first ordering in the alternation is load-bearing. Go's regexp
alternation is leftmost-first, so напомнить after напомни would never match.
Found while training the V-546 intent head, where the same three cases went
to system.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0117tgnmbgZpHVV3XSNw8Qua
The classifier learns models/seeds and the router is prompted with
routeSystem, and they held different definitions on 80 lines. Sensor and
host state was system in the seeds and is query in the prompt, which is the
V-374 edit the seeds never received. World questions were chat, written
before external search could answer them.
64/91 to 66/91 on the fixture. en-sys-002 and ru-query-011 gain, nothing
regresses, clarify counts unchanged.
The third disagreement is measured and rejected. Dropping the eight bare
reminder verbs scores 65, because a centroid is a shape to be near and the
bare verb phrase is part of that shape. A seed file and a prompt have
different jobs there.
The plan calls the labeled set the whole project and names the stage 0
grammars as the label functions. cmd/labelgen runs them, the real ones in
buildRouter order, so a rule change moves the training data with it.
Gemma labels the rest at 334ms/call with nothing unparsed, which matches the
plan's estimate. It agrees with the seed files on 197/277, and reading the
disagreements is the finding: the seeds and the router prompt hold different
definitions of system, of a world question and of a bare verb. V-626.
V-586 reported 64/91 on the RU routing fixture, unchanged. That number does not
bear on the change: the fixture holds three fact cases and all three miss on
intent, so DefaultFactParser is never reached and any parser edit scores as
"unchanged".
So the parser gets its own corpus, 91 cases, scored against BOTH
implementations — the closed classes that ship and legacyFactParse, a verbatim
copy of the substring parser at 0445693, frozen in the test file so the
comparison reruns. True positives 35/40 to 39/40, misfires rejected 8/15 to
14/15. The rewrite wins every case anyone argued about.
The third case class is the point: 36 sentences a person would plainly say
whose word is in no lexicon set. The old parser caught 3 by accident, the new
one catches 0. "ем суп", "вздремнул", "помылся", "перекур", "i napped". A
silent miss is this parser's worst failure mode and the corpus sizes it.
Two defects recorded rather than fixed, since this branch measures: "допил
воду" misses because the dictionary lemmatises допил to допилить, the same saw
collision drink_verbs carries пил for; and the oblique cases of душ go with the
exact match that keeps the soul out.
The LLM arm the original commit skipped is run here against gemma-4-12b on the
workstation at 192.168.1.105:8080 — it was reachable all along, the failure was
the shell's HTTP_PROXY. cascade+llm 85.7% to 86.8%, one case, same failing set,
variance. Full write-up in docs/evals/2026-08-06-fact-parser.md.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0117tgnmbgZpHVV3XSNw8Qua
Markdown only, and the pre-commit hook refuses master, so --no-verify.
All four of the owner's cases hold. The invented clock is gone and the agenda
question mid-flow now reaches the calendar. Two new defects, V-592 and V-593.
The pinned acceptance transcript is superseded by the rule the owner ruled
after writing it, and the doc says where they disagree.
Measured on the box against master at 70b32af. Six of nine scenarios answer the
interrupting utterance, which is what V-560 bought. Three fail: an agenda
question completes the reminder at the current clock, a note stated mid-flow is
dropped in silence, and a repair fragment writes a note nobody dictated.
V-572 is confirmed to hold for every spoken time form, not only the one it was
measured on, including the spelled-out девять.
--no-verify: the pre-commit hook refuses master, and this is an unattended run
where the merge target is master by the owner's instruction.
Item 2 was blocked because the resident llama-server binds --port 0 inside the
container, so no host process can reach it. Cleared by taking the first of the
three ways out the task listed: a second llama-server on the same gguf, on a
fixed host port.
Cascade + resident model scores 75.8% full and 80.2% intent-only at p50 1.19s
and p95 1.65s, on the fixture as it now stands at 91 cases. That is a new
baseline rather than a movement: 14 cases were added since the 77-case number
in CLAUDE.md.
The model alone scores 37.4% full against 61.5% intent-only. The gap is slots,
not routing. Every reminder case leaves the time to the daemon, which is what
the contract asks of it, and the cascade fills them.
Item 3: the ~6s figure recorded in the task was one sample through the whole
of POST /api/chat, not the router, and is not comparable.
Item 4 is still not run. Killing the resident llama-server needs a permission
this session does not have, and it now has a second half anyway, since with the
workstation up only killing both proves the classifier answers.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SoL7EBdYC5Mhz3DJd49GJy
V-405 measured reach with the classifier only, and the LLM router is the
deployed default, so 16/30 was the floor rather than the shipped behaviour.
TestReachWithLLMRouter scores the same 30 cases with the model, gated on
MAVEN_LLM_URL like TestLLMRouterBaseline.
The open question was whether the model writes a literal Praxis capability
into the fn slot and reaches a service the classifier structurally cannot. It
does not. Praxis is 0/12 with the model alone, exactly what the classifier
alone scores, and all twelve fail the same way: local, empty fn. Nothing in the
router prompt names a Praxis capability, so there is no string for it to write.
So V-516's stage-0 grammars are the only path to Praxis, not a determinism
argument. Through the cascade the model scores 28/30 with praxis 11/12, one
point above the classifier baseline. Hexis is 10/10 either way.
Overreach is 1 in both configurations, under the 4 the harness asserts.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SoL7EBdYC5Mhz3DJd49GJy
Four paths on the same 72 RU cases with the daemon's own router prompt.
Text in scores 90.3% intent-only. Whisper then route scores 84.7% at p50
1372ms. The workstation transcribing then routing scores 83.3% at p50 997ms.
One call from audio straight to a route scores 54.2%.
The one-call number is not a transcription failure. Four clips it
transcribes word for word it then routes wrong or refuses, and the emitted
slot holds the tail of the sentence with the interrogative head gone. A
3.5k-character classification prompt and an audio part compete for
attention, so transcription needs its own call with a short instruction.
The two speech-to-text paths differ by one case, which is noise on 72, so
the choice is latency and transcript quality. The workstation wins both.
mavgpud.json on the workstation is restored to its text-only args.
Verification, as the task asked. Drove что такое фотосинтез through
/api/chat with the search reachable, with the container stopped, and with
the host blackholed. Kiwix claims the turn in both failure cases, and a
stopped container costs nothing: DNS fails and the ZIM answers inside the
same second.
The blackhole is the case that hurts. The search waited its full 8-second
budget before the ZIM was asked and the turn took 15.4s against 3.5, which
he sits through with nothing being said. So the connect phase alone is now
capped at 1.5s. A reachable instance that is merely slow keeps the whole
budget, because it is fanning out to real engines.
The RU Wikipedia ZIM is on the box (owner moved it into the kiwix zims
dir), and kiwix-serve picked it up. A Cyrillic question now searches
book_ru verbatim and skips the RU->EN rewrite: that rewriter is the
workaround for an English book, and against a Russian one it is a
translation of his own words back at him. Catalog names come from the
filename, not the <name> field — books.name=wikipedia_ru_all returns
nothing.
Measurement in docs/evals/2026-08-05-kiwix-offline-fallback.md. The RU book
answering a driven turn needs a rebuild and is not verified yet.
V-539 said SearXNG claims every world question, including invented terms,
so Kiwix is never reached. Measured today against the configured instance:
seven of eight invented Russian questions now return zero results, and
Response.Empty() already passes those to the ZIM. The premise moved with the
upstream engine set in three days.
The three quality signals the task named were recorded per query and none
separate the sets. Token overlap is zero for the one bad claim and also zero
for "столица Франции", whose answer is Париж. Empty snippets never fire,
because ParseResponse already drops a hit with no text. SearXNG returned no
corrections or suggestions even for the query it silently respelled. So no
threshold is built: it would cost a real answer to save one invented word.
What ships is the second half. The claiming query source crosses the IPC seam
on ipc.ChatReply.Source and renders as a badge beside the reply on /chat. It
rides the context rather than a return value, because handleText answers every
reach through one string and the mic, telegram and the web all share it.
Chat now returns ChatReply instead of a bare string.
Full -race suite green.
Russian names a half hour by the hour it is entering, in the genitive, so
"половина восьмого" is 07:30 and never 08:30. Neither date parser read that
shape, so the reminder parsed to nothing.
rewriteHalfPast runs in front of the token pass in SpellOutDigits, so the
python parser and the stub both see "в 7:30". It also reads the contracted
"полвосьмого" and the quarter-to shape "без четверти восемь", which counts
from a cardinal and is 07:45. Minus one is in one place, clockHourBefore, with
twelve rather than zero before one.
Ordinals eleven and twelve added to the lexicon, because a clock reaches them.
Minutes a spoken clock does not use are left alone: a guess here is a missed
dose.
Classifier + onnx over the routing fixture 58/82 to 62/87, three new cases,
none regressed. Python dateparser is not installed on this host, so only the
stub was measured. See docs/evals/2026-08-05-half-past-hours.md.
The router names a position ("2", "last") or a demonstrative ("this"),
because only the daemon has the list. surfacedItems records the item ids
she read out, in the order she said them, and only for items she could
actually say: one Praxis returned without a title has no position in what
he heard.
resolveSurfacedPosition maps the reference to an id before dispatch, and
its second return says whether the turn is still Praxis's. A position that
names nothing keeps the turn and clears the slot, so the capability asks
which пункт -- he said "второй пункт" and deserves to hear there is no
second one. A demonstrative that resolves to nothing gives the turn BACK,
because "я это сделал" was probably never about a пункт. "это" also needs
the list to hold exactly one item: pointing at one of five is a guess, and
a wrong guess here transitions the wrong item.
No TTL, unlike the pending confirmation. A stale position resolves to an
item Praxis will report as already acknowledged, which is a harmless
answer, where a stale confirmation would execute something.
Measured, make eval-reach, classifier + ONNX: 16/30 -> 27/30 overall,
praxis 0/12 -> 11/12, lifecycle 0/5 -> 5/5, attention 0/7 -> 6/7, hexis
and none unchanged, p50 20.6ms -> 16.5ms. make eval-router: 60/84, 0 false
clarifies, and no failure in that list comes from a stage-0 decision.
Details and the two judgement calls in docs/evals/2026-08-05-praxis-reach.md.
Four temperatures, three runs each, on the 36-case talk fixture. 0.40 leads the
mean by 5.6 points and the spread inside one temperature is 11, so three runs
cannot tell the effect from the noise. The default stays 0.7.
The result worth having is not about temperature. Sixty failures across the
twelve runs are one parse error, every one of them in the reply family, two of
them in all twelve runs. That is deterministic and caps the fixture at 30/36.
Filed as V-537.
Praxis reach is zero on all twelve cases under both embedders, and it is
structurally impossible rather than merely weak: handlePraxisAct dispatches
on fn equality, and no praxis alias can ever enter the fn slot, because that
slot is filled from the deployment's tool allowlist.
Hexis reach is 9/10. All three services are up and answer; both praxis feeds
are empty, so the gap is entirely on Maven's side of the wire.
The swap itself already landed: deploy loads
models/embedder/multilingual-e5-small/model_quantized.onnx, and
onnxembedder.go grew EmbedQuery/EmbedPassage with the query:/passage:
prefixes the model was trained with. What was missing is the half of #371
that says "re-run make eval-recall and compare against the recorded numbers",
so nothing in the repo says whether it worked.
It worked, on every axis at once. recall@1 60.0% → 70.4%, recall@3 80.0% →
85.2%, answered after the gate 48.0% → 63.0%, false recall 1/5 → 0/5, and
latency p50 59ms → 23ms because the quantized file is 118MB against the 470MB
fp32 one the old config loaded. The guitar-chords note no longer beats the
docker-logs note.
One premise of the task did not come true and the new doc says so. #371
expected a better retriever to separate the score distributions and make
query_min_score tunable. It did not: right-first top-1 runs 0.791-0.890 and
must-stay-silent runs 0.795-0.835, still overlapping, just higher and
tighter. The margin separates them instead — 0.024 median against 0.002 — and
0.008 is the knee where all five silent cases are silenced at no cost. The
score gate is close to inert now; the margin is the live dial. Neither is
changed here, since #412 is where a sweep belongs.
docs/evals/2026-08-04-recall-e5-small.md is the dated measurement.
rearchitecture.md's "upgrade MiniLM → bge-m3 later" is now done and says so,
CLAUDE.md names the retriever and the prefix rule where it already promises
the embedder never leaves homesrv, and the Makefile comment points at this
eval instead of the one that asked for the swap.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
#496 asked to skip the veto when the question and the hit are in
different scripts, so an English question stops losing a Russian note.
Measured first: the fixture has no cross-language case, and en-hard-024
is an English question against an English note. Both proposed fixes are
no-ops.
What the veto actually does on the fixture, with the real embedder: it
costs en-hard-024 and buys ru-silent-029. Pass count is 22/32 either
way; false recall is 0/5 with it and 1/5 without. The two cases are one
lexical class, so no rule cheap enough for RecallAllowed separates them.
Accepts the loss and pins both sides in a test.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The forwarded log named the cause in one line: the prompt cache limit
defaults to 8192 MiB. llama-server saves the full KV state of every idle
slot it evicts, 112 kiB per token, so RSS climbed about 170MB per
distinct prompt until the deployed server held 7.9GB for a 1.1GB model.
Measured on homesrv today, uncapped versus `--cache-ram 512`: RSS
plateaus at 932MB from the fourth distinct prompt instead of climbing.
The task's leading guess was wrong. `-ngl 99` costs almost no RSS,
because RADV keeps device memory outside the process. Numbers and method
in docs/evals/2026-08-03-llama-prompt-cache.md.
`-c 4096` is untouched. The knob is `phraser.cache_ram_mib`, unset means
512, negative passes no flag for a llama-server too old to know it.
The deploy still runs the old image, so the box keeps its 8 GiB default
until mavend is rebuilt.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
"что я говорил про бэкапы?" is his data by definition, and nothing outside the
box has ever heard him say anything. The boundary matched possession words only,
so the question walked past it into SearXNG and came back answered out of a Habr
article about somebody else's backups.
A speech-verb marker class was written first and dropped. Russian gives every
verb a dozen surface forms and the "как я говорил, ..." preamble list has no end,
so each form the lexicon missed was one more question reaching the world, and a
missing verb looks exactly like no bug.
The boundary now embeds two frozen seed sets and scores the turn's own query
vector, already computed upstream, against both. Nearest side wins. The
possession markers stay as the offline floor for a handler with no embedder.
19/19 held-out utterances correct against multilingual-e5-small; see
docs/evals/2026-08-03-personal-boundary.md. The live probe on the deployed box is
not done.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
claude