kami/Maven
1
0
Fork 0
Code Issues 1 Pull Requests 1 Actions Packages Projects Releases Wiki Activity

Give the stage 0 grammar set one home (V-693)

Browse Source 
buildRouter held the real set and baselineGrammars in eval_test.go restated it
by hand, in the daemon's order, with its own comment saying so. Three test files
score against the fixture and nothing compared the two lists. They had already
drifted: BareCaptureGrammar went into the daemon with V-557 and never into the
fixture, so every routing measurement since has scored a set nobody runs. That
is the failure CLAUDE.md warns about by name, and a diff test would have caught
it one grammar late.

The list moves to router.StageZeroGrammars in internal/router/stagezero.go, with
the ordering comments, which are the load-bearing part. buildRouter and the
fixture both call it. One list cannot drift from itself.

Measured before and after on the 96-case fixture: classifier+onnx 72/96, 75.0%
intent, 33.3% destination, identical either way, and the deterministic claim and
reach hash ratchets do not move. So the missing grammar cost no measurable
accuracy. That is the point rather than a reprieve: the fixture had been scoring
the wrong set for four days and nothing could say so.

The invariants caveat is deleted, both entries, since V-692 landed the other
guard in the previous commit. The reasoning for both now sits in docs/routing.md
beside the subsystem, which is where a fix's durable record belongs.

Unrelated and pre-existing: TestONNXPersonalBoundary fails on "я рассказывал
тебе про байкал?" (personal 0.9068, world 0.9413) at the merge base too.
This commit is contained in:
claude
2026-08-11 21:02:31 +04:00
parent d8efb667c7
commit 240d53a96a
7 changed files with 98 additions and 105 deletions
Download Patch File Download Diff File Expand all files Collapse all files
CLAUDE.md
+4 -2
View File
@@ -115,13 +115,15 @@ classifier. Every stage may decline and the next one answers.
- **The classifier is the floor, not dead code.** It answers when the resident
model is off, absent, or erroring. **Any model error falls through.**
- **`baselineGrammars` in `eval_test.go` mirrors `buildRouter`.** A grammar
added to one belongs in both, or the fixture scores a set nobody runs.
- **The stage 0 set lives in `router.StageZeroGrammars`**, and both `buildRouter`
and the eval fixture call it. Add a grammar there, in the right place, and read
the comment above the line you insert after. Do not restate the list anywhere.
- **Go's `\b` is ASCII-only** and never fires after a Cyrillic letter. A Russian
pattern needs an explicit `(\s|[?!.]|$)`.
- **`PraxisGrammars()` is the only path to Praxis**, not a faster one.
- **`voice.embedder.heads_path` must never point at `model_path`.** Recall
depends on the resident e5-small scoring what it scored. Fine-tune a copy.
Refused at config load since V-692, symlinks included.
- **Routing traces are retained 14 days**, enforced on write and again on start.
- **Bump `tokenizerRev` on any change to what `encodeWord` emits**, so a
tokenizer fix triggers `ReembedAll` the way swapping the model file does.
cmd/mavend/voicewire.go
+3 -39
View File
@@ -469,45 +469,9 @@ func buildRouter(emb router.Embedder, acts router.ActMatcher, threshold float64,
llmR *router.LLMRouter, heads *router.RouterHeads) *router.Router {
cls := router.NewClassifier(emb)
seedClassifier(cls)
grammars := router.DefaultGrammars(acts)
grammars = append(grammars, router.SystemTimeDateGrammars()...)
// After the time/date rules on purpose: "какой сегодня день" is a clock
// question and must keep reaching replySystem, while "что у меня сегодня"
// is an agenda question and must not.
grammars = append(grammars, router.AgendaQueryGrammars()...)
// Same reason as the agenda rules, for the feeds: "что нового в лентах?"
// routed system and answered "пока не умею" (Vikunja #474).
// After the agenda rules, which are the narrower claim, and BEFORE the feed
// and list rules, which are not: "что такое лента" is a definition question
// and the feed rule would take it on the noun alone (V-655).
grammars = append(grammars, router.WorldQueryGrammars()...)
grammars = append(grammars, router.FeedQueryGrammar())
// The list side of the same exposure: a phrasing with no possessive in it
// ("список дел") routed system and never reached queryTasks (Vikunja #467).
grammars = append(grammars, router.TaskListGrammar())
grammars = append(grammars, router.ListGrammars()...)
grammars = append(grammars, router.ReminderGrammar())
// Before the capture marker, because "отметь" is a capture verb and "отметь
// второй пункт" is not a note. The Praxis rules are the narrower claim — a
// lifecycle verb AND an item named — so they get first refusal (Vikunja #516).
grammars = append(grammars, router.PraxisGrammars()...)
// Last, and it matches any utterance shape — its Build is the filter. An
// explicit capture marker beats the model, which called it an act and
// rewrote the task text (Vikunja #467). After the rules above because a
// marker never collides with a clock or agenda question.
// After Praxis, whose bare "закрой" claim this rule cannot reach (it needs the
// board noun), and before the capture marker, which would otherwise read
// "убери из задач купить молоко" as a new task (Vikunja #512).
grammars = append(grammars, router.TaskStatusGrammar())
// Before the capture markers, which all need an object. A capture verb
// alone is a fact with no key, and the clarify path asks for it rather than
// letting the model invent an answer (Vikunja #557).
grammars = append(grammars, router.BareCaptureGrammar()...)
grammars = append(grammars, router.TaskCaptureGrammar())
// After the capture marker, so "запиши" still wins over "расскажи", and
// last overall because it matches on the first word alone: "расскажи про
// X" is a world question the model called a fact (Vikunja #498).
grammars = append(grammars, router.NarrativeQueryGrammars()...)
// The stage 0 set, in the router package, so the eval fixture runs the rules
// the daemon runs (V-693). Order and reasoning live with the list.
grammars := router.StageZeroGrammars(acts)
return router.New(router.Config{
Grammars: grammars,
Classifier: cls,
docs/caveats/CLAUDE.md
+4 -4
View File
@@ -22,10 +22,12 @@ caveat is the pointer between them plus the trigger.
Every entry below came from the 2026-08-10 deep audit
(`docs/evals/2026-08-10-repo-audit.md`), except the last, which came from wiring
the gate the audit asked for. Three of the twenty findings are fixed and have no
the gate the audit asked for. Five of the twenty findings are fixed and have no
entry. The unauthenticated mavgpud proxy was V-673. The 20 reachable advisories
in the toolchain and `x/text` were V-682. The missing analyzers were V-694, and
what they now report is the baseline entry under V-701.
what they now report is the baseline entry under V-701. The two unguarded
invariants were V-692 and V-693, and their guards are described in
`docs/routing.md`.
| limit | severity |
| --- | --- |
@@ -41,8 +43,6 @@ what they now report is the baseline entry under V-701.
| [mavweb errors cannot be traced](transport.md#errors) | medium |
| [Fact enrichment is a 20-call serial waterfall](workers.md#enrichment) | medium |
| [A suppressed nudge is phrased anyway](workers.md#nudges) | medium |
| [heads_path may equal model_path](invariants.md#heads) | medium |
| [baselineGrammars is mirrored by hand](invariants.md#grammars) | medium |
| [Committed absolute paths pin the build to this box](config.md#paths) | medium |
| [The env example omits deployed variables](config.md#secrets) | medium |
| [The analyzers pass against a baseline, not zero](dependencies.md#baseline) | medium |
docs/caveats/invariants.md
-27
View File
@@ -1,27 +0,0 @@
# Unguarded invariants
`CLAUDE.md` names these as load-bearing. Nothing enforces either one. A rule
that lives only in prose gets broken by whoever did not read the prose. Both of
these fail silently when broken.
`tokenizerRev` and `preRouteLadder` were checked and need nothing. The rev is
baked into the embedder key, so a bump triggers re-embedding. A missing ladder
rung is observable in the decision record.
## heads_path may equal model_path [#692] {#heads}
Costs: the routing heads then score with the same graph the resident e5-small
uses, and recall degrades. There is no error and no log line, so it reads as
ordinary drift rather than a misconfiguration.
Revisit when: `deploy/mavend.json` is edited by hand, or a fine-tuned heads
graph is swapped in.
Workaround: check the two keys by eye. That is the whole guard today.
## baselineGrammars is mirrored by hand [#693] {#grammars}
Costs: the eval fixture restates the stage 0 rule set in the daemon's order,
and its own comment says so. Three test files score against it. A grammar added
to `buildRouter` alone means every routing measurement scores a set nobody
runs. `CLAUDE.md` warns about this failure by name.
Revisit when: the next stage 0 grammar is added. That is when it bites.
Workaround: add to both lists, which is what the rule already says.
docs/routing.md
+17 -5
View File
@@ -1,6 +1,6 @@
# Routing
*Last verified: 2026-08-09 @ 31b5093*
*Last verified: 2026-08-11 @ 25ed201*
How an utterance becomes a `Decision`, why each stage exists, and what every
stage has measured. `CLAUDE.md` carries the rules an agent must not break. This
@@ -137,10 +137,15 @@ below.
Go's `\b` is ASCII-only and never fires after a Cyrillic letter. A pattern needs
an explicit `(\s|[?!.]|$)`.
`baselineGrammars` in `eval_test.go` mirrors `buildRouter` and has drifted before.
`WorldQueryGrammars` was wired into the daemon by V-655 and not into the mirror,
so the fixture scored a grammar set nobody runs. Fixed by V-659, worth 3 points
of destination.
The stage 0 set lives in `router.StageZeroGrammars` (`internal/router/stagezero.go`).
Both `buildRouter` and the eval fixture call it. The daemon and the measurement
cannot disagree about which rules exist, or in what order.
It was two lists until V-693 and it drifted twice. V-655 wired
`WorldQueryGrammars` into the daemon and not into the fixture. That cost 3 points
of destination and V-659 fixed it. `BareCaptureGrammar` then did the same thing,
from V-557 until V-693 found it. That one moved no number, which is the point:
the fixture had been scoring a set nobody ran and nothing said so.
### Praxis lifecycle rules
@@ -270,6 +275,13 @@ Three rules around it, each measured:
means the heads are nil. The cascade is then byte-for-byte what shipped before
them.
Pointing it at `model_path` is refused at config load (V-692). An unloadable
weights file is not fatal, because the heads are an accelerator. A working file
in the wrong role is a different thing. The heads then score with the graph the
resident embedder scored with, and recall degrades with no log line. The check
cleans and absolutises both paths, then compares them with `os.SameFile`, so a
symlinked copy is caught too.
### The tokenizer bug the heads found
`encodeWord` in `onnxembedder.go` read every long word backwards until 2026-08-08.
internal/router/eval/eval_test.go
+5 -28
View File
@@ -255,35 +255,12 @@ func newBaselineClassifier(t *testing.T, emb router.Embedder) *router.Classifier
return cls
}
// baselineGrammars — the stage-0 rule set in the daemon's order (buildRouter in
// cmd/mavend/voicewire.go). Split out of newBaselineRouter so the claim
// measurement can run the same rules one at a time and see which of them
// contend for the same utterance, which the cascade hides by stopping at the
// first match.
// baselineGrammars — the stage 0 rule set the daemon runs, from the one place
// it is written down (V-693). It used to restate the list by hand, and by the
// time the guard was written the two had already drifted by one grammar.
// Kept as a name because the claim measurement reads it as "the baseline set".
func baselineGrammars(acts router.ActMatcher) []router.Grammar {
grammars := router.DefaultGrammars(acts)
grammars = append(grammars, router.SystemTimeDateGrammars()...)
// Same order as buildRouter (voicewire.go). The fixture is only worth
// anything while its grammar set is the daemon's grammar set.
grammars = append(grammars, router.AgendaQueryGrammars()...)
// After the agenda rules and before the feed and list rules, same as
// voicewire.go: "что такое лента" is a definition question and the feed
// rule would claim it on the noun alone (V-655). Missing here until V-659,
// so the fixture was scoring a grammar set the daemon does not run.
grammars = append(grammars, router.WorldQueryGrammars()...)
grammars = append(grammars, router.FeedQueryGrammar())
// The list side of the same exposure: a phrasing with no possessive in it
// ("список дел") routed system and never reached queryTasks (Vikunja #467).
grammars = append(grammars, router.TaskListGrammar())
grammars = append(grammars, router.ListGrammars()...)
grammars = append(grammars, router.ReminderGrammar())
grammars = append(grammars, router.PraxisGrammars()...)
grammars = append(grammars, router.TaskStatusGrammar())
grammars = append(grammars, router.TaskCaptureGrammar())
// "расскажи про X" is a world question the model called a fact, and the
// rule goes last because it matches on the first word alone (Vikunja #498).
grammars = append(grammars, router.NarrativeQueryGrammars()...)
return grammars
return router.StageZeroGrammars(acts)
}
// seedOrder — fixed iteration order over the corpus. Not cosmetic: a few
internal/router/stagezero.go
+65
View File
@@ -0,0 +1,65 @@
package router
// StageZeroGrammars — the stage 0 rule set, in the order the daemon runs it.
//
// It lives here because it used to live in two places (V-693). `buildRouter` in
// cmd/mavend/voicewire.go held the real set and `baselineGrammars` in
// internal/router/eval/eval_test.go restated it by hand, in the daemon's order,
// with its own comment saying so. Three test files score against the fixture,
// and nothing compared the two lists. By 2026-08-11 they had already drifted:
// BareCaptureGrammar was in the daemon and not in the fixture, so every routing
// measurement scored a set nobody ran. That is the failure CLAUDE.md warned
// about by name, and a diff test would have caught it one grammar late. One
// list cannot drift from itself.
//
// The order is the contract, not the membership. Each rule below says why it
// sits where it sits, and a rule inserted in the wrong place changes which
// utterances the cascade never reaches. Read the comment above a line before
// moving it, and read docs/routing.md before adding one.
//
// The classifier, the extractor, the threshold and the model arm are the
// daemon's to assemble. This function returns the rules and nothing else, so
// the fixture can also run them one at a time and see which of them contend for
// the same utterance, which the cascade hides by stopping at the first match.
func StageZeroGrammars(acts ActMatcher) []Grammar {
grammars := DefaultGrammars(acts)
grammars = append(grammars, SystemTimeDateGrammars()...)
// After the time/date rules on purpose: "какой сегодня день" is a clock
// question and must keep reaching replySystem, while "что у меня сегодня"
// is an agenda question and must not.
grammars = append(grammars, AgendaQueryGrammars()...)
// Same reason as the agenda rules, for the feeds: "что нового в лентах?"
// routed system and answered "пока не умею" (Vikunja #474).
// After the agenda rules, which are the narrower claim, and BEFORE the feed
// and list rules, which are not: "что такое лента" is a definition question
// and the feed rule would take it on the noun alone (V-655).
grammars = append(grammars, WorldQueryGrammars()...)
grammars = append(grammars, FeedQueryGrammar())
// The list side of the same exposure: a phrasing with no possessive in it
// ("список дел") routed system and never reached queryTasks (Vikunja #467).
grammars = append(grammars, TaskListGrammar())
grammars = append(grammars, ListGrammars()...)
grammars = append(grammars, ReminderGrammar())
// Before the capture marker, because "отметь" is a capture verb and "отметь
// второй пункт" is not a note. The Praxis rules are the narrower claim — a
// lifecycle verb AND an item named — so they get first refusal (Vikunja #516).
grammars = append(grammars, PraxisGrammars()...)
// After Praxis, whose bare "закрой" claim this rule cannot reach (it needs the
// board noun), and before the capture marker, which would otherwise read
// "убери из задач купить молоко" as a new task (Vikunja #512).
grammars = append(grammars, TaskStatusGrammar())
// Before the capture markers, which all need an object. A capture verb
// alone is a fact with no key, and the clarify path asks for it rather than
// letting the model invent an answer (Vikunja #557).
grammars = append(grammars, BareCaptureGrammar()...)
// Last, and it matches any utterance shape — its Build is the filter. An
// explicit capture marker beats the model, which called it an act and
// rewrote the task text (Vikunja #467). After the rules above because a
// marker never collides with a clock or agenda question.
grammars = append(grammars, TaskCaptureGrammar())
// After the capture marker, so "запиши" still wins over "расскажи", and
// last overall because it matches on the first word alone: "расскажи про
// X" is a world question the model called a fact (Vikunja #498).
grammars = append(grammars, NarrativeQueryGrammars()...)
return grammars
}