Commit Graph
28 Commits
Author SHA1 Message Date
Ilker ManapandClaude Opus 5.5 c09d5ad4ee Outside services read answers redacted and whole; a card's own lanes
CI / test (push) Successful in 9m16s
From the second review.
- An output service reads the answer with the values restore put back
  taken out again (guard.Subject.Restored), as an input service reads the
  request redacted.
- With n above 1, the service reads the first choice once it is whole,
  not at whichever choice finished first.
- Restored values go into an answer asked for as JSON (response_format)
  escaped for it, as into tool-call arguments.
- The embedding failure is logged once a minute through an outage, not
  once.
- A card's own lane count and fastest generation come from the kernel
  (max_link_width, max_link_speed): the driver's width.max is what card
  and slot allow together, so an x16 card in an x8 slot read x8 of x8 and
  was never said.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-29 13:22:39 +01:00
Ilker ManapandClaude Opus 5.5 5eac62d556 Guardrail services merged the stricter way; the request log keeps no restored value
CI / test (push) Failing after 7m44s
From review.
- A service two policies name is asked the stricter way of the two, both
  directions and block over flag, rather than as the first one says.
- A service answering null where it says unsafe gave no answer, not safe.
- With restore: true, the kept answer had the request's redacted values in
  it beside the redacted request; they are taken out again before it is
  kept.
- guardrails.md says what fail: open lets through, and that of n answers a
  service reads the first.
- The similar cache's embedding failures are logged once a minute, not per
  request; the page's test reads the gateway's local MCP handler under its
  lock; org/name:Q4_K_M is not taken for a Hugging Face id to read as
  safetensors.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-29 11:03:57 +01:00
Ilker ManapandClaude Opus 5.5 54c6b66cb9 Fixes from review: the similar cache, restore, outside services, latency
CI / test (push) Successful in 9m48s
- The similar-question cache embeds a question as its project would: a
  project kept in the cluster has external fallbacks refused, and an
  embedding model outside the cluster is not asked at all. The vectors are
  compared outside the cache's lock. The question's bound is in characters.
- The restorer puts values back in streamed tool calls' arguments (escaped
  for the JSON they are), completions' text and reasoning, each held across
  events on its own, and releases a held tail with the finish event. A
  whole answer is walked as JSON, so a value with a quote no longer breaks
  tool-call arguments.
- An outside guardrail service reads the user's turn as redacted, and a
  streamed answer once rather than at the finish chunk and again at the
  end. guardrails.md says it can stop a stream only under buffer_all.
- A replica's smoothed time to first token is forgotten after five minutes
  without an answer: one found slow and then given nothing was never tried
  again.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-29 08:15:43 +01:00
Ilker ManapandClaude Opus 5.5 783dc9e9f6 A guardrail policy can ask someone else's service
CI / test (push) Failing after 7m39s
A site may be required to use a particular guardrail: Azure Content Safety,
Lakera, NeMo, a classifier of its own. A policy's services asks one over
HTTP — a body template given the text, header values from files read each
call, dotted paths to where the answer says unsafe and why — on input,
output or both, to block or flag. On input it reads the latest user turn
before the judge; on output the whole answer once at its end. A service
that does not answer is what the policy's fail says. It is told the text
only, never who sent it.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-25 16:49:56 +01:00
Ilker ManapandClaude Opus 5.5 81685ecd75 A request's redacted values are put back in its answer
CI / test (push) Successful in 8m49s
With restore: true on a guardrail policy's input, what redact rewrote as
placeholders comes back in the answer: the model reads "[EMAIL_1]", the
client reads the address it sent. A streamed placeholder split across
events is held until it is whole, and what is held is sent at the end.
The swap runs before the output filter, so the output rules judge the
answer as the client reads it and a placeholder they make of a value the
model invented is never taken for the request's own. The map lives in the
Verdict for the request only; nothing stores, logs or traces it.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-25 16:41:53 +01:00
Ilker ManapandClaude Opus 5.5 85f135311e How fast the judge is, measured, and said at install
CI / test (push) Successful in 5m29s
ctl guard judge check, and POST /api/v1/guard/judges/check, time three
judgments of a 280-token prompt, then a burst at once, against the
policies' wait: ok, tight or slow, about how many requests a second can
be judged, and what to do. server init runs it once the judge is up and
says a machine too slow before the next steps; the wait is not changed
without asking. BACKLOG reordered by what goes wrong where installed.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
2026-10-20 22:35:19 +02:00
Ilker ManapandClaude Opus 5.5 5a4b022dcd A lost store machine is given up on in seconds, and a new leader waits for its judge
CI / test (push) Failing after 30m6s
Measured on three VMs: cutting the store primary's power cost about 45 s of
hanging requests. New connections tried the dead member first with no
connect timeout, and Patroni kept its 30 s leader key. Cutting the leader of
both cost a further 17 s of 503s while the new leader knew no judge.

The store dials each member for 3 s and closes connections to a machine that
is gone within about 10 s (keepalive, TCP_USER_TIMEOUT). Patroni runs with
ttl 20, loop_wait 5 and retry_timeout 5, applied in place to existing
clusters through its API. Judges that all fail at once are asked again
every 250 ms while the policy's budget lasts, and reported once.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
2026-10-15 20:33:13 +02:00
Ilker ManapandClaude Opus 5.5 b0273ffb5f Qwen3Guard is the judge, and judges on a GPU and on CPUs are asked in turn
Llama Guard 3 1B could not tell ordinary coding requests from harmful ones:
on the PoC it scored "continue" 0.51, "continue the code above" 0.89 and
"how do I kill all child processes in bash" 0.92 unsafe, against 0.97 to
1.00 for harmful requests, at Q4_K_M and Q8_0 alike, with its own template.
A site had "continue" refused. Qwen3Guard-Gen 0.6B gave the ordinary ones
0.15 at the most and the harmful ones 1.00, in 0.8 GB on CPUs.

- format qwen3guard: its prompt written out as its template writes it, for
  requests and answers, and its Safe / Unsafe / Controversial verdict read;
  controversial counts as the policy's judge.controversial says, safe by
  default.
- judge.models: judges asked in turn, the first that answers deciding; a
  failure is recorded against the judge that failed. guard judge use sets
  the order, replacing categories none of the judges names.
- known judges (Qwen3Guard 0.6B/4B/8B, Llama Guard 1B/8B): guard judge
  enable serves any of them, --gpu the 4B on the roomiest GPU class as a
  standard model; the leader fetches and serves one a policy names that the
  catalog lacks.
- server init installs qwen3guard-0.6b with the baseline's categories.
- a judge reads a long text by its start and end, which fit its context,
  rather than overflowing it and refusing the request under fail: closed.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
2026-10-08 10:22:47 +02:00
Ilker ManapandClaude Opus 5.5 4721b8e353 Llama Guard 3's verdicts carry their hazards' names without a categories list
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
2026-10-06 12:44:53 +02:00
Ilker ManapandClaude Opus 5.5 249a91aeec Guardrails turn off and on; a judge's refusal says which hazard; ctl guard judge ask
guard.enabled (unset is on) makes every policy check nothing while false,
live: llmgw ctl guard off/on set it in the cluster's settings, and ctl
status, ctl guard list and the guardrails page say when they are off.

A refusal by a judge names the hazard it saw - judge: S6,Specialized
Advice - and the verdict is kept with the request in a new usage column,
guard_rule, shown by ctl request, the request log and the guardrails page.

llmgw ctl guard judge ask has a judge read a text, past the cache, and shows
its raw answer and how it is read; the endpoint is kept out of the audit so
no text is recorded.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
2026-10-06 12:41:08 +02:00
Ilker ManapandClaude Opus 5.5 a8bb417618 The guardrails' judge has its own command, and its trouble is shown
llmgw ctl guard judge status lists each judge model the policies ask:
in the catalog, as a judge, replicas ready, its last failure to answer, and
what is wrong in words. enable serves the missing one through model serve -
Llama Guard 3 1B on CPUs, role guard, under the name the policies give it -
with --replace for a model of that name that is not a judge; the console has
it under Do. GET /api/v1/guard/judges, and the summary's judges, put the
problem on ctl status and the console's Overview.

A Llama Guard file served from the mirror page is added as a judge, under
the name the policies ask, rather than as a model clients use the policies
never find. ctl says plainly when the API answers with a page.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
2026-10-06 10:47:37 +02:00
Ilker ManapandClaude Opus 5.5 6289b01ed6 The cluster's settings live in the store, and every change takes effect
Everything in server.yaml that is not one machine's own is now a versioned
document in the store (cluster_settings), the same on every control plane.
The first start against a store imports the file's cluster keys, with the
baseline guard policy when there is none; after that those keys in a file
are ignored, with a warning.

Each control plane polls the document's version and its own server.yaml's
hash. A change is validated and applied: guard live (Gate.Replace), anything
else by a graceful restart that systemd brings back, standbys first and the
leader 30s later. A version that does not load on a machine is not applied;
at start the newest one that loads is used, so a bad edit cannot keep every
control plane down.

llmgw ctl config show/get/set/unset/edit/history/rollback, the admin API
under /api/v1/settings, and the console's Settings screen, which layers the
store's document over the machine's file, write it. Commands beside the
control plane that read cluster keys (init, node add, server add, backup,
upgrade) read the layered configuration; init --bundle writes offline to
the store once it holds the settings.

A standby that becomes the leader re-reads models, engine choices and
facility holds from the store.

Retires appendGuardIfMissing: the import does it.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
2026-10-05 15:25:23 +02:00
Ilker ManapandClaude Opus 5.5 9136bc71c2 More of OpenAI's API: Responses structured output, images and previous_response_id; rerank, score and audio passed through; moderations by the judge
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
2026-10-04 22:56:25 +02:00
Ilker ManapandClaude Opus 5.5 c553f2041a The baseline fails closed with two judges, and a judge not answering is a notification
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
2026-10-04 19:19:49 +02:00
Ilker ManapandClaude Opus 5.5 dbcecd1e85 Judge an answer with Llama Guard 3's own prompt, as a completion, not a conversation ending on the answer
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
2026-10-04 15:35:57 +02:00
Ilker ManapandClaude Opus 5.5 742d6633e2 Read a judge's verdict after the answer llama.cpp goes on from, and log a verdict that cannot be read
Found on the PoC: llama.cpp takes a conversation ending on the assistant's
turn as a reply to continue, so Llama Guard answered "Rome.safe" and every
answer judged was served unjudged.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
2026-10-04 15:33:23 +02:00
Ilker ManapandClaude Opus 5.5 62038484d6 Never lose a word held back at one release when the next has nothing to rewrite
Found on the PoC: a streamed list came out as "1. 2. 3." — each item cut at
one release and dropped at the next, which sent the engine's own events.
A property test now cuts an answer every way and checks it arrives whole.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
2026-10-04 15:24:22 +02:00
Ilker ManapandClaude Opus 5.5 2cc52fd1a7 Read an unwritten guard mode as enforce when merging, not as the weakest
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
2026-10-04 15:22:07 +02:00
Ilker ManapandClaude Opus 5.5 47e847423d Guardrail metrics and alerts: checks, latency, judge failures, fail-open, redactions, stopped streams, tenant blocks
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
2026-10-04 15:11:50 +02:00
Ilker ManapandClaude Opus 5.5 43f6a585a6 A judge model on the cluster: role guard in the catalog, judge in a guard policy
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
2026-10-04 15:09:31 +02:00
Ilker ManapandClaude Opus 5.5 dc788e5cda Opt-in capability checks and structured-answer validation in guard policies
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
2026-10-04 12:39:40 +02:00
Ilker ManapandClaude Opus 5.5 ec737cffb5 no_store: refuse conversations, notes and models outside the cluster under a policy that keeps nothing
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
2026-10-04 12:36:02 +02:00
Ilker ManapandClaude Opus 5.5 13e6f19301 Attach guard policies to scopes: store, admin API and ctl guard; run the gate when server.yaml defines policies
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
2026-10-04 12:34:47 +02:00
Ilker ManapandClaude Opus 5.5 b4bdd250af The gate: judge requests and answers by the policies in force, buffered or a sentence at a time
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
2026-10-04 12:31:55 +02:00
Ilker ManapandClaude Opus 5.5 ccdc07fe0c Guard policies: a library in server.yaml, merged so nothing below loosens what is above
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
2026-10-04 12:27:23 +02:00
Ilker ManapandClaude Opus 5.5 e3ef5b18c1 Test national phone numbers and bare digit runs
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
2026-10-04 12:25:32 +02:00
Ilker ManapandClaude Opus 5.5 ee33524599 Find personal data and secrets by validated pattern, and prompt injection through disguises
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
2026-10-04 12:25:20 +02:00
Ilker ManapandClaude Opus 5.5 e701f92806 Add the guard interface, and filter hooks in the proxy for buffered and streamed answers
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
2026-10-04 12:16:00 +02:00