Add Laya backend and an opt-in training log #1

Merged

james merged 5 commits from laya-backend into main 2026-10-02 21:18:46 +00:00

Conversation 1 Commits 5 Files changed 18 +499 -17

james commented 2026-10-02 20:12:18 +00:00

Owner

Copy link

Adds a laya decision backend and an opt-in training log, so Centinela can run the fine-tuned Laya model in monitor mode and collect real traffic for a third training round.

What changes

• CENTINELA_BACKEND=laya: calls a self-hosted Laya server at CENTINELA_LAYA_URL. Laya speaks the Jev /v1/systemone wire format, so this reuses the TypeSafe decider; the API key is optional (CENTINELA_LAYA_API_KEY).
• Question wording is pinned: internal/decider/testdata/questions.json is the exact set the checkpoint was trained on, and a test fails if the battery drifts from it.
• CENTINELA_TRAIN_LOG: one JSON line per model call with the exact state sent, the seven answers, verdict, reasons, id, fingerprint, timestamp and latency. Off by default, owner-only file, stops at CENTINELA_TRAIN_LOG_MAX_BYTES (1 GiB). The client IP is never written. Cache hits and gated requests are not recorded.
• centinela-eval -backend laya, plus README / DEPLOY updates.

Privacy

The logged state is the full model input, so it includes the body excerpt and CRS-matched values and can hold passwords and personal data. That is intended (password-like values are one of the model's main false-positive classes); the file should live on a dedicated volume and be removed when the collection window ends.

Verification

• make vet, make test (race) and make cover pass locally; total coverage 38.5% -> 44.3%.
• centinela-eval -backend laya against laya-centinela-r2 on the RX 7900 XTX: 24/29 attacks blocked, 7/39 benign blocked, 85 / 125 ms p50 / p99.
• Ran the binary with CENTINELA_BACKEND=laya CENTINELA_WAF=embedded CENTINELA_POLICY=monitor CENTINELA_BUDGET=150ms and the training log on: model calls were recorded, cache hits and gated requests were not, no client IP or cookie in the file.

Not in this PR

• No cluster manifests. Deploying needs a Laya model service (checkpoint hosting and GPU are undecided) and a volume for the training log.
• The model misses all five recon probes in the eval and its thresholds are uncalibrated: keep CENTINELA_POLICY=monitor.

Adds a `laya` decision backend and an opt-in training log, so Centinela can run the fine-tuned Laya model in monitor mode and collect real traffic for a third training round. ## What changes - **`CENTINELA_BACKEND=laya`**: calls a self-hosted Laya server at `CENTINELA_LAYA_URL`. Laya speaks the Jev `/v1/systemone` wire format, so this reuses the TypeSafe decider; the API key is optional (`CENTINELA_LAYA_API_KEY`). - **Question wording is pinned**: `internal/decider/testdata/questions.json` is the exact set the checkpoint was trained on, and a test fails if the battery drifts from it. - **`CENTINELA_TRAIN_LOG`**: one JSON line per model call with the exact `state` sent, the seven answers, verdict, reasons, id, fingerprint, timestamp and latency. Off by default, owner-only file, stops at `CENTINELA_TRAIN_LOG_MAX_BYTES` (1 GiB). The client IP is never written. Cache hits and gated requests are not recorded. - `centinela-eval -backend laya`, plus README / DEPLOY updates. ## Privacy The logged `state` is the full model input, so it includes the body excerpt and CRS-matched values and can hold passwords and personal data. That is intended (password-like values are one of the model's main false-positive classes); the file should live on a dedicated volume and be removed when the collection window ends. ## Verification - `make vet`, `make test` (race) and `make cover` pass locally; total coverage 38.5% -> 44.3%. - `centinela-eval -backend laya` against `laya-centinela-r2` on the RX 7900 XTX: 24/29 attacks blocked, 7/39 benign blocked, 85 / 125 ms p50 / p99. - Ran the binary with `CENTINELA_BACKEND=laya CENTINELA_WAF=embedded CENTINELA_POLICY=monitor CENTINELA_BUDGET=150ms` and the training log on: model calls were recorded, cache hits and gated requests were not, no client IP or cookie in the file. ## Not in this PR - No cluster manifests. Deploying needs a Laya model service (checkpoint hosting and GPU are undecided) and a volume for the training log. - The model misses all five recon probes in the eval and its thresholds are uncalibrated: keep `CENTINELA_POLICY=monitor`.

james added 1 commit 2026-10-02 20:12:19 +00:00

Add Laya backend and an opt-in training log ... 6fb1aff956

Laya is a small encoder fine-tuned on Centinela's seven questions. It
speaks the Jev /v1/systemone wire format, so the backend reuses the
TypeSafe decider with its own URL and an optional API key.

- config: CENTINELA_BACKEND=laya with CENTINELA_LAYA_URL, _LAYA_MODEL
  and optional CENTINELA_LAYA_API_KEY
- decider: testdata/questions.json pins the question wording the
  checkpoint was trained on; the Authorization header is only sent when
  a key is set
- trainlog: CENTINELA_TRAIN_LOG appends one JSON line per model call
  with the exact state sent, the seven answers, verdict, id and
  timestamp, for re-labelling and retraining. Owner-only file, capped
  by CENTINELA_TRAIN_LOG_MAX_BYTES (1 GiB), no client IP. Off by default
- centinela-eval: -backend laya

On the probe set (PL3, strict, RX 7900 XTX) laya-centinela-r2 blocks
24/29 attacks and 7/39 benign at 85 / 125 ms p50 / p99. It misses all
five recon probes, so it is for monitor mode until retrained.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

james added 1 commit 2026-10-02 20:12:38 +00:00

README: plan to move the training log to stdout with minimization ... 83ae7eefb9

Co-Authored-By: Claude Opus 5.5 <[email protected]>

forgejo-actions commented 2026-10-02 20:13:01 +00:00

Copy link

🔎 ojo scan results

Severity	Count
🟢 LOW	1

Details (1)

Type	Severity	ID/Rule	Location	Description
misconfig	🟢 LOW	dockerfile-no-healthcheck	Dockerfile:1	image has no HEALTHCHECK

<!-- ojo-scan-summary --> ### 🔎 ojo scan results | Severity | Count | |---|---| | 🟢 LOW | 1 | <details><summary>Details (1)</summary> | Type | Severity | ID/Rule | Location | Description | |---|---|---|---|---| | misconfig | 🟢 LOW | dockerfile-no-healthcheck | <a href="https://git.colibrisec.org/ColibriSec/centinela/src/commit/ae3fa73843e6a65de1b74e8573d897e1d15d7f6b/Dockerfile#L1" target="_blank" rel="noopener noreferrer">Dockerfile:1</a> | image has no HEALTHCHECK | </details>

james added 1 commit 2026-10-02 20:17:24 +00:00

Dockerfile: set the nonroot user explicitly ... c93f7c80f7

The distroless nonroot base already runs as 65532; the scanner flags the
image as root unless USER is stated.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

james added 1 commit 2026-10-02 21:03:07 +00:00

Default the laya backend to laya-serve's port and model name ... 891d883cab

laya-serve answers on port 8000 and is fastest when asked for
"multilingual"; other model names give the same answers on a slower
path.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

james added 1 commit 2026-10-02 21:03:14 +00:00

gofmt config.go ... ae3fa73843

Co-Authored-By: Claude Opus 5.5 <[email protected]>

james merged commit 642250ac69 into main 2026-10-02 21:18:46 +00:00

james referenced this pull request from a commit 2026-10-02 21:18:47 +00:00

Merge pull request 'Add Laya backend and an opt-in training log' (#1) from laya-backend into main

james deleted branch laya-backend 2026-10-02 21:19:11 +00:00

james referenced this pull request from a commit 2026-10-02 22:00:13 +00:00

Merge pull request 'Add Laya backend and an opt-in training log' (#1) from laya-backend into main

Sign in to join this conversation.

Reviewers

No reviewers

Labels

Clear labels

No items

No labels

Milestone

Clear milestone

No items

No milestone

Projects

Clear projects

No items

No project

Assignees

Clear assignees

No assignees

2 participants

Notifications

Subscribe

Due date

The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference

ColibriSec/centinela!1

No description provided.