# CLAUDE.md — rules to keep front of mind

Plant Operations Assistant for the **Waterloo Road Pump Station (WRPS)**, deployed onto an
existing, live, shared Docker host. Full detail: `BUILD-AI-CONTAINERS.md` (the build spec) and
`YAU_Linux_Host_Onboarding.md` (the host brief). **Where the two conflict, the host brief wins.**

## The three lines this system does not cross

1. **No instructions for safety-critical actions.** Procedural questions get document identity,
   revision, effective date and verbatim prerequisites — never synthesised steps.
2. **No recommended setpoints or operating parameters.** Advisory questions get evidence, ranges,
   outcomes, documented limits, then an explicit deferral. Never a number presented as the answer.
3. **No answers outside the evidence.** Zero rows means "no records found", never a plausible figure.

These are **code paths, not prompt instructions.** A safety rule living only in a prompt is not
implemented. The classifier runs first; each class has a Pydantic contract validated in Python
after generation and before returning. Contract failure -> regenerate once -> error. Never return.

When class is uncertain, choose the **more restrictive** class. Procedural beats Reference.
Advisory beats Historical. Partly-advisory is advisory.

## Host rules — inherited, non-negotiable

- Growing data goes on **`/datadisk`**, never `/`. Root is 62 GB and has hit 100% before.
- **No published host ports for anything we build.** Join the external `proxy` network, reach it
  through Caddy. Some existing containers do publish (`caddy`, `wireguard`, `mosquitto`,
  `chirpstack-gateway-bridge`, `openplc-runtime`) — they carry non-HTTP protocols. We do not.
- **Never omit `import authelia`** from a Caddyfile block — it silently makes a service public.
  One deliberate exception, applied 2026-08-28 at the customer's direction: `ai.yokogawa.tech`
  admits `remote_ip 10.0.0.21` (`cicore1`) unauthenticated and 403s everything else. It does
  not generalise — do not copy it, and do not widen it. `api.yokogawa.tech` stays gated.
- `~/authelia/configuration.yml` is **root-owned**: `sudo`, back up as `.bak-<purpose>-<date>`.
  Restarting Authelia **logs out every active user** — announce it first.
- AD group membership must be **DIRECT**. Nested membership silently fails.
- `pg-ai` and `cube` are pinned — **do not add them to Watchtower's update list**.
- **`openplc-runtime` is live control for this demo.** Never restart, update or reconfigure it as
  a side effect of AI work. Do not copy its published-port pattern, and never change its
  `10.0.0.17` binding to `0.0.0.0` — that binding is what keeps unauthenticated Modbus off the
  internet.
- Never install on, write to or restart `cicore1`. Never exceed read-only on `imh`.
- **No secrets in Git, compose files, logs or error messages.** `0600` env files under `~/ai/`.

## Verification

`docker ps` showing "Up" is **not** proof. `curl -sI` the public URL, expect `302` to the auth
portal, and read the container logs. Prove each layer without the LLM first: prove Cube returns
the right number by hand, prove retrieval finds the right procedure by hand, *then* wire the agent.

## Working conventions

- Prefer additive changes. Snapshot config before editing.
- **Do not invent schema.** `imh` is pending — inspect it, or ask. Fixtures are marked as fixtures.
  Before connecting it, work through `db/README-standin-historian.md` — the stand-in's shape is
  reasoned from the SCADA config, but **every SQL Server detail in it is a guess**, and two of the
  twelve listed assumptions fail silently rather than erroring.
- Store UTC. Convert to `SITE_TIMEZONE` exactly once, in Cube. Never do timezone maths in a prompt.
- Fix failures in the classifier, Cube or ingestion — **not by adding instructions to the prompt.**
- **A document becomes citable only after a human confirms its number, revision and effective date** —
  `confirm_header()` at a terminal, or the Phase 9 review screen. Never add a path that ingests an
  unconfirmed header. The web-facing role may only make a document **less** citable (withdraw);
  publishing and restoring go through `ai-docs-worker`, which has no HTTP surface.
- When something fails, add the failing case to `eval/testset.jsonl` *before* fixing it.
- Small commits, one concern each. If a change alters an accepted phase, re-run that phase's gate.

## The plant

Waterloo Road Pump Station: a three-pump wastewater station. Wet well `WW-101` (0–7000 mm,
120 m³/m), duty/assist/assist pumps `PU-301/302/303` on a common VSD speed reference, discharging
through manifold `MAN-301` against 22 m static lift. Spill weir at 6000 mm, `LSHH-102` at 5500 mm.
Control runs on `openplc-runtime`; Yokogawa CI Server on `cicore1` polls it over Modbus TCP and
historises the result.

**Four namespaces name the same measurement. Know which one you are holding.**

| Layer | Example | Source of truth |
|---|---|---|
| Instrument tag | `LIT-101` | `WRPS/01-design-doc` |
| PLC symbol + address | `%QW0` | `WRPS/04-plc/register-map.csv` |
| CI Server point | `PS_STN_WET_WELL_LEVEL` | `WRPS/05-scada/modbus/scada-points.csv` |
| **CI Server item** | **`AID.WRPS.STN.LEVEL`** | `WRPS/05-scada/modbus/wrps_item_df.qli` |

**The historian is keyed on the item, and only the item.** Modbus carries register numbers,
not names, so the point layer and the item layer are free to differ — and they do. Keying
history on a point name is what produced all three Phase 5 findings. `db/seed/tags.csv` owns
plant facts; `db/seed/historian_items.csv` owns the item-to-tag mapping and is **generated**
by `scripts/gen_historian_items.py` — never hand-edited.

- **Equipment is asserted in exactly one place: `tags.equipment_id`.** Nothing in the
  history carries an equipment column. CI Server's section tree stops at the station and the
  three pumps; it has no wet well, weir, manifold or switchboard.
- **The historian retains seven days.** Zero rows outside that window means "the historian
  does not go back that far", never "nothing happened". `metrics.HISTORY_RETENTION_DAYS`.
- **Never hardcode a sample interval.** Read `scan_interval_seconds` from the item. The
  analogue groups run at 5 s and 30 s; a measure that assumed 60 s was wrong by twelvefold.