diff --git a/packages/ui/docs/places-llm-filters.md b/packages/ui/docs/places-llm-filters.md
index bdda7846..73597087 100644
--- a/packages/ui/docs/places-llm-filters.md
+++ b/packages/ui/docs/places-llm-filters.md
@@ -1,274 +1,181 @@
-# Places LLM “AI Filters” — investigation & draft plan
+# Places AI grid filters
 
-This document captures the current codebase touchpoints and a proposed architecture for **customizable grid-level LLM filters**: user-defined prompts, **target fields** exposed as **MUI DataGrid** columns/filters, persisted **cache** on `places.meta`, definitions in **user secrets**, and **batch** execution with **SSE** progress—without bloating `GridSearchResults.tsx` further.
+This document describes the **AI grid filters** feature: user-defined prompts stored in **user secrets**, batch LLM evaluation over grid search results, results cached on **`places.meta`**, and **MUI DataGrid** columns driven by those definitions.
 
-### Implementation order (agreed)
-
-1. **Server first** — New HTTP handlers and kbot batch/single logic live in **`server/src/products/places/llm.ts`** (same module family as `handleGetPlaceInfo` / `handleGetLlmRegionInfo`). Register routes next to existing place LLM routes in **`routes.ts`** / `PlacesProduct` as today. Split into e.g. `llm-grid-filters.ts` only if the file grows unwieldy.
-2. **UI second** — New components for the **filter editor**: same interaction pattern as the **Context** editor (toolbar control → **collapsible panel** below the toolbar with textarea / fields, commit-on-blur where appropriate). **Do not** park this next to the view-mode buttons as a second “AI” row long term.
-3. **Placement** — **Collapse AI filters under “Grid Filters”**: the **Grid filters** control (`SlidersHorizontal` + label in `GridSearchResults.tsx`) expands a section that includes both **saved type excludes** (current behavior) and **AI filter definitions** (add/remove/enable, prompts, target field, run on selection / all). Optionally fold or relocate the standalone **Sparkles / Context** strip into this panel so grid-scoped AI lives in one place.
-4. **MUI last** — **`extraColumns`**, **`useAiGridColumns`**, and **`CompetitorsGridView`** wiring once API + editor exist.
+It replaces the older **per-search “potential customer” context** on place summaries; scoring and extraction are **per filter definition** under `meta.llm_grid_filters`.
 
 ---
 
-## 1. Goals
+## Overview
 
-1. **Remove** campaign-specific **`potentialCustomer`** from the place LLM summary flow (prompt + client usage) in favor of **generic, user-defined** “AI filters.” : done
-2. **Extend** the existing **“Grid Filters”** UX (today: toggle + saved **type excludes** from secrets + MUI filters) with **“AI Filters”**: add/remove/enable filter definitions; run them over **current search** and/or **selected rows**.
-3. **Persist filter definitions** in **`user_secrets.settings`** (same pattern as `gridsearch_exclude_types`).
-4. **Persist per-place LLM outputs** in **`public.places.meta`** (JSONB), including **prompt identity**, **result**, **`computed_at`**, and **`source_user_id`** (to allow reuse / shared cache semantics).
-5. **Server**: a **generic batch** runner (many places × many filters) with **SSE** notifications to the client, aligned with existing grid search streaming.
-6. **Frontend**: extract **dynamic columns + row projection** into dedicated modules so `GridSearchResults.tsx` stays orchestration-only.
+1. **Definitions** live in `user_secrets.settings.places_ai_grid_filters` (array), edited in the **Grid filters** panel when the grid search is expanded.
+2. **Runs** call `POST /api/places/llm-filters/run` with `placeIds` and `filterIds`; the server evaluates each pair, merges into `places.meta.llm_grid_filters`, and emits **SSE** events on `GET /api/places/llm-filters/stream?runId=…`.
+3. **Columns** are built client-side (`useAiGridColumns`) from enabled definitions: `GridColDef.field === targetField`, cell value from `row.llm_grid_filters[filterId].value` (by **filter UUID**, not by `targetField` key in meta — see below).
+
+```mermaid
+flowchart LR
+  subgraph secrets["User secrets"]
+    defs["places_ai_grid_filters[]"]
+  end
+  subgraph server["Server"]
+    run["POST /llm-filters/run"]
+    llm["grid-filter.md + LLM"]
+    meta["places.meta.llm_grid_filters"]
+  end
+  subgraph client["Client"]
+    sse["EventSource stream"]
+    grid["DataGrid extraColumns"]
+  end
+  defs --> run
+  run --> llm --> meta
+  run --> sse
+  sse --> grid
+  meta --> grid
+```
 
 ---
 
-## 2. Current state (inventory)
+## Filter definitions (`places_ai_grid_filters`)
 
-### 2.1 Grid “Grid Filters” toggle (`GridSearchResults.tsx`)
+Stored as JSON via existing **`/api/me/secrets`** (same pattern as `gridsearch_exclude_types`). TypeScript: `PlacesAiGridFilterDef` in `src/modules/places/client-gridsearch.ts`.
 
-- State: `applySavedExcludesInGrid` toggles between:
-  - **Off**: `filteredCompetitors` (server/client prefilter) + URL-persisted grid filters.
-  - **On**: full `competitors` list + **seeded MUI column filters** on `types` via `seedExcludeTypeFilters={excludedTypes}` (`PlacesGridView` / `useGridColumns` custom operator `excludesAnyOf`).
-- **Saved excludes** are loaded/stored via **`/api/me/secrets`** as `gridsearch_exclude_types` (see `client-gridsearch.ts` and `PlacesLibrary.postPlacesGridSearchExpand` reading `getUserSecrets`).
+| Field | Description |
+|--------|-------------|
+| `id` | Stable UUID; **key** for `meta.llm_grid_filters[id]` and SSE payloads. |
+| `label` | Optional display name for the column header. If empty or the legacy default “New filter”, the UI falls back to **`targetField`**. |
+| `prompt` | User instructions combined with the server template `grid-filter.md`. |
+| `targetField` | Slug for **`GridColDef.field`** (must match `^[a-z][a-z0-9_]{0,63}$`). Must **not** collide with built-in columns (`types`, `title`, … — see `RESERVED_GRID_FIELDS` in `llm.ts`). |
+| `enabled` | Optional; default **on**. Disabled filters are skipped by the runner. |
+| `valueType` | Optional: **`auto`** (default), **`number`**, or **`string`**. Controls coercion of the LLM `value` for sorting/filtering and MUI column `type`. Included in **prompt hash** so changing it invalidates cached cells until you re-run. |
 
-**Target UX:** Introduce a **`gridFiltersOpen`** (name TBD) collapsible **below** the main toolbar row—same structural pattern as **`aiContextOpen`** + violet panel (see ~`{aiContextOpen && (…textarea…)}`). **Grid filters** becomes the parent affordance: expanded panel holds **type-exclude / MUI hint copy** and the **AI filter editor** (not a duplicate top-level “AI” entry). Today the **Sparkles** control sits beside view modes; plan is to **merge** grid-related AI into this **Grid filters** section.
-
-**Implication:** “AI Filters” **compose** with existing grid behavior: same DataGrid, extra **dynamic columns** whose `field` names match **target fields** stored under `meta`.
-
-### 2.2 MUI grid columns (`useGridColumns.tsx`, `PlacesGridView.tsx`)
-
-- Columns are **mostly static** (`thumbnail`, `title`, `types`, …).
-- **Community DataGrid** constraint: **one filter item** in the filter model in some modes; type excludes are encoded as a **single** synthetic filter (`types` + `excludesAnyOf`).
-- **Extension point:** `GridColDef[]` can be **concatenated** with dynamic defs built from enabled AI filters (string/number columns + `valueGetter` reading `row.meta?.llm_filters?.[key]` or a flattened path).
-
-### 2.3 Place LLM today (`server/src/products/places/llm.ts`, `place-info.md`)
-
-- `GET` handler resolves **`runId`** / manual **`context`**, builds `${grid_search_context}`, runs kbot completion, returns summary JSON for the place detail tab.
-- **Product direction:** campaign fit moves to **per-filter** outputs in **`meta.llm_grid_filters`**, not a dedicated **`potentialCustomer`** field on the summary response (removed from prompt/UI per §7).
-
-### 2.4 `places.meta` (`server/src/products/places/db-places.ts`, `places.ts`)
-
-- **`updatePlaceMeta(placeId, meta)`** merges/replaces JSONB as implemented today.
-- Many code paths **strip** heavy keys (`pages`, `pageErrors`, `bodyHtml`) when sending to clients; any new subtree should be **namespaced** (e.g. `meta.llm_grid_filters`) to avoid accidental huge payloads if logs/debug include raw pages.
-
-### 2.5 SSE today (`PlacesProduct.handleGetPlacesGridSearchStream`)
-
-- Route: **`/api/places/gridsearch/{id}/stream`** (`routes.ts`).
-- Uses Hono **`streamSSE`**, subscribes to **`EventBus`**: `job:progress`, `job:complete`, `job:failed`.
-- Events already include **`location`**, **`node`**, **`job_result`**, with **hydration** from Postgres for fresh `meta` / top-level fields.
-
-**Implication:** batch AI filter runs can either:
-
-- **A)** Emit new **`EventBus`** event types (e.g. `ai_filter_progress`) **on the same `jobId` stream** (client multiplexes), or  
-- **B)** Add a **dedicated SSE route** (e.g. `/api/places/llm-filters/stream?jobId=…`) that only carries filter events.
-
-Option **A** minimizes concurrent connections if the user already holds the grid search stream open; **B** is simpler to reason about if batch jobs are decoupled from the C++ pipeline. **Recommendation:** start with **B** or a **shared helper** that writes SSE frames, then **optionally merge** into the grid stream once event shapes stabilize.
+**UI:** `GridFiltersPanel.tsx` — add/remove filters, save definitions, run on selection or all rows (owner-only).
 
 ---
 
-## 3. Proposed data model
+## Per-place cache (`meta.llm_grid_filters`)
 
-### 3.1 User secrets — filter definitions
-
-Store under e.g. **`settings.places_ai_grid_filters`** (name TBD):
+Merged with `updatePlaceMeta`. Shape per filter id:
 
 ```ts
-type AiGridFilterDef = {
-  id: string;              // stable UUID, used in meta keys
-  label: string;           // UI label
-  prompt: string;          // user text; combined with server base template
-  targetField: string;     // slug: must match GridColDef.field, e.g. "ai_fit_seo"
-  enabled: boolean;        // default on/off for new sessions
-  // optional: valueType: 'string' | 'number' | 'enum'
-  // optional: enumOptions for MUI singleSelect
+meta.llm_grid_filters[filterId] = {
+  value: string | number;       // normalized for the grid (see below)
+  raw: { value, value_type?, detail? };  // parsed LLM JSON (value may match stored value after normalization)
+  promptHash: string;             // invalidates when template version, prompt, targetField, or valueType changes
+  model: string;                  // resolved model id at run time
+  computedAt: string;             // ISO timestamp
+  sourceUserId: string;           // user who ran the job
 };
 ```
 
-- **Validation:** `targetField` — regex `^[a-z][a-z0-9_]{0,63}$`, no collision with built-in fields (`types`, `title`, …).
-- **API:** extend existing **`PATCH /api/me/secrets`** merge behavior (same as `gridsearch_exclude_types`).
-
-### 3.2 Per–grid-search activation
-
-Persist **which filters are active for this run** without a DB migration:
-
-- **`grid_search_runs.settings`** JSON (already used for `is_public`, etc.) **or**
-- **localStorage** key `gridsearch_ai_filters_${jobId}` listing enabled `filter id`s.
-
-**Tradeoff:** DB settings sync across devices; localStorage is faster to ship. Plan: **settings JSON** if multi-device consistency matters; else **localStorage** for v1.
-
-### 3.3 `places.meta` cache subtree
-
-Proposed shape (exact names flexible):
-
-```ts
-meta.llm_grid_filters = {
-  [filterId: string]: {
-    value: string | number | null;  // primary cell value for MUI + quick filter
-    raw?: unknown;                     // optional full model JSON if needed
-    promptHash: string;                // hash of def.prompt + base template version
-    model: string;                     // LLM model id
-    computedAt: string;                // ISO
-    sourceUserId: string;              // user who paid for the LLM call / first writer
-    // optional: shareable: boolean
-  };
-};
-```
-
-**Sharing / reuse:** If another user runs the **same** `(place_id, filterId, promptHash)` and a row already exists, the server may **skip** LLM and **copy** or **reference** the cached value (policy: always allow read if place is visible to user; **write** only if owner or explicit share). **`sourceUserId`** documents provenance for audits.
+**Skip / idempotency:** If `promptHash` for this filter already matches the stored entry, the server **skips** another LLM call for that `(placeId, filterId)` until the definition or template version changes.
 
 ---
 
-## 4. Server API sketch
+## LLM contract
 
-**Home for handlers:** implement in **`server/src/products/places/llm.ts`** (re-export or register from `PlacesProduct` like existing LLM routes). Keeps kbot tasks, Zod `format`, and prompt path helpers in one place.
+### Prompt template
 
-| Endpoint | Purpose |
-|----------|---------|
-| `POST /api/places/llm-filters/run` | Body: `{ placeIds: string[], filterIds: string[], jobId?: string }`. Enqueues work, returns `{ runId }`. |
-| `GET /api/places/llm-filters/stream?runId=` **or** reuse grid SSE | SSE: `progress` (placeId, filterId, status), `cell` (placeId, filterId, value), `complete`, `error`. |
-| `GET /api/places/{id}` | Existing; ensure `meta.llm_grid_filters` is included after merge (respect current stripping rules). |
+- Default path: `server/data/products/places/llm/grid-filter.md`.
+- Override: env **`LLM_GRID_FILTER_PROMPT_PATH`** (absolute or relative to server cwd as implemented by `resolvePromptPath`).
+- Substitutions include place fields from `buildPlaceSubstituteVars`, plus `filter_prompt`, `filter_label`, `target_field`, **`value_type_hint`** (from `valueType`: auto / number / string).
 
-**Internals:**
+### JSON response (`AiGridFilterResultSchema`)
 
-- Drive outputs through **kbot structured data** (see §4.1), not ad-hoc “JSON in prose” parsing only.
-- **Concurrency:** bounded pool (env `LLM_GRID_FILTER_CONCURRENCY`); **idempotent** writes to `meta` (compare `promptHash`).
-- **Auth:** same as place read; batch aborted if any `place_id` not accessible.
+| Field | Type | Description |
+|--------|------|-------------|
+| `value` | `string \| number` | Primary cell value (short; no PII per prompt). |
+| `value_type` | optional `"number" \| "string"` | **Guided** hint: use `"number"` for numeric scales (1–5) so the server coerces and DataGrid can filter numerically. |
+| `detail` | optional `string` | Longer rationale. |
 
-### 4.1 Kbot structured output (`format`) — align with `kbot/src/zod_schema.ts`
+After parse, **`normalizeAiGridFilterValue`** applies **user `valueType`** and/or **`value_type`** and/or numeric-string heuristics so `meta` stores a **number** when appropriate (see `server/src/products/places/llm.ts`).
 
-The kbot CLI and **`run()`** (`@polymech/kbot-d`, used from `server/src/products/places/llm.ts`) support a **`format`** option on **`IKBotTask`**, declared in **`packages/kbot/src/zod_schema.ts`** (`OptionsSchema` → generated `IKBotOptions`).
+### Template version
 
-**What it does**
-
-- **`format`** is transformed (CLI) from:
-  - a path to a **JSON Schema** `.json` file,
-  - a **JSON Schema** object string,
-  - a **Zod** schema string, or
-  - an in-memory **Zod** schema (`zodResponseFormat` branch in `zod_schema.ts`).
-
-- Under the hood, OpenAI’s **`zodResponseFormat(zodSchema, name)`** (`openai/helpers/zod`) produces a **`response_format`** payload for **chat.completions**.
-
-- **`run-completion.ts`** passes it through:
-
-  ```ts
-  await client.chat.completions.create({
-    model: options.model,
-    messages: params.messages,
-    response_format: options.format as any,
-  })
-  ```
-
-So the **model is constrained by the API** to emit JSON matching the schema, not only by prompt text.
-
-**Recommendation for AI grid filters**
-
-1. **Fixed Zod schema** in pm-pics (or shared), e.g. `AiGridFilterResultSchema`:
-
-   - `value`: `string` (primary cell; short, filterable) or `z.union([z.string(), z.number()])` if you need numeric columns.
-   - Optional: `detail`, `confidence`, `labels`, etc., as needed for tooltips or future UI.
-
-2. Build the task roughly as:
-
-   - `import { zodResponseFormat } from 'openai/helpers/zod'`
-   - `format: zodResponseFormat(AiGridFilterResultSchema, 'ai_grid_filter')` (name stable for provider logs).
-   - `mode: 'completion'`, `prompt` = base markdown + substitutions + user **`filter_prompt`**.
-
-3. **Server-side**: still **`schema.parse(JSON.parse(content))`** (or equivalent) after the completion returns, so bad provider output fails fast and never writes garbage to **`meta`**.
-
-4. **promptHash** should include a hash of:
-   - base template text,
-   - **serialized Zod/JSON Schema** (or a fixed **schema version** constant),
-   - user filter prompt,
-   so cache invalidation tracks **schema** changes, not only wording.
-
-**What to avoid**
-
-- Relying on **`runLlmJsonCompletion`**-style **regex strip of ` ```json `** as the **only** contract; keep it as a fallback only if **`format`** is unsupported for a given router/model.
-- Letting users paste **arbitrary JSON Schema** into secrets **without validation** (injection / oversized schema). Prefer **one server-owned Zod schema** for v1; optional **preset enum variants** (e.g. “numeric score 1–5”) as separate fixed schemas later.
-
-**Optional asset**
-
-- Check in **`data/products/places/llm/grid-filter.schema.json`** mirroring the Zod shape if you want CLI parity with kbot’s **file path** `format` loading (`zod_schema.ts` lines 298–308).
+- Constant **`LLM_GRID_FILTER_TEMPLATE_VERSION`** (currently **`2`**) is part of **`promptHash`**. Bump it when the base template or normalization rules change in a way that should force re-computation.
 
 ---
 
-## 5. Prompt strategy (`place-base-filter.md` or not)
+## Environment variables (grid filter LLM)
 
-**Recommendation:**
+Resolution order for **provider**:
 
-- **One server-owned base template** (e.g. `data/products/places/llm/grid-filter.md`) that:
-  - Injects **place facts** (reuse `buildPlaceSubstituteVars` minus campaign context, or a slimmer variant).
-  - Injects **`${filter_prompt}`** from the user definition.
-  - States **privacy / no PII in output** (same spirit as `place-info.md`).
-- **Shape of the answer** is enforced by **`format` + Zod** (§4.1), not by asking for “raw JSON only” in the prompt alone.
-- **Per-filter `.md` files** are **not** required for v1; user prompt in secrets is enough. Add optional **advanced** “prompt file” path later if needed.
+1. `LLM_GRID_FILTER_PROVIDER`
+2. `LLM_PLACE_FILTERS_PROVIDER`
+3. `LLM_PLACE_PROVIDER`
+4. `LLM_REGION_PROVIDER`
+5. default `openrouter`
 
-This avoids proliferating files while keeping **consistent** privacy rules and **machine-verifiable** outputs.
+Resolution order for **model**:
+
+1. `LLM_GRID_FILTER_MODEL`
+2. `LLM_PLACE_FILTERS_MODEL`
+3. `LLM_PLACE_MODEL`
+4. `LLM_REGION_MODEL`
+5. default `openai/gpt-5.2`
+
+Set e.g. `LLM_PLACE_FILTERS_PROVIDER` / `LLM_PLACE_FILTERS_MODEL` in `server/.env` when you want grid filters to use a dedicated model without changing place-info or region LLMs.
 
 ---
 
-## 6. Frontend (after server): Grid Filters panel + components
+## HTTP API
 
-**Pattern:** Reuse the **Context** editor UX: a **toolbar button** toggles open state; when open, a **full-width panel** under the toolbar (bordered card, compact labels) holds forms. Apply the same for **AI filters** inside the **Grid filters** region—**not** a separate floating section next to view modes.
+| Method | Path | Purpose |
+|--------|------|---------|
+| `POST` | `/api/places/llm-filters/run` | Body: `{ placeIds: string[], filterIds: string[], lang?: string }`. Returns **`202`** with `{ data: { runId } }`. Max **80** place ids, **20** filter ids (see route validation). |
+| `GET` | `/api/places/llm-filters/stream?runId=` | **SSE** (`text/event-stream`). For browsers, pass **`?token=`** (auth) when using `EventSource`. Events include progress, per-cell updates, completion, errors. |
 
-Suggested new files under `src/modules/places/gridsearch/`:
-
-| Module | Responsibility |
-|--------|------------------|
-| `GridFiltersPanel.tsx` (or split) | Collapsible content for **Grid filters**: type-exclude explanation, **AI filter list** (add/remove/enable), prompts, target field, run actions. Consumed when `gridFiltersOpen` (or merged state with `applySavedExcludesInGrid`). |
-| `aiGridFiltersState.ts` | Load/save defs from secrets; enabled set per `jobId`; localStorage helpers if used. |
-| `useAiGridColumns.ts` | Build `GridColDef[]` from enabled defs + `valueGetter` from `meta.llm_grid_filters`. |
-| `PlacesGridView.tsx` (or wrapper) | Accept **optional** `extraColumns: GridColDef[]` merged after static columns. |
-
-`GridSearchResults.tsx` stays layout-only: view modes, **`GridFiltersPanel`** slot, **`CompetitorsGridView`** props.
-
-**Phase order:** build **Grid Filters** collapsible + editor **after** endpoints exist (mock with static data if needed); add **MUI dynamic columns** once `meta.llm_grid_filters` flows from the server.
+Handlers: `handlePostLlmFiltersRun`, `handleGetLlmFiltersStream` in `server/src/products/places/llm.ts`; route definitions in `server/src/products/places/routes.ts`.
 
 ---
 
-## 7. Migration from `potentialCustomer`
+## Frontend
 
-1. Remove **`potentialCustomer`** from `place-info.md` and **Zod/UI** types in `LocationDetail` summary tab. : done
-2. If product still needs a **single** “campaign fit” score, implement it as **one AI filter def** with a default prompt (stored in secrets or shipped as preset), not a special-case column in the LLM summary. : eventually yes, we might have default wizard driven filters
+| Area | Behavior |
+|------|----------|
+| `GridSearchResults.tsx` | Loads filter defs (`getPlacesAiGridFilters`), passes **`extraColumns={useAiGridColumns(aiFilterDefs)}`** to the grid for **owners**; wires `GridFiltersPanel`, SSE refresh hooks. |
+| `useAiGridColumns.tsx` | Builds `GridColDef[]`: `field: targetField`, `headerName` from label / targetField, **`type: 'number' \| 'string'`** when `valueType` is set, `valueGetter` reads `row.llm_grid_filters[def.id].value`. |
+| `PlacesGridView.tsx` | Merges **`extraColumns`** after base columns; filter model sync when toggling **“Apply saved type excludes as DataGrid filters”** without remounting the grid (preserves column visibility/order). |
+| `GridFiltersPanel.tsx` | Edit definitions, save to secrets, run batch. |
 
+**Auth:** Only the search **owner** sees the AI filter editor and dynamic columns; viewers use the standard grid without those columns.
 
 ---
 
-## 8. Risks & open questions
+## Numeric scores (1–5, etc.)
 
-- **GET URL length:** batch IDs may require **POST**-only APIs (already directionally true for run).
-- **MUI single-filter constraint:** adding many AI columns may interact badly with **community** grid limits; may need to **lift** filter encoding (documented in `useGridColumns`) when AI columns need filtering simultaneously with types.
-- **Meta size:** cap string length per cell; strip `raw` in list views if large.
-- **Cost / abuse:** rate limits per user; optional confirmation before “run on all.”
-- **Public / shared grid searches:** hide or disable **AI filter** columns and editor affordances for viewers when `is_public` and `!isOwner` (same spirit as hiding sensitive tools); only owner runs LLM fills that write `meta`.
+1. Prefer **`valueType: number`** on the filter (dropdown in the panel) so the column is typed as numeric in MUI.
+2. In the prompt, ask for a **JSON number** and optionally **`"value_type":"number"`** in the model output (see `grid-filter.md`).
+3. **Re-run** AI filters after changing `valueType` or normalization logic so old string cells are replaced.
 
 ---
 
-## 9. Suggested implementation phases (matches § order at top)
+## Related features
 
-1. **Server (`llm.ts` + routes):** Zod **`AiGridFilterResultSchema`** + **`zodResponseFormat`**; single-place single-filter handler (debug); **`POST …/run`** + **`GET …/stream`** (or EventBus bridge); **`updatePlaceMeta`** merge with **`promptHash`** checks.
-2. **Client — Grid Filters UI:** **`GridFiltersPanel`** (or equivalent); filter editor **like Context** (collapsible under **Grid filters**); wire secrets **`PATCH /api/me/secrets`** for definitions; run actions calling batch API + SSE.
-3. **Client — MUI:** **`useAiGridColumns`** + **`CompetitorsGridView`** `extraColumns`; row refresh on SSE / refetch.
-4. **Product cleanup:** `potentialCustomer` removal from place-info pipeline — **done** where noted in §1 / §7.
+- **Saved type excludes:** Same **Grid filters** panel; `gridsearch_exclude_types` + optional **seeded** `types` column filter when “Apply saved type excludes as DataGrid filters” is on (`seedExcludeTypeFilters` / `PlacesGridView`).
+- **Place tab LLM** (`place-info.md`): Separate flow; does not use `places_ai_grid_filters` definitions.
 
 ---
 
-## 10. File reference cheat sheet
+## File reference
 
 | Area | Files |
 |------|--------|
-| Grid toggle & toolbar | `src/modules/places/gridsearch/GridSearchResults.tsx` |
-| Column defs | `src/modules/places/useGridColumns.tsx`, `PlacesGridView.tsx` |
-| Secrets (exclude types) | `src/modules/places/client-gridsearch.ts`, `server/.../places/places.ts` |
-| Place LLM & **new grid-filter LLM** | `server/src/products/places/llm.ts` (handlers + kbot tasks), `server/data/products/places/llm/place-info.md`, `grid-filter.md` (TBD) |
-| Grid Filters UI / Context pattern | `src/modules/places/gridsearch/GridSearchResults.tsx` (toolbar; collapse under Grid filters per §6) |
-| SSE | `server/src/products/places/index.ts` (`handleGetPlacesGridSearchStream`), `server/src/products/places/routes.ts` |
-| Meta persistence | `server/src/products/places/db-places.ts` (`updatePlaceMeta`) |
-| Shared place type | `packages/ui/shared/src/competitors/schemas.ts` (`PlaceSchemaFull`) |
-| **Kbot `format` / options** | `polymech-mono/packages/kbot/src/zod_schema.ts` (`OptionsSchema`, `format` → `zodResponseFormat`) |
-| **Kbot completion → API** | `polymech-mono/packages/kbot/src/commands/run-completion.ts` (`response_format: options.format`) |
-| **Structured output example** | `polymech-mono/packages/kbot/src/examples/core/iterator-markdown-example.ts` (JSON Schema in `format`) |
+| Server logic & SSE | `server/src/products/places/llm.ts` |
+| Routes | `server/src/products/places/routes.ts`, `server/src/products/places/index.ts` |
+| Prompt | `server/data/products/places/llm/grid-filter.md` |
+| Client API & types | `src/modules/places/client-gridsearch.ts` |
+| Grid UI | `src/modules/places/gridsearch/GridSearchResults.tsx`, `GridFiltersPanel.tsx` |
+| Dynamic columns | `src/modules/places/gridsearch/useAiGridColumns.tsx` |
+| Grid shell | `src/modules/places/PlacesGridView.tsx`, `useGridColumns.tsx` |
 
 ---
 
-*Living plan — implementation follows server (`llm.ts`) → Grid Filters panel + editor → MUI columns.*
+## Operational notes
+
+- **Cost / abuse:** Batch size is capped; consider rate limits per deployment.
+- **Meta size:** Keep `value` / `detail` short; avoid large blobs in `raw` if list payloads grow.
+- **Public grid searches:** Non-owners do not get AI filter columns or the editor; LLM writes require an authenticated owner context for runs tied to their secrets.
+
+---
+
+*Implementation reference — behavior is defined by `server/src/products/places/llm.ts` and the routes above.*