diff --git a/docs/advisor_framework_strategy.md b/docs/advisor_framework_strategy.md
new file mode 100644
index 0000000..58b66ed
--- /dev/null
+++ b/docs/advisor_framework_strategy.md
@@ -0,0 +1,197 @@
+# Advisor framework strategy (multi-source; push-first as a working hypothesis)
+
+This doc is **product/architecture oriented** and intended as a jumping-off point for engineering investigation. It’s written to be OSS-friendly (conceptual, not a binding implementation spec).
+
+It does *not* propose adding non-SQL “advisor engines” into Splinter. Instead, Splinter remains a **SQL-only** producer of Postgres findings, while other sources (logs/traces/metrics/edge/humans) can create Advisor issues via an ingestion API.
+
+References:
+- Splinter docs: `https://supabase.github.io/splinter/`
+- Supabase Database Advisors docs: `https://supabase.com/docs/guides/database/database-advisors?queryGroups=lint&lint=0012_auth_allow_anonymous_sign_ins`
+
+## Key design choice: push vs pull
+
+### Push (often a good fit)
+Producers evaluate signals “close to the data” and **push** Advisor issues:
+- Logs/traces: Logflare/ClickHouse queries + thresholding in a job or alerting layer
+- Metrics: **VictoriaMetrics + vmalert** (or Alertmanager) → webhook to the ingestion API
+- Product services: plan/usage thresholds, resource exhaustion, quota events
+- Edge Functions: detect misconfigurations and push issues
+- Humans / Assistant: manual issues (“broadcasting important updates”)
+
+Pros:
+- Uses each system’s native evaluation semantics (PromQL + alert windows, etc.)
+- Avoids Supabase embedding every query engine + auth model
+- Reduces load on Postgres/telemetry backends from “always-on polling”
+
+### Pull (fallback / managed checks)
+Supabase periodically queries sources (ClickHouse, PromQL, etc.) and synthesizes issues.
+
+Pros:
+- Great for onboarding (“turnkey checks”)
+- Works even if users haven’t wired alerting/webhooks
+
+Cons:
+- Forces Supabase to own connectors, auth, safe query execution, and alert semantics
+
+**Working hypothesis:** provide a **push-first ingestion API** and optionally layer **managed pull-based checks** behind the same ingestion + issue model.
+
+## Advisor API surface (conceptual capabilities)
+
+Minimum viable operations (regardless of URL shape):
+- **Upsert** an issue (create or update, typically setting state to `open`)
+- **Resolve** an issue (set state to `resolved`)
+- **Dismiss** an issue (set state to `dismissed`)
+- **Snooze** an issue (set state to `snoozed` until a timestamp)
+
+You can also unify create/resolve as a single “upsert current state” operation if the payload includes `state`.
+
+## Webhooks as first-class plumbing
+
+If Supabase is investing in “more webhooks” functionality, it fits the Advisor strategy in **two** complementary ways:
+
+### 1) Inbound webhooks (create/update Advisor issues)
+Treat a webhook delivery as a **producer** of Advisor issues. Examples:
+- VictoriaMetrics/vmalert/Alertmanager firing events → create/update/resolve Advisor issues
+- log-based alerting (Logflare/ClickHouse) → create/update/resolve Advisor issues
+- internal platform events (resource exhaustion, quota events, deploy events) → create/update/resolve Advisor issues
+- schema change events (e.g. “policy changed”, “RLS disabled”) → create Advisor issues or prompt re-run of Splinter
+
+This aligns with your goal: “Advisor as a flexible issue creation and triggers framework”.
+
+### 2) Outbound webhooks (broadcast Advisor issues)
+When an Advisor issue is created/updated/resolved/dismissed, Supabase can emit **outbound** events to user-configured destinations:
+- Slack, email, PagerDuty, generic webhooks, etc.
+
+This is separate from issue *creation* but part of the end-to-end “alerting/notifications + advisor” experience:
+- Advisor issue = the durable, discussable, dismissable record
+- Outbound webhooks = delivery/routing based on user preferences
+
+Practical implementation detail:
+- model outbound delivery rules separately from issue types
+- support per-project routing and per-type routing (e.g. only `ERROR` to PagerDuty)
+
+## Issue identity: generalize Splinter’s `cache_key`
+
+Splinter already encodes the correct idea:
+- each finding has a stable `cache_key` that can be used for suppression/dedupe
+
+Generalize this for *all* sources:
+- **type** (aka check id): e.g. `splinter/unindexed_foreign_keys`, `logs/http_5xx_spike`, `metrics/cpu_hot`
+- **fingerprint**: stable instance key within a type
+
+Store uniqueness as:
+- `(project_ref, type, fingerprint)`
+
+### Lifecycle state
+- **open**: currently firing / present
+- **resolved**: no longer firing / no longer present
+- **dismissed**: user doesn’t want to see it (until explicitly reopened)
+- **snoozed**: suppressed until a timestamp
+
+## Suggested ingestion payload (conceptual)
+
+An ingestion payload should be able to carry:
+
+- `type` (string)
+- `fingerprint` (string)
+- `severity` (`ERROR|WARN|INFO`)
+- `title` (string)
+- `detail` (string)
+- `categories` (string[])
+- `remediation` (url/string, optional)
+- `metadata` (object, optional)
+- `source` (object, optional): `{ kind: "splinter|clickhouse|vmalert|edge|manual", ref?: string }`
+- `state` (optional): if absent, default to `open`
+- `observed_at` (timestamp, optional)
+
+On the backend, this becomes a durable issue record with timestamps (`first_seen`, `last_seen`, etc.).
+
+## How sources map to the ingestion model
+
+### Splinter (Postgres schema advisors)
+Splinter stays SQL-only:
+- lints in `lints/` emit rows with `cache_key`
+- `splinter.sql` unions all lints
+
+Producer job options:
+- UI job (on-demand “rerun advisors”)
+- backend cron (e.g. daily) to keep issues warm
+
+Flow:
+1) run `splinter.sql`
+2) convert each row to `type = "splinter/{name}"` and `fingerprint = cache_key`
+3) upsert issues
+4) resolve issues for that `type` that were not returned (or use an explicit `resolve` call)
+
+### Logs/traces (Logflare → ClickHouse)
+Push-first options:
+- a scheduled job queries ClickHouse for the last N minutes and pushes issues
+- a logs alerting layer (if present) calls the ingestion endpoint directly
+
+Example types:
+- `logs/http_5xx_spike` (fingerprint: `service=X,route=Y`)
+- `traces/p95_latency_regression` (fingerprint: `service=X,operation=Y`)
+
+### Metrics (VictoriaMetrics)
+Preferred path: **vmalert → webhook**.
+
+Why:
+- vmalert already encodes windowing, alert state, de-dupe, and silence/inhibition concepts
+- Supabase only needs to map alert events into Advisor issues
+
+Fallback path: managed pull checks.
+- Supabase runs curated PromQL on a schedule
+- returns series above a threshold → upsert issues
+- no longer returned → resolve issues
+
+### VictoriaMetrics SaaS alerting (managed alert rules)
+If you’re using VictoriaMetrics’ SaaS/managed offerings for alerting (rule management UI, managed alert evaluation, etc.), treat it the same as vmalert conceptually:
+- alert evaluation happens in VictoriaMetrics’ alerting system
+- alerts deliver to a webhook receiver
+- the Advisor ingestion API is one such receiver
+
+The key mapping is still:
+- `type`: stable alert name (namespaced, e.g. `metrics/cpu_hot`)
+- `fingerprint`: stable label-set identity (e.g. `instance=db-01` or `cluster=X:instance=Y`)
+- `state`: open when firing, resolved when cleared
+
+What we likely need to confirm (since SaaS feature sets vary):
+- what webhook payload shape is emitted (Alertmanager-compatible vs custom)
+- whether “resolved” notifications are delivered
+- how de-dupe/inhibition/silences are represented
+
+## Standardized Advisor types (registry) vs custom
+
+You likely want both:
+- **Curated registry** of known types for UI, docs, and consistent severity/category semantics
+- **Custom types** for user-defined rules and internal experimentation
+
+Practical compromise:
+- accept any `type` string
+- Studio treats known prefixes (e.g. `splinter/`, `logs/`, `traces/`, `metrics/`, `billing/`) as first-class
+- optionally expose `GET /advisor/types` for metadata (title, docs link, default severity)
+
+## What belongs in Splinter (and what doesn’t)
+
+Splinter should remain focused on what it’s good at:
+- deterministic Postgres catalog/state checks
+- SQL queries that return structured findings
+- stable fingerprints (`cache_key`)
+
+It should *not* embed:
+- alerting engines
+- time-windowed evaluation for logs/traces/metrics
+- notification routing logic
+
+## Open questions (for an engineering spike)
+
+These are intentionally left open to avoid over-prescribing implementation:
+
+- **Inbound webhook shape**: will inbound alerts be Alertmanager-compatible, vendor-specific, or both?
+- **Resolved semantics**: do upstream systems send “resolved” events (or do we infer resolution)?
+- **Auth**: how do producers authenticate to the ingestion API (per-project secrets, signed webhooks, service tokens)?
+- **Multi-tenancy**: how does a producer target the correct `project_ref` (explicit routing vs token-bound project)?
+- **Dedupe**: what’s the recommended fingerprint scheme per source (labelset hashing, stable IDs, etc.)?
+- **Type registry**: do we want a curated set of “known types” (for UI/UX), plus custom types (for extensibility)?
+- **Outbound routing**: how should outbound webhooks/notifications relate to issues (per type, per severity, per project)?
+
diff --git a/docs/architecture.md b/docs/architecture.md
new file mode 100644
index 0000000..9887aa1
--- /dev/null
+++ b/docs/architecture.md
@@ -0,0 +1,114 @@
+# Architecture: Splinter and Supabase Database Advisors
+
+This document explains how this repository works today (Splinter), how it maps to Supabase’s “Database Advisors” UI, and why the current approach is inherently **SQL/Postgres-only**.
+
+For end-user facing documentation of individual lints, see the lint pages on the Splinter docs site (`https://supabase.github.io/splinter/`) and Supabase docs ([Database Advisors](https://supabase.com/docs/guides/database/database-advisors?queryGroups=lint&lint=0012_auth_allow_anonymous_sign_ins)).
+
+## What Splinter is (and is not)
+
+Splinter is a **collection of SQL lints** intended to be run against a Supabase Postgres database. Each lint is implemented as a view that returns a standardized “issue row” shape.
+
+Splinter is **not** a generic alerting/advisory system. It does not:
+- schedule periodic evaluations
+- store issue lifecycle state (open/resolved/dismissed)
+- evaluate non-Postgres sources (logs, traces, metrics, client errors, billing/usage)
+
+## The lint interface (the contract)
+
+Each lint view returns the same columns (see `README.md`):
+- `name` (text) — stable check id (e.g. `unindexed_foreign_keys`)
+- `title` (text) — human title
+- `level` (text) — `ERROR` / `WARN` / `INFO`
+- `facing` (text) — `INTERNAL` / `EXTERNAL`
+- `categories` (text[]) — e.g. `{SECURITY}` or `{PERFORMANCE}`
+- `description` (text) — why this matters
+- `detail` (text) — specific instance detail (table/view/function etc.)
+- `remediation` (text, optional) — documentation URL or guidance
+- `metadata` (jsonb, optional) — structured data for UI rendering or deep links
+- `cache_key` (text) — **stable fingerprint** used for de-dupe/suppression
+
+**Key point:** `cache_key` is the only stable identifier for “the same finding” across runs. It’s the hook needed for an issue lifecycle system.
+
+## How lints are authored
+
+Lints live in `lints/` as one `.sql` file per lint, each creating a view under the `lint` schema:
+
+- Example: `lints/0001_unindexed_foreign_keys.sql` creates `lint."0001_unindexed_foreign_keys"`.
+- Tests read from these views directly (see `test/sql/`).
+
+## How Splinter is “compiled” into one query
+
+Some consumers (like a UI) want to run a single query and receive a single result set containing all findings.
+
+Splinter ships that as `splinter.sql`, produced by `bin/compile.py`:
+- reads `lints/*.sql`
+- strips the `create view` line and semicolons
+- wraps each lint query in parentheses
+- concatenates them with `UNION ALL`
+- prefixes `set local search_path = '';` to enforce schema qualification
+
+Net: **`splinter.sql` is “all lints as one unionable query”.**
+
+## How Splinter is tested
+
+Tests are Postgres regression tests that validate each lint has at least a true-positive case:
+- `test/fixtures.sql` provisions minimal schemas/roles/functions needed for lint execution.
+- `test/sql/*.sql` sets up example schema states and selects from `lint."00xx_*"`.
+- `test/expected/*.out` captures expected output.
+- `bin/installcheck` runs pg_regress and also validates that `splinter.sql` runs.
+
+## How this maps to Supabase “Database Advisors”
+
+Supabase’s “Performance and Security Advisors” are a set of checks surfaced in the dashboard UI ([Database Advisors docs](https://supabase.com/docs/guides/database/database-advisors?queryGroups=lint&lint=0012_auth_allow_anonymous_sign_ins)).
+
+The “DB schema / security / performance” subset of those checks can be implemented as Splinter lints because they are answerable by querying Postgres system catalogs and Supabase schemas.
+
+Operationally, a consumer (such as Supabase Studio) can:
+- run `splinter.sql` (or an equivalent union query)
+- render each row as an “advisor finding”
+- allow users to dismiss/suppress findings via `cache_key`
+
+## When are lints run (triggering) and where do results live?
+
+Splinter itself does **not** define scheduling, triggers, or persistence. It only defines the SQL that produces findings.
+
+In Supabase’s product, there are typically two different “execution contexts” to be aware of:
+
+### 1) Interactive / on-demand (e.g. Supabase Studio)
+In Supabase Studio (in the `supabase/supabase` repo), the current “lints/advisors” flow is implemented as an **on-demand SQL query** executed against the project’s Postgres database when the UI needs it (for example, when you open the Advisors page or click a re-run action).
+
+Concretely, Studio sends the lint SQL to a platform endpoint that proxies to `pg-meta`, executing the query against the project database (see `executeSql` in Studio which posts to `/platform/pg-meta/{ref}/query` in `supabase/supabase`).
+
+This is why Splinter focuses on:
+- deterministic SQL
+- predictable output shape
+- stable `cache_key` values (so consumers can de-dupe and suppress repeated findings)
+
+### 2) Scheduled / aggregated (analytics, fleet-level views)
+Separately from the UI, Supabase may run similar checks on a schedule and/or export results for analytics. For example, you mentioned querying a BigQuery table named `supabase-etl-prod-eu.dbt.project_lints`.
+
+That table name is a good example of historical terminology: the underlying concept is “advisor findings/issues”, but earlier naming often used “lints”. It’s worth calling out explicitly as Supabase’s long-term direction expands Advisors beyond SQL lints.
+
+Because this repo is OSS and Splinter-only, the precise scheduling/ETL details live outside this repo (Studio/backend/data pipelines). A reasonable mental model is:
+- Splinter provides the *finding generator* (SQL)
+- Supabase systems decide *when to run it* and *whether/how to store it* (UI state, issue lifecycle, analytics exports)
+
+## A concrete limitation example (docs vs SQL)
+
+This repo contains documentation for check `0012_auth_allow_anonymous_sign_ins` (`docs/0012_auth_allow_anonymous_sign_ins.md`), but there is no corresponding SQL lint in `lints/`.
+
+That’s a useful reminder: some “advisor” checks depend on **non-Postgres state** (e.g., auth configuration) or **time-windowed telemetry** (logs/traces/metrics), which can’t be expressed as a static Postgres catalog query.
+
+## Why this matters for “multi-source Advisors”
+
+If you want Advisors that can come from logs/traces/metrics/UI errors/etc., you need something beyond “run a SQL union query against Postgres”.
+
+Splinter’s lint rows are a good *finding format* for DB checks, but a full Advisor system also needs:
+- source-specific evaluation (ClickHouse, PromQL, etc.)
+- scheduling and/or event triggers
+- a stable issue model + dedupe
+- lifecycle storage (open/resolved/dismissed/snoozed)
+- routing/notification integrations
+
+
+