diff --git a/docs/observability.md b/docs/observability.md
new file mode 100644
index 0000000..1021c33
--- /dev/null
+++ b/docs/observability.md
@@ -0,0 +1,215 @@
+# Observability & Tracing Operations Guide
+
+## Purpose
+
+This guide describes the end-to-end observability stack in Disciplr backend: how request context flows from inbound HTTP through the Express middleware, the Horizon listener, and background jobs; what Prometheus metrics and structured logs are emitted; and how to correlate traces, logs, and metrics during an incident.
+
+**Target audience:** On-call engineers, service owners, incident commanders.
+**Related runbooks:**
+- [On-Call SLO and Alerting Runbook](runbooks/on-call-slo.md) — SLO targets, alert rules, escalation procedures.
+- [Database Metrics & Operations](operations-metrics.md) — Prometheus scraper auth, gauge definitions, pool health.
+
+---
+
+## Stack Summary
+
+| Layer | Mechanism | Emitter |
+|------|-----------|---------|
+| Request context | Correlation ID (generated or forwarded) | `src/middleware/logger.ts`, `src/middleware/requestLogger.ts` |
+| Structured logs | Pino JSON with redaction | `src/middleware/logger.ts` |
+| HTTP metrics (RED) | `prom-client` Counter + Histogram | `src/observability/httpMetrics.ts` |
+| System/job/listener metrics | `prom-client` Gauges (custom registry) | `src/routes/metrics.ts` |
+| Default process metrics | `prom-client` built-in collectors | `src/routes/metrics.ts` (`collectDefaultMetrics`) |
+
+**Tracing status:** The codebase does **not** currently use OpenTelemetry. Request identity is carried via a `correlationId` (not a W3C `traceparent`). There is no span-based distributed trace export. The correlation ID is the closest equivalent to a trace ID for log correlation; see [Correlation ID ↔ Trace ID](#correlation-id--trace-id).
+
+---
+
+## Context Propagation Path
+
+### 1. Inbound HTTP → Express Middleware
+
+1. `httpMetricsMiddleware` (`src/app.ts:19`, `src/observability/httpMetrics.ts:38`) records request start time and emits `http_requests_total` + `http_request_duration_seconds` on `res.finish`.
+2. `requestLogger` (`src/middleware/requestLogger.ts:49`) extracts or generates a `correlationId` via `getOrGenerateCorrelationId` (`src/middleware/logger.ts:103`), attaches it to `req.correlationId` and creates a Pino child logger at `req.logger`.
+3. The correlation ID is derived from inbound headers in priority order:
+   - `x-correlation-id`
+   - `x-request-id`
+   - Generated fallback: `${Date.now()}-${random}`
+4. `errorHandler` (`src/middleware/errorHandler.ts`) echoes `x-request-id` back in error responses — note it does **not** echo the generated correlation ID, so clients receiving a fallback ID cannot map the response to a log line by ID alone.
+
+### 2. Application Code
+
+Handlers should use `req.logger` (the correlation-scoped Pino child) so log lines carry the `correlationId` field. The root `logger` singleton (`src/middleware/logger.ts:88`) is shared by modules without request context (e.g., `src/security/abuse-monitor.ts`).
+
+**Log-line fields** (common envelope):
+
+```
+{ level, service: 'disciplr-backend', correlationId, event, timestamp, ... }
+```
+
+- `http.request` events (`requestLogger.ts:106`): `req.{method,url,path,headers,body,...}`, `res.{statusCode,headers}`, `durationMs`.
+- `horizon.*` events: `contractAddress`, `lastLedger`, `attempts`, `error`, `stack`.
+- `security.cors_rejected` (`src/app.ts:142`): `origin`.
+- `app_error` / `unhandled_error` (`errorHandler.ts`): `code`, `status`, `method`, `path`, `requestId`, `message`.
+
+### 3. Horizon Listener (separate process)
+
+The Horizon listener runs as a **separate Node process** (`src/services/horizonListenerMain.ts`), launched via `tsx src/services/horizonListenerMain.ts`. Because it is a separate process:
+
+- It inherits **no** Express request context, no `req.correlationId`, no `req.logger`.
+- All logging is via `console.log/warn/error(JSON.stringify({...}))` with no correlation ID and no trace ID.
+- The `HorizonListener` constructor (`src/services/horizonListener.ts:40`) accepts only `config, eventProcessor, db, checkpointStore` — no logger or context bag.
+
+**Implication:** Events processed by the listener cannot be correlated back to an originating HTTP request via the current correlation-ID mechanism. The listener's logs are correlated by `eventId`, `contractAddress`, and `lastLedger` only.
+
+### 4. Background Jobs
+
+The in-memory job queue (`src/jobs/queue.ts`) and `BackgroundJobSystem` (`src/jobs/system.ts`) carry **no trace context**:
+
+- `InternalQueuedJob` (`src/jobs/queue.ts:10`) has fields `id, type, payload, attempt, maxAttempts, createdAt, runAt, leasedAt` — no `traceId`/`correlationId`.
+- `runJob()` (`src/jobs/queue.ts:434`) passes only `{ jobId, attempt }` to handlers.
+- Recurring scheduler jobs (`src/jobs/system.ts:276`) are timer-triggered and have no originating request.
+
+**Implication:** Job executions are not correlated to the request that enqueued them. Job logs must be correlated by `jobId` and `type`.
+
+---
+
+## Correlation ID ↔ Trace ID
+
+There is **no OpenTelemetry trace ID** in this codebase. The `correlationId` field on Pino log lines serves as the de-facto trace identifier for log correlation within the API process.
+
+| Concept | Exists? | Location |
+|--------|---------|----------|
+| W3C `traceparent` header parsing | No | — |
+| OTel span context propagation | No | — |
+| `correlationId` (log-scoped) | Yes | `src/middleware/logger.ts:93-111` |
+| `correlationId` on `req` | Yes | `src/middleware/requestLogger.ts:62` |
+| `x-request-id` echoed in error responses | Yes | `src/middleware/errorHandler.ts` |
+| `correlationId` echoed in responses | **No** | — |
+| Correlation ID propagated to Horizon listener | **No** | Separate process |
+| Correlation ID propagated to jobs | **No** | No context field on `InternalQueuedJob` |
+
+**Triage guidance:** To correlate a request during an incident, use the `correlationId` from the `http.request` log line. Search logs for that exact `correlationId` value. For the Horizon listener and jobs, correlation must be done via time window and `eventId`/`jobId` rather than a shared trace ID.
+
+---
+
+## Exported Metrics
+
+All custom metrics are registered on a dedicated `prom-client` Registry (`src/routes/metrics.ts:10`, exported as `metricsRegistry`). Default process metrics (`process_*`, `nodejs_*`, `gc_*`) are also collected on this registry via `collectDefaultMetrics`.
+
+### HTTP RED Metrics
+
+Defined in `src/observability/httpMetrics.ts`. **Note:** these are registered on the **default global registry**, not the application's `metricsRegistry`, so they are not currently exposed on the `/api/metrics` scrape endpoint.
+
+| Name | Type | Labels | Description |
+|------|------|--------|-------------|
+| `http_requests_total` | Counter | `method`, `route`, `status_class` | Total inbound HTTP requests handled. |
+| `http_request_duration_seconds` | Histogram | `method`, `route`, `status_class` | Latency distribution in seconds. Buckets: 0.005–10s. |
+
+### System Gauges
+
+All defined in `src/routes/metrics.ts`. Aggregate-only — no tenant/org/user labels.
+
+| Name | Type | Description | SLO / Alert Use |
+|------|------|-------------|-----------------|
+| `disciplr_job_queue_depth` | Gauge | Current depth of the background job queue | [SLO 1](runbooks/on-call-slo.md#slo-1-job-queue-depth): warning >100/5m, critical >500/2m |
+| `disciplr_job_failed_total` | Gauge | Total number of failed jobs | Trend monitoring (note: semantically a counter; emitted as gauge) |
+| `disciplr_db_available_connections` | Gauge | Available DB connections in pool | [SLO 2](runbooks/on-call-slo.md#slo-2-db-connection-pool): warning >80%/5m, critical >95%/1m |
+| `disciplr_db_waiting_clients` | Gauge | Clients waiting for a DB connection | [SLO 2](runbooks/on-call-slo.md#slo-2-db-connection-pool): warning >2/5m |
+| `disciplr_horizon_listener_lag` | Gauge | Listener lag in ledgers | [SLO 3](runbooks/on-call-slo.md#slo-3-listener-lag): warning >30/5m, critical >120/2m |
+| `disciplr_outbox_relay_lag_seconds` | Gauge | Age of oldest unprocessed outbox row (s) | [SLO 4](runbooks/on-call-slo.md#slo-4-outbox-relay-lag-informational): warning >300s/10m, critical >3600s/5m |
+| `disciplr_webhook_breaker_closed` | Gauge | Subscribers with closed circuit breaker | Capacity planning |
+| `disciplr_webhook_breaker_open` | Gauge | Subscribers with open circuit breaker | Incident detection |
+| `disciplr_webhook_breaker_half_open` | Gauge | Subscribers with half-open circuit breaker | Recovery monitoring |
+| `disciplr_webhook_dispatch_in_flight` | Gauge | Webhook deliveries currently in flight | Saturation monitoring |
+| `disciplr_webhook_dispatch_queue_depth` | Gauge | Webhook deliveries waiting in queue | Backpressure detection |
+| `disciplr_event_throughput_events_per_sec` | Gauge | Event processing throughput (events/sec) | Throughput trending |
+
+### Default Process Metrics
+
+Enabled via `client.collectDefaultMetrics({ register })` at `src/routes/metrics.ts:13`. Includes `process_cpu_seconds_total`, `process_resident_memory_bytes`, `nodejs_heap_size_total_bytes`, `nodejs_eventloop_lag_seconds`, and GC metrics. See [prom-client docs](https://github.com/siimon/prom-client#default-metrics) for the full list.
+
+---
+
+## Scrape Endpoint
+
+`GET /api/metrics` is mounted at `src/app.ts:177` behind JWT auth, admin RBAC, and a rate limiter (20 req/min). It returns the contents of the custom `metricsRegistry`.
+
+For scraper authentication details (token, IP allowlist, audit logging), see [Database Metrics & Operations](operations-metrics.md#prometheus-scraper-authentication).
+
+---
+
+## Correlating Traces, Logs, and Metrics During Triage
+
+### Scenario: Investigating a slow request
+
+1. **Find the request in logs.** Search Pino JSON logs for the `correlationId` (from `x-correlation-id` header or the `http.request` log line):
+   ```
+   jq 'select(.correlationId == "1719600000-abc123xyz")' < logs.jsonl
+   ```
+2. **Check HTTP RED metrics.** Query Prometheus for the route and status class:
+   ```
+   rate(http_request_duration_seconds_sum{route="/api/vaults/:id"}[5m])
+   /
+   rate(http_request_duration_seconds_count{route="/api/vaults/:id"}[5m])
+   ```
+3. **Check system gauges at the time window.** Did the slow request coincide with pool exhaustion or queue growth?
+   ```
+   disciplr_db_available_connections
+   disciplr_job_queue_depth
+   ```
+4. **Cross-reference the SLO runbook.** See [On-Call SLO and Alerting Runbook](runbooks/on-call-slo.md) for thresholds and remediation.
+
+### Scenario: Alert fired on `disciplr_horizon_listener_lag`
+
+1. Confirm the alert value in Prometheus.
+2. Check listener logs (stdout of the `horizonListenerMain` process) for `horizon.connection_error` or `horizon.stream_error` events around the fire time.
+3. Check `disciplr_db_available_connections` — listener lag often coincides with pool exhaustion.
+4. Follow [SLO 3 remediation](runbooks/on-call-slo.md#slo-3-listener-lag).
+
+### Scenario: Alert fired on `disciplr_job_queue_depth`
+
+1. Check `disciplr_db_waiting_clients` and `disciplr_db_available_connections` for pool exhaustion.
+2. Check `disciplr_horizon_listener_lag` — a stalled listener can starve job-producing events.
+3. Inspect worker process health and logs.
+4. Follow [SLO 1 remediation](runbooks/on-call-slo.md#slo-1-job-queue-depth).
+
+### Correlation limitations
+
+- **Horizon listener logs** have no `correlationId`. Correlate by time window and `eventId`/`lastLedger`.
+- **Job logs** have no `correlationId`. Correlate by `jobId` and `type`.
+- **HTTP RED metrics** are on the default registry, not `/api/metrics`. They must be scraped separately if needed.
+
+---
+
+## Log Sampling & Debug Overrides
+
+`requestLogger` (`src/middleware/requestLogger.ts:37`) applies tail-based sampling:
+
+- Always logged: errors (≥500), slow requests (>`LOG_SLOW_THRESHOLD_MS`, default 1000ms), statuses in `LOG_ALWAYS_LOG_STATUS` (default `500,502,503`).
+- Other requests: sampled at `LOG_SAMPLE_RATE` (default `1` = 100%).
+
+Admin debug overrides (require `ADMIN_API_KEY`):
+- `x-debug-trace: <ADMIN_API_KEY>` forces debug-level logging.
+- `x-log-level: <level>` with `x-admin-key: <ADMIN_API_KEY>` overrides the log level.
+
+---
+
+## Known Gaps & Future Work
+
+1. **No OpenTelemetry.** No packages, SDK, or instrumentation. Correlation ID is log-only and not a W3C trace context.
+2. **Correlation ID not returned in responses.** `errorHandler` echoes `x-request-id` but not the generated `correlationId`.
+3. **HTTP RED metrics on wrong registry.** `http_requests_total` / `http_request_duration_seconds` are on the default registry, not the app's `metricsRegistry`, so they are not exposed on `/api/metrics`.
+4. **Duplicate metric registration.** `disciplr_webhook_dispatch_in_flight` and `disciplr_webhook_dispatch_queue_depth` are declared in both `src/routes/metrics.ts` and `src/services/boundedWebhookDispatcher.ts`.
+5. **`job_failed_total` typed as Gauge** despite monotonic-counter semantics.
+6. **No context propagation** to the Horizon listener (separate process) or background jobs (no context field on `InternalQueuedJob`).
+
+---
+
+## Cross-References
+
+- [On-Call SLO and Alerting Runbook](runbooks/on-call-slo.md) — SLO targets, alert rules, escalation.
+- [Database Metrics & Operations](operations-metrics.md) — Prometheus scraper auth, gauge definitions, pool health, slow-query ring buffer.
+- [Horizon Listener Architecture](horizon-listener.md) — Listener design, recovery scenarios.
+- [Jobs System](jobs.md) — Background job design and configuration.
+- [Operations](operations.md) — General operational procedures.
diff --git a/src/tests/docs.observability.test.ts b/src/tests/docs.observability.test.ts
new file mode 100644
index 0000000..9d4878c
--- /dev/null
+++ b/src/tests/docs.observability.test.ts
@@ -0,0 +1,121 @@
+import * as fs from 'fs';
+import * as path from 'path';
+import { metricsRegistry } from '../routes/metrics.js';
+import { httpRequestsTotal, httpRequestDurationSeconds } from '../observability/httpMetrics.js';
+
+describe('Observability Documentation Sync', () => {
+  const docPath = path.resolve(__dirname, '../../docs/observability.md');
+  const doc = fs.readFileSync(docPath, 'utf8');
+
+  // Metric names documented in the "System Gauges" and "HTTP RED Metrics" tables.
+  // These must match the names registered in code.
+  const documentedMetrics = [
+    'http_requests_total',
+    'http_request_duration_seconds',
+    'disciplr_job_queue_depth',
+    'disciplr_job_failed_total',
+    'disciplr_db_available_connections',
+    'disciplr_db_waiting_clients',
+    'disciplr_horizon_listener_lag',
+    'disciplr_outbox_relay_lag_seconds',
+    'disciplr_webhook_breaker_closed',
+    'disciplr_webhook_breaker_open',
+    'disciplr_webhook_breaker_half_open',
+    'disciplr_webhook_dispatch_in_flight',
+    'disciplr_webhook_dispatch_queue_depth',
+    'disciplr_event_throughput_events_per_sec',
+  ];
+
+  describe('documented metric names exist in the registry', () => {
+    it('should expose every documented custom metric on the application registry', async () => {
+      const registryMetrics = await metricsRegistry.getMetricsAsJSON();
+      const registryNames = registryMetrics.map((m) => m.name);
+
+      for (const name of documentedMetrics) {
+        // http_requests_total / http_request_duration_seconds live on the default
+        // registry (see observability.md "Known Gaps"); all others on the app registry.
+        if (name === 'http_requests_total' || name === 'http_request_duration_seconds') {
+          expect(httpRequestsTotal).toBeDefined();
+          expect(httpRequestDurationSeconds).toBeDefined();
+          continue;
+        }
+        expect(registryNames).toContain(name);
+      }
+    });
+
+    it('should not document any metric that is absent from code', () => {
+      // Extract metric names from markdown table rows: | `name` | ...
+      const tableRegex = /^\|\s*`([a-z0-9_]+)`\s*\|/gim;
+      const found = new Set<string>();
+      let match: RegExpExecArray | null;
+      while ((match = tableRegex.exec(doc)) !== null) {
+        found.add(match[1]);
+      }
+
+      // Every metric name appearing in a docs table must be one we know about.
+      const known = new Set(documentedMetrics);
+      for (const name of found) {
+        expect(known.has(name)).toBe(true);
+      }
+    });
+  });
+
+  describe('required sections present', () => {
+    const requiredSections = [
+      '## Purpose',
+      '## Stack Summary',
+      '## Context Propagation Path',
+      '## Correlation ID',
+      '## Exported Metrics',
+      '## Scrape Endpoint',
+      '## Correlating Traces, Logs, and Metrics',
+      '## Log Sampling',
+      '## Known Gaps',
+      '## Cross-References',
+    ];
+
+    for (const section of requiredSections) {
+      it(`should contain the "${section}" section`, () => {
+        expect(doc).toContain(section);
+      });
+    }
+  });
+
+  describe('runbook cross-link valid', () => {
+    it('should link to the on-call SLO runbook with a valid anchor', () => {
+      expect(doc).toMatch(/\[.*?\]\(runbooks\/on-call-slo\.md/);
+    });
+
+    it('should link to the operations-metrics doc', () => {
+      expect(doc).toMatch(/\[.*?\]\(operations-metrics\.md/);
+    });
+
+    it('should reference all four SLO runbook anchors', () => {
+      const anchors = [
+        'runbooks/on-call-slo.md#slo-1-job-queue-depth',
+        'runbooks/on-call-slo.md#slo-2-db-connection-pool',
+        'runbooks/on-call-slo.md#slo-3-listener-lag',
+        'runbooks/on-call-slo.md#slo-4-outbox-relay-lag-informational',
+      ];
+      for (const anchor of anchors) {
+        expect(doc).toContain(anchor);
+      }
+    });
+  });
+
+  describe('documented metric metadata matches code', () => {
+    it('should document job_queue_depth as a gauge with the correct help', async () => {
+      const registryMetrics = await metricsRegistry.getMetricsAsJSON();
+      const m = registryMetrics.find((x) => x.name === 'disciplr_job_queue_depth');
+      expect(m).toBeDefined();
+      expect(m?.type).toBe('gauge');
+    });
+
+    it('should document horizon listener lag as a gauge', async () => {
+      const registryMetrics = await metricsRegistry.getMetricsAsJSON();
+      const m = registryMetrics.find((x) => x.name === 'disciplr_horizon_listener_lag');
+      expect(m).toBeDefined();
+      expect(m?.type).toBe('gauge');
+    });
+  });
+});