@apex-inc/mcp-server 0.10.0 → 0.12.0

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@apex-inc/mcp-server",
-  "version": "0.10.0",
+  "version": "0.12.0",
   "description": "MCP server for Apex — currently in private beta.",
   "type": "module",
   "main": "dist/index.js",

package/skills/apex-experimentation/SKILL.md

@@ -114,7 +114,31 @@ Match Apex UI terminology in user-facing copy: e.g. **Adaptive traffic allocatio
 - Changing variant implementation mid-flight without versioning (invalidates analysis).
 - Only using p-values without effect size or business context.
 
+## Prove incrementality — "is our growth real?"
+
+A winning experiment proves variant B beat variant A. It does NOT prove Apex beat *doing nothing* — maybe that growth would have happened anyway. The **global holdout** answers that: a persistent slice of users Apex leaves completely alone (journeys suppressed, experiments forced to control). The lift of everyone else vs that do-nothing cohort, per goal, is the number a CFO can't argue with.
+
+Recipe:
+
+1. **Confirm the holdout is on.** Settings → Journeys → "Global holdout (measurement)" (default 5%). It governs journeys AND experiments — one cohort proves your whole Apex footprint.
+2. **Read the proof.** Call `get_growth_reality({ days })` (or `apex.getGrowthReality()` / `GET /api/growth/reality`). Each goal returns lift vs holdout + a 95% CI + `isDecisive`.
+3. **Respect the honest states.** `isDecisive: false` means "still collecting" — report it as directional, never as proof. Counting is forward-only from when the holdout went live; there is no backfill.
+4. **Reconcile credit.** Realized revenue credit is measured, significance-gated, net of refunds, and excludes guardrail-blocked experiments. Don't quote a raw point-estimate projection as "attributed revenue" — quote the holdout-gated number.
+
+When a user asks "did Apex actually work?" or "is this growth real?", reach for `get_growth_reality`, not a per-experiment p-value.
+
+## Editing an experiment (fix a draft; fork after data)
+
+Editability is gated by lifecycle, keyed to DATA:
+
+- **Draft, no exposures yet:** fully editable. Use `update_experiment({ experimentId, ... })` to fix a mislabeled surface, rename, or correct the hypothesis/metric/guardrail. This is where typo and surface fixes belong.
+- **Running, or any exposures recorded:** the pre-registration (hypothesis, metric, guardrails) and bucketing (surface, variants) FREEZE. `update_experiment` returns `experiment_locked` with the reason — that's intentional, so results stay valid. Do not fight it.
+- **After freeze:** call `fork_experiment({ experimentId })`. It duplicates the design into a fresh draft (no data carries over, lineage recorded), which you then edit and activate.
+
+Surface specifically: `create_experiment` auto-detects mobile vs web from the workspace's data sources; if a Capacitor app got mislabeled `web`, fix it on the draft with `update_experiment({ experimentId, surface: "mobile" })` before the first exposure, or fork if it's already running.
+
 ## Related
 
 - Growth tracking and identity: see **apex-growth-tracking** skill (`@apex-inc/mcp-server` / `@apex-inc/sdk`).
 - Full methodology (beliefs, calibration): use MCP resources `apex://beliefs`, `apex://calibration`, and tools `create_belief`, `record_outcome`.
+- Global holdout + the four-way measurement: `/docs/concepts/global-holdout`.