Publisher playtest telemetry — July 16–19, 2026 (baseline)¶
The first real-cohort test of the released build. This is the baseline the action pivot
must beat — cited by docs/design/action-pivot/ and the Phase-3 metrics gate in its
implementation plan.
Build under test: the pre-joystick tap-to-move stealth build (last client commit Jul 15; the virtual joystick landed Jul 21, after the test window). The cohorts played the stealth game with its full verb set — tap-to-move, sprint, swipe-dodge, tap-hunt. Data interpretation must account for this: the stealth loop failed with its tools, not because the joystick stripped them.
Cohorts¶
250 installs across four daily cohorts (Jul 16–19, ~58–75 each). Not all Jul 18–19 data had arrived at analysis time. Small-n caveat: D2+ cells are single-digit users — treat everything past D1 as directional only. Cohorts 3–4 (Sat/Sun) showed ~2× the D0 playtime of cohorts 1–2 (Thu/Fri) with no client change between them (verified in git) — likely a weekend / creative-mix effect, not a product signal.
Headline metrics (weighted means)¶
| Metric | Value | Note |
|---|---|---|
| CPI | $1.52 | Within benchmark — the visual set sells. Best creative: standard camera, B&W environment, colored characters |
| D1 retention | 22.5% | Cohort range 15.3–24.1%; significantly below target |
| D2 / D3 retention | 16.4% / 3.45% | D2→D3 collapse; single-digit users |
| D0 playtime | 20.6 min | Over 2.2 sessions → ≈9.4 min/session |
| Session ≈ raid | ~1 raid/session | 9.4 min ≈ roundTime 600 s — the raid length dictates the session |
| D1 playtime (returners) | 25.5 min / ~3.1 sessions | Returners are engaged — the kernel works for the ~20% it lands with |
Funnel¶
Tutorial (7 stages, steps 0-1 … 0-7): 93–98% completion per step, cumulative 75.5%. No single broken step — the loss is the accumulated price of seven sequential gates. The only tutorial step with meaningful fails is 0-5 Open Combat (fail/complete 3.11).
Real floors: 17–35% churn per floor. 65.2% complete 1-1 · 25.7% reach biome 2 · 15.9% complete 2-1 · 6.0% complete 3-1.
Fail-rate curve maps to content introductions¶
| Floor | Fail/complete | What enters there (mobIntro / config) |
|---|---|---|
| 1-1 | 1.32 | — baseline melee (spider, goblin) |
| 1-2 | 13.79 | Thorn Hound — pack-flagged, the first real melee pressure (correction 2026-07-23: an earlier note blamed its authored speed 162, but non-boss mobs move at player parity via enemyScale.enemySpeedMul — authored speeds are inert) |
| 1-3 | 9.38 | Emberling — first ranged enemy |
| 1-4 | 12.99 | Bone Thug |
| 1-5 | 43.86 | Boss floor — Brute hp 1500 × mobHpMult 2 |
| 2-1 | 18.92 | Rivals appear (prog.botsFromBiome 1) + Warden — first mob that out-sees the player's fog (visR 340 > 300); churn 35% |
The difficulty problem is spiky, not uniform — a blanket nerf would flatten the game along with the spikes. The spikes are individually addressable (hound entry floor, boss encounter design, the 2-1 triple-stack of new pressures).
Publisher read¶
Kill the game or make major changes to gameplay and core mechanics; CPI is good enough to
test hypotheses quickly. Directive: a new extraction game reusing the current assets and
part of the mechanics — not an iteration of the current game. The action pivot
(docs/design/action-pivot/) is the response.
Pivot targets derived from this baseline (Phase-3 / G3 gate)¶
| Metric | Baseline | Target |
|---|---|---|
| FTUE completion | 75.5% | ~88%+ (3 stages at the same ~95%/step) |
| D1 retention | 22.5% | 30%+ |
| D0 sessions / raids per session | 2.2 / ~1 | 3+ / 2–3 |
| Fail/complete on 1-1 and 1-2 | 1.3 / 13.8 | < 5 (the "designed win") |
| Reach biome 2 | 25.7% | 45%+ |