Tactical Action · Guides & deep dives

0011: Per-Cell Kill / Defer Criteria and De-Scope Sequence

ADR 0002 commits V4 to six launch cells and states that removing a launch cell "requires a phase-cut ADR" — but no document defines when such an ADR is warranted, what evidence it needs, or which cell goes first.

3sections7 minread1table
Markdown source
On this page

Status: Proposed

Date: 2026-06-12

Context#

ADR 0002 commits V4 to six launch cells and states that removing a launch cell "requires a phase-cut ADR" — but no document defines when such an ADR is warranted, what evidence it needs, or which cell goes first. The 2026-06-12 plan-set audit (§6.2) flagged this as a verified gap: the program's single biggest bet (six AAA-feel cells at once, risk R-01 in V4/docs/planning/program-risk-register.md) had no contingency mechanics.

The raw materials for criteria already exist:

  • Feel-test gates are quantitative and release-blocking per cell (V4/V4_features.md:343-384): e.g. Hitman sandbox at 60 fps with 120 NPCs and a sub-12-minute fresh-player loop; Wukong 6-frame ±1 parry windows; RTS openings replicable within ±2% of SC2 timing; Contra 60 fps couch co-op. They run in nightly Gauntlet from V4/ue/Source/V4Tests/FeelTests/, and "a failing feel test blocks the affected ruleset's release" (V4/V4_features.md:382-384).
  • Content inventory rows per cell are committed (V4/V4_features.md:1576-1603).
  • Launch milestones derive from V4/V4_features.md:2722-2726 and the fixed launch date (V4/liveops/live-service-calendar.json, 2026-10-01): Closed Beta 2026-04-01, Open Beta 2026-08-03, cert submission window opening 2026-08-17, Soak 2026-09-24 (dates are planning assumptions adopted 2026-06-12 where not directly specced).
  • "No coming soon" is absolute (V4/V4_features.md:290-291, exit criterion :2736): a deferred cell must vanish from the advertised product, not linger as a teaser panel.

This ADR turns those materials into decision criteria. It does not cut any cell; each actual cut still gets its own phase-cut ADR per ADR 0002, citing this ADR's criteria and attaching the evidence pack.

Decision#

1. Gates and feel-test failure thresholds#

Each cell is measured at four gates. "Green" means the cell's feel tests (V4/V4_features.md:343-384) pass in nightly Gauntlet; percentages are over the stated trailing window of nightlies. Thresholds are planning assumptions adopted 2026-06-12.

Gate Date Feel-test threshold (per cell) Additional bar
G1 — Closed Beta entry 2026-04-01 All feel tests implemented in Gauntlet; ≥80% green, trailing 21 nightlies, PC Perf scenarios within 20% of target on console-proxy hardware
G2 — Open Beta entry 2026-08-03 ≥95% green, trailing 28 nightlies, on PC + PS5 + Series X Quantitative feel numbers met on primary platforms (60 fps scenarios ≥58 fps p95); Closed-Beta survey "feels like the inspiration" top-2-box ≥70% for the cell
G3 — Cert submission 2026-08-17 100% green, trailing 21 nightlies, on every platform that cooks the cell Cell's content-inventory rows 100% content-complete; platforms that strip the cell per ADR 0005 cook rules are recorded as per-platform deferrals in the cert plan
G4 — Soak / RC 2026-09-24 Zero feel-test regressions on the release-candidate build Per-cell crash rate <0.1% during Soak (aligned with V4/V4_features.md:2743)

Defer trigger (feel): a cell that (a) misses the same gate twice — initial miss plus a re-test after a 14-day cure period — or (b) is red on >25% of nightlies over any trailing 30-day window inside the 60 days before a gate, enters mandatory defer review (§4).

2. Production-cost overrun triggers#

Each cell carries a cost baseline (estimate-at-completion) re-set at G1 and re-estimated monthly. Defer review is mandatory when either:

  • the cell's estimate-to-complete exceeds 130% of its G1 baseline in two consecutive monthly re-estimates, or
  • the cell consumes >1.4× its planned share of pooled production capacity (mocap stage hours, VO studio hours per V4/docs/planning/program-risk-register.md R-03, cinematics minutes) in a quarter such that another cell's gate-critical content slips.

Content-completion floor: independent of cost, a cell below 70% of its content-inventory rows (V4/V4_features.md:1576-1603) at content-complete quality at G2 enters defer review — content debt that late cannot be bought back by money alone.

3. De-scope sequence (the order, argued from shared-tech leverage)#

If a defer review concludes "defer," cells defer in this order. The ordering principle: defer the cell whose removal frees the most schedule while destroying the least shared technology and the least of the product promise.

  1. 2D Run-and-Gun (Contra) — first. Shares the least infrastructure of any cell: its own side-scroll camera/renderer path, no use of V4Perception/V4Crowd/V4Tactical (the shared surfaces named at V4/V4_features.md:390-394), couch-co-op netcode unique to it, the smallest content row (8 stages, :1595), no ranked ladder (:2020-2024), no esports pool. Its deferral has the smallest blast radius and breaks no other cell's dependency. (Corollary: it also frees the least capacity — deferring it is a signal, not a rescue.)
  2. Real-Time Strategy — second. The RTS cell is the most isolated netcode risk in the program: 25 Hz deterministic lockstep over 32.32 fixed-point simulation (V4/V4_ARCHITECTURE.md:890-932) shares nothing with the client-server or rollback stacks and uniquely requires cross-platform determinism on all 9 targets (risk R-04). It also carries the largest map inventory (52 maps across SC-style/AoE-style/TD, V4/V4_features.md:1592-1594) and two of the five ranked ladders. Deferring RTS removes the single largest concentrated technical risk and the second-largest content burden in one decision. Costs to absorb: three of the five factions are RTS factions (:1979-1985) — the faction system must launch with the two Shadow War blocs and add the RTS factions with the cell; the esports circuit drops two ladders (apply risk R-15's collapse-into-Wildcard mitigation).
  3. Action RPG (Wukong) — third. Shares the character/animation/GAS combat pipeline with the Tactical cell, so its engine work is mostly not wasted by deferral, and it is co-op-only (V4/V4_features.md:339) so it adds no PvP service load — but for the same reasons its deferral saves less server and netcode risk. It defers only on content grounds (30 bosses + 12 areas, :1590-1591, is the program's heaviest bespoke-encounter load) when boss content misses the §2 floor.
  4. Real-Time Stealth Tactics — fourth. RTST runs on V4Stealth/V4Perception shared with the Tactical cell's Splinter Cell and Hitman sub-rulesets (V4/V4_features.md:390-394); if RTST is failing feel, the shared stealth stack is usually the cause, and the Tactical stealth modes are already sick. Deferring RTST therefore saves little marginal engineering — it is ordered above only the cells that must not defer.
  5. Stealth — effectively never alone. The Stealth cell's plugins sit directly on the Tactical cell's shared surfaces; a Stealth deferral is a partial Tactical deferral and breaks the core promise wording (V4/V4_features.md:280-285). It may only defer as part of a program-level re-plan, not via this ladder.
  6. Tactical FPS — not deferrable. It is the spine: the largest design footprint (V4/V4_features.md:388-394), the host of the shared systems, the PvP/monetization/esports backbone, and the beachhead cell in V4/docs/planning/market-analysis.md §5. If Tactical misses gates, the correct lever is the launch date or program scope (risk R-06), decided at Studio Head level — not a cell cut.

The ladder is an order, not a quota: a defer review may conclude "hold" (extend cure period at G1/G2 only) or skip a rung upward only (a higher-numbered cell may not be deferred while a lower-numbered one is in unresolved defer review, except Tactical/Stealth which never enter the ladder normally).

4. Defer-review protocol and what "defer" means#

  • Evidence pack (assembled by the Cell Production Lead within 5 business days of a trigger): nightly Gauntlet history for the trailing 60 days, perf captures per platform, EAC/ETC history, content-inventory burn-up, beta survey data.
  • Recommendation by the Program Director; decision by Game Director + Studio Head jointly; recorded as a phase-cut ADR (per ADR 0002) within 5 business days of the decision, citing this ADR.
  • Defer semantics: the cell's plugins are stripped from cook per ADR 0005; the cell is removed from all advertised modes, store copy, and marketing within 5 business days (no "coming soon" — V4/V4_features.md:290-291, :2736); its content lands in a named post-launch season, no earlier than Season 2 (launch +90 days) and stated publicly with a date only after it has passed G2-equivalent thresholds in the post-launch track; SKU/pricing review runs if the deferral lands within 60 days of launch.
  • Kill (vs defer): a deferred cell is killed — removed from the roadmap — if it misses two consecutive committed post-launch season windows, or if its refreshed cost-to-complete exceeds its own 18-month revenue forecast (per-cell figures in V4/docs/planning/concurrency-and-monetization-forecast.md §5 scenario table). A kill is its own ADR with portfolio-level signoff.

Consequences#

  • The six-cell bet now has a pre-agreed, evidence-driven failure mode; the defer decision is mechanical and fast instead of political and late.
  • Nightly Gauntlet feel-test history becomes gate-of-record data and must be retained ≥120 days and tamper-evident.
  • Faction system, esports circuit, and SKU plans must each document their behavior under an RTS or 2D deferral (the two most likely rungs).
  • Marketing may not advertise any cell that has an open defer review at the time copy locks.
  • This ADR supplies the criteria ADR 0002 presupposed; any actual cut still produces its own phase-cut ADR referencing the evidence pack.