# Stepwise gameplay validation

Latest checkpoint: **verified win at frame 59,492**, clean exit, using the
isolated stepwise candidate. The original full pass and limitations were recorded
in `STEPWISE_WIN_2026-09-20.md`, retained in the private recovery archive.
Final binary also lost once; broad reliability and demo readiness remain open.

User direction: one scoped change, accelerated real match, audit, then advance.
Battlecruiser dominance is a desired outcome, not a reason to declare an
unvalidated tech path successful or delay an available ground win.

## Promotion discipline

- Keep the preserved winning artifact and experimental work intact.
- Record the exact native binary and controller used. A native-only control
  with a newer Python coach is not a full historical-stack reproduction.
- Make one behavior change at a time; do not simultaneously replace combat,
  macro, scouting and composition.
- Run the real autonomous Steamhammer with UI enabled, accelerated simulation,
  capped paid requests and no assistant-issued in-game orders.
- Audit result, construction, worker/production activity, supply, resource
  spending, damage conversion and termination. A timeout is not a win.
- An isolated capability can pass its behavioral gate while the match fails;
  record both. Do not promote it as a winning replacement without match proof.
- Preserve HUD/coach-strip work. Do not change presentation to hide failures.

## Step 0: native baseline control

Evidence: `runtime/evidence/stepwise-baseline-01/`.

Loaded the preserved `focused-resume-02/winning-module.dylib`, SHA-256
`929c5e73ffc5f6715b75bb064c6a13529d1f11c15ff877700e7c287dd5cdd9da`.
Current Python bridge/coach; fixed FACTORY_EXPAND, four paid requests, all
selected GROUND development; real Steamhammer/13Nexus, Destination 1.1.
Start randomness was not pinned. Source/staged-build hashes stayed stable.

**Result: loss at frame 45,356; engine verified result; cleanup exit 0.**

- Native supply-blocked frames: 356; peak minerals: 1,187.
- Observed damage: 49,584; building damage: 25,257; worker damage: 1,511.
- Enemy destruction callbacks included 121 Probes, 130 Dragoons, 19 Reavers,
  10 Gateways and 3 Nexuses. These counts do not establish net economic lead.
- Four Factories, eleven Depots and two Barracks materialized and completed.
- No sampled continuously stalled incomplete shell. This is not proof of
  successful expansion or perfect construction.
- Expansion materialized at frame 9,584, builder 163 died at 9,679, replacement
  assigned at 9,688, Command Center 261 destroyed at 9,866. Replacement builder
  182 died at 10,217. No replacement expansion completed.
- The construction lease later emitted `failed_to_materialize` at 10,828.
  That label is misleading here: engine creation/destruction and sampled HP
  prove the structure existed and was destroyed.
- At 23,040 only 412 local minerals remained. At 28,800 local minerals were
  zero with 21 workers, 13 banked minerals and 898 gas. No mineral-income
  recovery followed; workers subsequently died.

The control disproves any assumption that the preserved binary reliably wins.
Its earlier win remains valid. The demonstrated failure is loss of expansion
followed by failure to restore income; exact reasons for later site/funding
rejection require instrumentation rather than an invented resource threshold.

## Next single experiment

Expansion recovery on the baseline combat policy. First expose candidate-site
eligibility, threat evidence, funding/reservations and construction lifecycle.
Then fix the evidenced blocker without importing the whole experimental army
allocator or air-tech system. Gate: after a failed expansion, a safe replacement
is constructed and actually mined before income collapses, with ongoing army
production and measured combat outcomes. Repeat against comparable conditions;
one favorable start does not establish a win rate.

## Subsequent gated capabilities

Order may follow new match evidence, but only one behavior changes per experiment:

1. Reliable expansion/income continuation.
2. Scout lifecycle and return to productive work.
3. Army cohesion/lane passage, then separately independent squad assignments.
4. Sustained multi-base gas and composition-responsive production.
5. Executable air prerequisites and useful air-unit control.
6. Evidence-justified Battlecruiser transition, deployment, repair and Yamato,
   while retaining the ground force and economy; verified match outcome.

These are validation gates, not deterministic in-game build orders. No new
gameplay implementation was promoted during Step 0.