Skip to content
MECHA SKIRMISHA game of parts & pilots Working prototypePrototype guide ↗
Explore Mecha Skirmish Menu +
Mecha Skirmish / Independent review: map diagnostic by mirror familyPrototype field notes / v0.6

Independent review: map diagnostic by mirror family

The retained map-diag-family-analysis.json and Markdown tables are approved as descriptive analysis of the completed old/fixed mirror experiment. This review used the retained observations only; no games were rerun and no simulation source changed.

Verified independently:

  • Both input SHA-256 hashes match the referenced files.
  • Each source contains the exact 18 map/family cells, seeds 800000–800063, orientation zero, aggression .75/.75, and both first-side assignments: 2,304 observations per source. All eight pairing keys match.
  • All 144 metric means and paired changes, and all 288 interval endpoints, reproduce from the raw rows. The 2,000 bootstrap samples resample whole shared seed blocks, retaining both first sides and all map/family cells.
  • Causeway armored physical-A payoff is 77.34375% old and 43.75% fixed. First-side payoff is 50.78125% and 23.4375%. Quarry mixed fixed physical-A payoff is 60.15625%.
  • All 128 Foundry melee observations in each source are draws at round eight. Every row has zero deaths, proactive Fires, melee starts, zone points and units making a first proactive attack. Movement-event counts are 33 old and 43 fixed. Describe these as non-engaging draws, not completely inactive games or evidence of healthy balance.

Pooled map means can conceal these roster-specific observations. The intervals describe old-to-fixed changes, not confidence intervals for a fixed-map payoff's departure from 50%. They are exploratory and unadjusted across the many displayed metrics. The intervention combines terrain/objective repairs, including three to four scoring zones; it does not isolate terrain-only causality or establish human fairness.

One script-hardening follow-up was reported: its initial checks require matching keys, 2,304 rows and 64 seeds but do not by themselves enforce the entire declared grid. The independent check above verifies that the retained files do contain that grid. Before reusing the parser on other inputs, require the exact map/family/seed/first-side/policy set so equally malformed old/fixed files cannot pass together.

Evidence: docs/tuning/evidence/2026-09-09-multivariant/map-diag-family-analysis.{json,md} and its referenced map-diag-old-run/baseline.json / map-diag-fixed-run/baseline.json. Analyzer: cli/analyze-map-diagnostic-families.py.