MobilityLab science battery

name: MobilityLab science battery overview: "Extend MobilityLab from a single treated/CF contrast into a reproducible science battery: tax shock with baseline window, dose–response grid, high-tax placebo twin, multi-seed ensemble, and economy/fiscal estimands—still app-local on Wave 1 kernels." todos:

• id: study-spec

content: Add StudySpec, ArmKind, tax schedule/shock fields; BatteryRunner executing primary/CF/placebo/dose/seeds status: completed

• id: estimands

content: "Extend ScientificEvaluator: event-study, fiscal/prod ATT, placebo + dose + ensemble aggregates and study-level checks" status: completed

• id: report-ui

content: BatteryMarkdownReport + MainWindow Battery mode (dose chart, shock marker, study scorecard); headless --headless / --single status: completed

• id: docs-verify

content: README study design; build + headless 48 smoke for dose/placebo/ensemble signs status: completed isProject: false

Problem

Today’s session answers one binary question well (ATT of α tax 0.38 vs CF). That confirms a kernel coupling; it does not explore a policy surface, test symmetry, separate levels vs shocks, or report economy/fiscal consequences. PASS/FAIL stays easy because the model is built to pass.

Design goal (locked)

Keep formulas in Civics/Economy/Geopolitics. MobilityLab becomes a design + estimation harness: multiple worlds per study, shared evaluators, comparative markdown. No kernel Wave 2 (unrest/HD/productivity) in this tranche—document those as later.

flowchart TB
  study [StudySpec]
  battery [BatteryRunner]
  worlds [World runs: baseline shock dose placebo seeds]
  eval [ScientificEvaluator per world]
  aggregate [BatteryReport ATT curves placebo ensemble]
  study --> battery --> worlds --> eval --> aggregate

Core abstractions (new)

Under `d:\novolis\novolis-dogfooding\apps\avalonia\MobilityLab\Experiment\`:

Extend `ExperimentSpec`:

• BaselineMonths, ShockMonth (0 = static treatment from t0; default shock at month 12 after baseline at β tax)
• AlphaTaxSchedule resolved by host each month (baseline → treatment), still locked after optional agents
• Keep existing static mode as ShockMonth = 0 with constant α tax for back-compat

Arm battery (always run on Study Run)

1. Primary — α = treatment tax (static or post-shock); β/γ low; war/agents off
2. Counterfactual — α = β tax (existing ATT baseline)
3. Placebo high twin — α = β = treatment tax; γ unchanged. Tests whether “twin” is real or only Gamma gravity: expect both α and β to lose; DID ≈ 0; Gamma absorb still high
4. Dose grid — α ∈ {0.22, 0.28, 0.32, 0.38, 0.45} (straddle Economy tax-push ~0.28); each with CF at β tax; plot ATT pop % and ATT mean push vs tax
5. Seed ensemble — primary+CF for seeds {42,43,44}; report mean/min/max ATT pop % (robustness, not inference)

Default study: 48 months, baseline 12, shock at 12, primary α 0.38, grid as above, 3 seeds. Fast enough for session + headless.

Estimands (extend `EffectSizes` / evaluator)

Reuse `ScientificEvaluator`; add:

• Shock event-study (primary arm): mean α net mig / push / approval in [1, baseline] vs [shock, shock+12]; pre-trend check (α−β pop growth near 0 pre-shock)
• Economy/fiscal ATT (fields already on PolityFacts): ATT on cumulative TaxCollected, mean ProductionValue, end StateCash — so the report is not migration-only
• Placebo diagnostics: |DID| < ε under placebo high; fail identification-style check if placebo DID mimics primary
• Dose summary: tax at which ATT pop % first crosses −5% and −20%; monotonicity flag (Spearman ATT vs tax)

Coupling checks become study-level (not only single-arm 6/6):

UI / report

`MainWindow.cs`:

• Mode toggle: Single (current) vs Battery (default for Run)
• Battery controls: months, baseline/shock, dose grid on/off, seeds, placebo on/off
• Charts: (1) primary pop series with shock marker; (2) dose curve ATT pop % vs α tax; (3) ensemble ATT range
• Scorecard = study-level checks; metrics panel = primary ATT + placebo DID + dose threshold

`MarkdownReport.cs` → BatteryMarkdownReport:

• Spec + identification
• Primary ATT/DID + event-study table
• Dose–response table
• Placebo table
• Ensemble table
• Economy/fiscal ATT
• Series snapshot (primary only)

Headless: --headless 48 runs battery; --single keeps old one-arm path.

Docs

Update `README.md`: study design, arm definitions, how to read dose/placebo/ensemble, link demography-coupling. Explicitly: kernel Wave 2 (unrest/HD) remains out of scope.

Verification

dotnet build d:\novolis\novolis-dogfooding\apps\avalonia\MobilityLab\MobilityLab.csproj -p:NovolisUseProjectReferences=true
dotnet run --project d:\novolis\novolis-dogfooding\apps\avalonia\MobilityLab\MobilityLab.csproj -p:NovolisUseProjectReferences=true -- --headless 48

Expect: primary ATT negative; dose curve more negative at higher tax; placebo DID near zero relative to primary; ensemble same sign; pre-trend check PASS with shock design.