Campaign

Continuity-prior PINN for CSI density — the refutation, this time with rendered evidence

c-continuity-prior-csi-density-v2

Archive snapshot, as of 4 h ago — the run corpus is rebuilt once a day, so this page is not a live reading. The fleet panel is the live one; it refreshes every 30 s.

Sessions

Session	State	Runs	Synthesis	Criticism	Figures	Verdict
EZP7M9CR 2026-07-15T11:29	finished	48	✓	✓	2	EQUIVALENCE (physics prior inert). A 1D continuity+FD prior confers no practically-meaningful advantage over a fair smoothing prior for the CSI→density inverse in BOTH flow regimes: paired high-density Δ(smoothing−pinn)≈−0.003 density units, ~33× below the pre-registered min_effect_size 0.10, CI wholly inside the equivalence band [−0.10,+0.10] (6 seeds/arm, BH-FDR q small but practically null). The anti-physical FD closure matches the physical one to ≈0.0001 (closure_invariance≈0.004) → the physics is inert; any small regularization value is FD-agnostic. Directed honesty gate met (advection 0.975, ρ_peak≤2.03). Sim-only, one floor, Metal Float32 depth-3, moderate density; congested branch (ρ≳3) + real hardware untested. Confirms [[continuity-prior-regularizes-density]] is not a thesis pillar (already refuted/retired). Infra: first full IP-115 native-Metal campaign; Metal reproduces the settled Docker Float64 estimand to ±0.0001; fixed an IP-115 native-metal digest read-back bug. Adversarial critic verdict REVISE was applied before seal (equivalence reframe + per-seed table + BH-FDR + scope).
8TPR0VX6 2026-07-15T11:06	cancelled	0	✓	—	—	CANCELLED (0 runs, no compute) pending MCP-server reload — jamming fix (AVM model + give-up watchdog) is code-complete + unit-tested but the new WalkScenario fields (model/stuck_timeout_s/stuck_progress_m) are rejected by the live server's cached ParamsSchema. Reopen after restart.

Brief

Why this campaign exists

The audit's "worst evidence gap": c-continuity-prior-csi-density session 01KW074M… carried a decisive refutation (a 1D mass-conservation PINN prior does NOT beat plain smoothing for the CSI→density inverse: RMSE 1.56 / 0.91 / 1.24 / 1.24 / 1.23) entirely as an inline prose list — zero figures, no metrics.json. The natural 5-bar chart was never drawn; nothing is fetchable to verify the claim.

This re-run repeats the comparison and emits the RMSE table as metrics.json + a rendered, uploaded bar figure. The refutation is likely correct — it just has to be auditable. Sim-only; the continuity prior is a regularizer hypothesis, reported honestly whether it wins or loses.

How this executes (two phases)

Phase A — corpus. The campaign fans out coupled exp-csi-crowd runs across density regimes (crowd sizes) × ≥6 seeds, plus ≥1 advection-dominated (directed-flow) regime. where=local (sionna-csi-runner isn't on GHCR). Runs attach to the session.

Phase B — the 5-arm reduction (standalone). monad_knowledge/notebooks/python/continuity_density_pinn.py (5 arms: unconstrained · smoothing · pinn_weidmann · pinn_constant · pinn_antiphys) consumes the Phase-A run dirs and emits continuity_density_pinn.metrics.json (flat per-arm-per-regime RMSE + delta_smoothing_minus_pinn_weidmann_high + closure_invariance_high + advection_index) and the rendered fig_continuity_rmse_by_regime.png. Run it standalone (the container can't read Hetzner s3://), once per seed for the paired CI:

KMP_DUPLICATE_LIB_OK=TRUE OMP_NUM_THREADS=1 uv run --no-project --with numpy --with pandas \
  --with pyarrow --with matplotlib --with torch --with PyYAML \
  python monad_knowledge/notebooks/python/continuity_density_pinn.py <config-seed-k.yaml>

Reduce delta_smoothing_minus_pinn_weidmann_high across seeds with sim_reduction_run(multi_seed_ci) for the estimand's CI (<0 ⇒ smoothing wins). closure_invariance_high ≈ 0 is the killer argument (physics adds nothing over smoothing). The rendered figure + metrics.json must attach to the session before seal.

Pre-registered physics expectation (adversarial-review, 2026-07-15): the refutation is expected to replicate in the static/crush regime — that is a replication, not a discovery, and must be framed as such. The scientifically open question is narrower: does the continuity prior ever beat smoothing under strong directed flow, where mass-conservation is binding? The sweep must include that regime or the negative is scoped "static only". Watch the seed count — the prior ran single-seed; a rendered bar chart over one seed is a prettier point estimate, not evidence.

Prior: c-continuity-prior-csi-density (prose-only refutation — replicate with rendered evidence).