m monad-knowledge FIIT STU · open research vault
Campaign

Defend the L4½.5 mechanism claim across the EXP-S1 sandbox

c-exp-s1-mechanism-defense · exps1-continuity-penalty

Sessions

No sessions yet — this campaign has a brief but no execution.

Brief

Question

We need to defend the thesis claim L4½.5 — that adding a continuity-residual penalty $\lambda \cdot | \partial\rho/\partial t + \nabla\cdot(\rho\mathbf{v}) |_2^2$ to the supervised loss reduces per-cell occupancy MAE because of the physics prior, not because of an incidental regularisation effect.

The risk that reviewers will dismiss the result as "any regulariser would have done the same" is real. The defence has three parts:

  1. The lift is meaningfully large (not a 0.01 ped/cell rounding-error effect).
  2. The lift transfers across architectures (not a quirk of one model class).
  3. The lift survives a density transfer (not specific to the calibration regime).

Each part maps to one success_criterion in the frontmatter above.

What we already know

  • EXP-S1 Phase A has shown the JuPedSim + PedPy + log-distance-BLE harness runs end-to-end in under 5 min CPU on the seminar-room geometry. The Phase A smoke test is the verification artefact for IP-078; we inherit it here as the trust anchor.
  • The λ-sweep range $\lambda \in [0, 1]$ is reasonable per di2023_285b ↗. λ much larger destabilises training.
  • LGBM, PINN, and CNN+LSTM are the three architectures we plan to defend across (per the EXP-S1 model zoo).
  • A density of N=8 is the planned calibration regime (per the EXP-S1 setup). N=4 and N=12 bracket it for the transfer check.

What I want the supervisor to do

Plan a sub-experiment sequence that addresses the three criteria in order:

  1. Criterion 1 (lift magnitude) — sweep λ at one density (N=8) on Geometry A across all 3 model architectures with a small number of seeds (3 is sufficient given the architecture × λ Cartesian product is already wide). Report Δ-MAE(λ*) per architecture.

  2. Criterion 2 (lift sign transfer) — verify the sign is positive across all 3 architectures at λ*. This may be folded into Criterion 1's data if the sweep covers all three architectures cleanly; otherwise run a targeted 3-architecture batch at λ*.

  3. Criterion 3 (density transfer) — if Criteria 1 + 2 hold, scan N ∈ {4, 8, 12} at λ* on the strongest architecture from Criterion 1. If the strongest architecture is ambiguous, scan two architectures (this consumes more budget; skip if borderline budget).

Replan once: if Criterion 1's lift is borderline on one architecture, tighten the λ-grid around that architecture's apparent λ* and re-sweep. If the lift is decisively absent on one architecture, the corpus speaks — surface that in the synthesis rather than hunting for a passable λ.

Stop when the budget runs out or all three success criteria are decisively resolved.

Out of scope

  • Hardware fidelity. The EXP-S1 BLE model is a log-distance + log-normal-shadowing proxy, not measured BLE. Hardware-fidelity defence lives in EXP-F1 / EXP-F2 / EXP-F3, not here.
  • Sionna RT replacement of the CSI proxy. EXP-S1 Phase C is the venue for that follow-on; if Criterion 3 is borderline and the supervisor judges a Sionna spot-check might resolve it, mention that in the synthesis's "follow-on" paragraph rather than spawning a Sionna run from this campaign (the EXP-S1 Phase C harness is not in simulations_available).
  • Geometry transfer to a corridor / cross-geometry layout. Geometry B follow-on lives in EXP-S1 Phase B'; surface it as a future-considerations note in the synthesis if Criterion 2 is decisive on Geometry A.

Expected interpretation

If all three criteria resolve, the synthesis should read: "L4½.5 is supported on Geometry A across the model zoo and across the calibration-regime density transfer. The mechanism transfers; the model-specific magnitudes do not. The thesis chain link ch4-half-physics-informed-fusion → L4½.5 can be footnoted with this session as supporting evidence."

If Criterion 3 is the gap, the synthesis should read: "L4½.5 is supported at the calibration density (N=8); the density transfer remains the open gap. A follow-on session targeting the Geometry-B transfer or the high-density (N≥12) regime would close it. Defer to EXP-F once hardware is up."*

If Criteria 1 or 2 fail, the synthesis should read: "L4½.5 is not supported on the EXP-S1 sandbox at the tested parameters. The architecture is the variance source, not the physics prior. Reconsider the loss term or the architectural choice." — and the supervisor should label this honestly without trying to rescue the claim with parameter-tweaking.

Status — 2026-07-06 overhaul triage

Blocked (infrastructure). The reduction notebook (exps1_continuity_penalty.py, runner python-ml) needs torch + lightgbm, and no runner image ships them (see EXP-S1 continuity-penalty findings, 2026-06). Unblock condition: build the python-ml runner image with torch/lightgbm (and the macOS-only OMP guard), then this campaign is dispatchable as-is. Not archived — the mechanism-defense question remains open and chartered.