docs(research): swiss mc dlinear/patchtst results + research-critic audit

36-cell multi_channel results (read from real results.json). Audit-corrected
framing: DLinear shows no measurable identifier effect (flat, val mirrors
test); on PatchTST the INJECTION POINT drives it — add_after_patch embedding
helps (~-5%), pre-patch concat_to_x transparent fusion hurts (+30-60%). NOT
claimed: "transparent harmful" (embedding is also an identifier and helps) or
"DLinear immune" (can't separate model from projection). Key caveat: mc
transparent results reflect ChannelTransparentWrapper's fusion projection,
not intrinsic identifier value. Contrast table vs per_entity LSTM notes the
split-mode confound. Follow-up ablations listed.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Author: jajupmochi

docs/research/2026-06-14-swiss-mc-results.md

@@ -0,0 +1,92 @@
+# swiss-river × {DLinear, PatchTST} entity-identifier results (multi_channel, 2026-06-14)
+
+DLinear and PatchTST on the three Swiss datasets in **multi_channel** split
+(each station is a channel — identity is already implicit in the channel
+layout, unlike the per_entity LSTM where the identifier is the *only*
+identity signal). Single seed 2026, 50-trial HPO per cell, batch_size fixed
+32, 30 epochs. Numbers read from real `results.json` by
+`tools/plot_swiss_mc_results.py` (no hand-copied values);
+CSV: `figures/swiss-mc-2026-06-14/swiss-mc-rmse.csv`.
+
+> These results only became measurable after fixing four multi_channel bugs
+> on 2026-06-14 (commits `fdecec1`, `82f0903`): before that, the transparent
+> identifier modes were a silent no-op (== none) and the 2010/zurich cells
+> were all-NaN. The 2026-05 dlinear/patchtst transparent results are
+> **retracted** (they measured the no-op).
+
+## Results — test RMSE (°C, denormalized)
+
+![RMSE by mode](figures/swiss-mc-2026-06-14/swiss-mc-rmse-by-mode.png)
+
+| model | dataset | none | embedding | onehot | random | sinusoidal | coordinates |
+|---|---|---|---|---|---|---|---|
+| DLinear | 1990 | 1.281 | 1.286 | 1.286 | 1.283 | 1.282 | 1.279 |
+| DLinear | 2010 | 1.355 | 1.356 | 1.356 | 1.361 | 1.358 | 1.357 |
+| DLinear | zurich | 1.391 | 1.401 | 1.396 | 1.393 | 1.393 | 1.385 |
+| PatchTST | 1990 | 1.374 | **1.304** | 2.189 | 2.059 | 2.108 | 1.815 |
+| PatchTST | 2010 | 1.488 | **1.387** | 2.505 | 2.345 | 2.340 | 2.036 |
+| PatchTST | zurich | 1.480 | **1.388** | 2.738 | 2.631 | 2.690 | 2.377 |
+
+## What we can and cannot claim (research-critic audit)
+
+**Defensible (single seed; val mirrors test, so not overfitting):**
+
+- **DLinear shows no measurable identifier effect** on swiss multi_channel:
+  all six modes land within ~1% of `none` on every dataset (HPO val is
+  equally flat, e.g. 1990 val 0.1103–0.1105 across modes). A linear
+  decomposition model over channels has no head-room to exploit per-channel
+  identity, and the wrapper's fusion projection adds none.
+- **On PatchTST, the injection POINT — not the identifier type — drives the
+  result.** `embedding` via `add_after_patch` (injected *after* patch
+  embedding) helps (~−5% vs none). The transparent modes, injected via
+  `concat_to_x` + `ChannelTransparentWrapper`'s per-channel
+  `Linear(1+D,1)` fusion *before* patching, hurt substantially (+30–60%),
+  and they hurt in HPO val too (1990 onehot val 0.231 vs none 0.124).
+
+**NOT claimed (would over-reach):**
+
+- *"Transparent identifiers are harmful."* Too broad — `embedding` is also
+  an identifier and it helps. What hurts on PatchTST is the **pre-patch
+  `concat_to_x` fusion**, which perturbs the per-channel series that
+  PatchTST then slices into patch tokens. `add_after_patch` avoids this.
+- *"DLinear is immune to identity."* We observe no effect; we cannot tell
+  apart "the linear model can't use identity" from "the fusion projection
+  discards it" without inspecting the projection weights.
+- Anything cross-domain — all three datasets are Swiss water temperature.
+
+**Key caveat (read before citing):** the multi_channel transparent results
+reflect the behaviour of **`ChannelTransparentWrapper`** (concat + learnable
+per-channel fusion projection), *not* the intrinsic value of the identifier
+information. This wrapper is structurally different from the per_entity
+`EntityTransparentWrapper` (zero-parameter concat). Cross-split comparisons
+must account for this.
+
+## Contrast with the per_entity LSTM (2026-06-13)
+
+| split | model | identifier effect |
+|---|---|---|
+| per_entity | LSTM | every identifier helps (−25 to −35% vs none) — identity is the ONLY identity signal |
+| multi_channel | DLinear | none (≈ flat) — identity already in the channel layout |
+| multi_channel | PatchTST | `add_after_patch` embedding helps; pre-patch transparent injection hurts |
+
+The dominant variable across these is the **split mode**, not the model:
+per_entity makes each sample identity-blind (so the identifier is decisive),
+while multi_channel already encodes identity in the channel index (so an
+extra identifier is at best redundant, at worst disruptive to the
+architecture). This matches the split-mode confound flagged in the 2026-05
+report.
+
+## To harden / explain (follow-ups)
+
+- **Ablation isolating injection point from identifier type**: run PatchTST
+  + transparent + `add_after_patch` — if it stops hurting, the cause is the
+  pre-patch fusion, confirming the audit's reading.
+- **Inspect the DLinear fusion-projection weights** to tell "can't use" from
+  "discards".
+- **Multi-seed** (≥3) for variance bands (shared with the lstm follow-up #32).
+
+## Reproduce
+
+```bash
+python tools/plot_swiss_mc_results.py
+```

docs/research/figures/swiss-mc-2026-06-14/swiss-mc-rmse-by-mode.png

@@ -0,0 +1,7 @@
+model,dataset,none,embedding,onehot,random,sinusoidal,coordinates
+dlinear,swiss-river-1990,1.2814,1.2860,1.2860,1.2827,1.2817,1.2790
+dlinear,swiss-river-2010,1.3545,1.3560,1.3558,1.3606,1.3576,1.3572
+dlinear,swiss-river-zurich,1.3905,1.4008,1.3964,1.3927,1.3933,1.3851
+patchtst,swiss-river-1990,1.3737,1.3035,2.1889,2.0590,2.1075,1.8148
+patchtst,swiss-river-2010,1.4884,1.3865,2.5051,2.3452,2.3402,2.0356
+patchtst,swiss-river-zurich,1.4803,1.3884,2.7380,2.6311,2.6900,2.3773

docs/research/figures/swiss-mc-2026-06-14/swiss-mc-rmse.csv

@@ -0,0 +1,80 @@
+#!/usr/bin/env python3
+"""Plot swiss multi_channel (dlinear/patchtst) entity-identifier results.
+
+Reads denorm test RMSE from real results.json (no hardcoded numbers) and
+emits one grouped-bar figure per model (3 datasets × 6 modes) with the
+`none` baseline drawn as a reference line. Also writes a CSV.
+"""
+from __future__ import annotations
+
+import csv
+import glob
+import json
+from pathlib import Path
+
+import matplotlib
+
+matplotlib.use('Agg')
+import matplotlib.pyplot as plt
+
+ROOT = Path(__file__).resolve().parents[1]
+ART = ROOT / 'artifacts' / 'entity_identifier'
+OUT = ROOT / 'docs' / 'research' / 'figures' / 'swiss-mc-2026-06-14'
+OUT.mkdir(parents=True, exist_ok=True)
+
+DATASETS = ['swiss-river-1990', 'swiss-river-2010', 'swiss-river-zurich']
+MODELS = ['dlinear', 'patchtst']
+MODES = ['none', 'embedding', 'onehot', 'random', 'sinusoidal', 'coordinates']
+TAG = {d: f'swiss-mc-{d.split("-")[-1]}-20260614' for d in DATASETS}
+SHORT = {'swiss-river-1990': '1990 (28)', 'swiss-river-2010': '2010 (63)',
+         'swiss-river-zurich': 'zurich (15)'}
+COLORS = {'none': '#9aa0a6', 'embedding': '#1a73e8', 'onehot': '#34a853',
+          'random': '#a142f4', 'sinusoidal': '#ea8600', 'coordinates': '#d93025'}
+
+
+def rmse(tag, ds, model, mode):
+    hits = glob.glob(f'{ART}/{tag}/{ds}-{model}-{mode}-*/**/results.json', recursive=True)
+    if not hits:
+        return None
+    m = json.load(open(hits[0]))['metrics']['test']
+    return float(m.get('denorm_rmse', m.get('rmse')))
+
+
+data = {model: {ds: {mode: rmse(TAG[ds], ds, model, mode) for mode in MODES}
+                for ds in DATASETS} for model in MODELS}
+
+with open(OUT / 'swiss-mc-rmse.csv', 'w', newline='') as fh:
+    w = csv.writer(fh)
+    w.writerow(['model', 'dataset', *MODES])
+    for model in MODELS:
+        for ds in DATASETS:
+            cells = data[model][ds]
+            row = [f'{cells[m]:.4f}' if cells[m] is not None else '' for m in MODES]
+            w.writerow([model, ds, *row])
+print('wrote', OUT / 'swiss-mc-rmse.csv')
+
+fig, axes = plt.subplots(1, 2, figsize=(13, 4.8), sharey=False)
+for ax, model in zip(axes, MODELS):
+    n = len(MODES)
+    bw = 0.82 / n
+    for j, mode in enumerate(MODES):
+        xs = [i + (j - n / 2 + 0.5) * bw for i in range(len(DATASETS))]
+        ys = [data[model][ds][mode] for ds in DATASETS]
+        ax.bar(xs, ys, bw, label=mode, color=COLORS[mode], edgecolor='white', linewidth=0.5)
+    # none reference line per dataset (thin) — visual anchor
+    for i, ds in enumerate(DATASETS):
+        nb = data[model][ds]['none']
+        ax.plot([i - 0.45, i + 0.45], [nb, nb], color='#444', lw=0.8, ls='--', alpha=0.6)
+    ax.set_xticks(range(len(DATASETS)))
+    ax.set_xticklabels([SHORT[d] for d in DATASETS])
+    ax.set_title(f'{model}  (multi_channel)')
+    ax.set_ylabel('Test RMSE (°C)')
+    ax.grid(axis='y', alpha=0.3)
+    for sp in ('top', 'right'):
+        ax.spines[sp].set_visible(False)
+axes[0].legend(ncol=6, fontsize=8, loc='upper center', bbox_to_anchor=(1.05, -0.12), frameon=False)
+fig.suptitle('swiss-river × {dlinear, patchtst}: entity-identifier modes '
+             '(single seed 2026, 50-trial HPO; dashed = none baseline)', fontsize=11)
+fig.tight_layout()
+fig.savefig(OUT / 'swiss-mc-rmse-by-mode.png', dpi=150, bbox_inches='tight')
+print('wrote', OUT / 'swiss-mc-rmse-by-mode.png')