P1: science corrections in results and methods (UQ-05, DET-01, SCI-05, S-RE.2-4, SCI-06/INV-01, UQ-03/04)

literature/figs/demo_11_multiverse.json

@@ -2,8 +2,8 @@
  "figure": "demo_11_multiverse",
  "generator": "codameter.deviations.multiverse + fig_multiverse_full",
  "codameter_version": "0.4.0",
- "git_commit": "d3a5c18b08d95872cfa54fa796f080daec94d20a",
- "generated_utc": "2026-09-13T13:47:02+00:00",
+ "git_commit": "606f52826e348f55cf41c5c8cc97d83f7477e814",
+ "generated_utc": "2026-09-14T20:22:04+00:00",
  "python": "3.12.13",
  "numpy": "2.4.3",
  "matplotlib": "3.10.9",
@@ -1011,20 +1011,145 @@
    -0.0015953601640949972
   ],
   "sobol_rms": {
-   "estimator": 0.19172730464675428,
-   "band": 0.020594323416902805,
-   "window": 0.27548277786930814,
-   "stack": 0.1720939509947523,
-   "reference": 0.0002193875128073912
+   "estimator": 0.18463613802418932,
+   "band": 0.023801370403100576,
+   "window": 0.2695482936114043,
+   "stack": 0.16715781784273717,
+   "reference": 0.0013864692528178886
   },
   "sobol_drop": {
-   "estimator": 0.18311655669696053,
-   "band": 0.049651795492108027,
-   "window": 0.19170881230491268,
-   "stack": 0.2062488515155831,
-   "reference": 0.001916032998687386
+   "estimator": 0.06235364288721693,
+   "band": 0.02898266944784389,
+   "window": 0.055215473347404134,
+   "stack": 0.00045630955433450854,
+   "reference": 0.12776374206635088
   },
   "n_pipelines": 108,
+  "n_valid": 96,
+  "n_empty": 12,
+  "empty": [
+   {
+    "config": {
+     "estimator": "MWCS",
+     "band": "0.4\u20131",
+     "window": "4\u201314",
+     "stack": "1",
+     "reference": "fixed"
+    },
+    "reason": "MWCS needs at least three sub-windows for the delay-versus-lapse fit; the 4-14 s window holds 2 at 6 s length and 3 s step"
+   },
+   {
+    "config": {
+     "estimator": "MWCS",
+     "band": "0.4\u20131",
+     "window": "4\u201314",
+     "stack": "1",
+     "reference": "moving"
+    },
+    "reason": "MWCS needs at least three sub-windows for the delay-versus-lapse fit; the 4-14 s window holds 2 at 6 s length and 3 s step"
+   },
+   {
+    "config": {
+     "estimator": "MWCS",
+     "band": "0.4\u20131",
+     "window": "4\u201314",
+     "stack": "10",
+     "reference": "fixed"
+    },
+    "reason": "MWCS needs at least three sub-windows for the delay-versus-lapse fit; the 4-14 s window holds 2 at 6 s length and 3 s step"
+   },
+   {
+    "config": {
+     "estimator": "MWCS",
+     "band": "0.4\u20131",
+     "window": "4\u201314",
+     "stack": "10",
+     "reference": "moving"
+    },
+    "reason": "MWCS needs at least three sub-windows for the delay-versus-lapse fit; the 4-14 s window holds 2 at 6 s length and 3 s step"
+   },
+   {
+    "config": {
+     "estimator": "MWCS",
+     "band": "0.4\u20131",
+     "window": "4\u201314",
+     "stack": "45",
+     "reference": "fixed"
+    },
+    "reason": "MWCS needs at least three sub-windows for the delay-versus-lapse fit; the 4-14 s window holds 2 at 6 s length and 3 s step"
+   },
+   {
+    "config": {
+     "estimator": "MWCS",
+     "band": "0.4\u20131",
+     "window": "4\u201314",
+     "stack": "45",
+     "reference": "moving"
+    },
+    "reason": "MWCS needs at least three sub-windows for the delay-versus-lapse fit; the 4-14 s window holds 2 at 6 s length and 3 s step"
+   },
+   {
+    "config": {
+     "estimator": "MWCS",
+     "band": "0.8\u20132",
+     "window": "4\u201314",
+     "stack": "1",
+     "reference": "fixed"
+    },
+    "reason": "MWCS needs at least three sub-windows for the delay-versus-lapse fit; the 4-14 s window holds 2 at 6 s length and 3 s step"
+   },
+   {
+    "config": {
+     "estimator": "MWCS",
+     "band": "0.8\u20132",
+     "window": "4\u201314",
+     "stack": "1",
+     "reference": "moving"
+    },
+    "reason": "MWCS needs at least three sub-windows for the delay-versus-lapse fit; the 4-14 s window holds 2 at 6 s length and 3 s step"
+   },
+   {
+    "config": {
+     "estimator": "MWCS",
+     "band": "0.8\u20132",
+     "window": "4\u201314",
+     "stack": "10",
+     "reference": "fixed"
+    },
+    "reason": "MWCS needs at least three sub-windows for the delay-versus-lapse fit; the 4-14 s window holds 2 at 6 s length and 3 s step"
+   },
+   {
+    "config": {
+     "estimator": "MWCS",
+     "band": "0.8\u20132",
+     "window": "4\u201314",
+     "stack": "10",
+     "reference": "moving"
+    },
+    "reason": "MWCS needs at least three sub-windows for the delay-versus-lapse fit; the 4-14 s window holds 2 at 6 s length and 3 s step"
+   },
+   {
+    "config": {
+     "estimator": "MWCS",
+     "band": "0.8\u20132",
+     "window": "4\u201314",
+     "stack": "45",
+     "reference": "fixed"
+    },
+    "reason": "MWCS needs at least three sub-windows for the delay-versus-lapse fit; the 4-14 s window holds 2 at 6 s length and 3 s step"
+   },
+   {
+    "config": {
+     "estimator": "MWCS",
+     "band": "0.8\u20132",
+     "window": "4\u201314",
+     "stack": "45",
+     "reference": "moving"
+    },
+    "reason": "MWCS needs at least three sub-windows for the delay-versus-lapse fit; the 4-14 s window holds 2 at 6 s length and 3 s step"
+   }
+  ],
+  "cadence": 3,
   "axes": [
    "estimator",
    "band",

literature/figs/demo_12_bayes.json

@@ -2,8 +2,8 @@
  "figure": "demo_12_bayes",
  "generator": "codameter.uq_bayes._build_bayes + _fig_bayes",
  "codameter_version": "0.4.0",
- "git_commit": "d3a5c18b08d95872cfa54fa796f080daec94d20a",
- "generated_utc": "2026-09-13T13:41:32+00:00",
+ "git_commit": "c0f91f1f5d72ab3d7b4ed4be9294e7b023a372aa",
+ "generated_utc": "2026-09-13T17:05:28+00:00",
  "python": "3.12.13",
  "numpy": "2.4.3",
  "matplotlib": "3.10.9",
@@ -61,7 +61,7 @@
   },
   {
    "axes": "ax3",
-   "title": "(c) time-dependent error;  $N_{eff}$=23/229",
+   "title": "(c) time-dependent error;  $N_{eff}$=29/229",
    "xlabel": "time (years)",
    "ylabel": "$\\sigma$ (dv/v, %)",
    "lines": [
@@ -143,8 +143,8 @@
  "bayes": {
   "tau": 0.0003612992348785833,
   "s": 12.098861625892763,
-  "corr_length_days": 39.20671143105746,
-  "n_eff": 22.760567115510465,
+  "corr_length_days": 5.702100823925634,
+  "n_eff": 28.950925940431706,
   "n_epochs": 229,
   "labels": [
    "stretching (TS) 0.4-1Hz 8-28s stack10d fixed",

literature/figs/demo_18_stack_coherence.json

@@ -2,8 +2,8 @@
  "figure": "demo_18_stack_coherence",
  "generator": "codameter.synthetic_demo.fig_stack_coherence",
  "codameter_version": "0.4.0",
- "git_commit": "141e7bc25bb439420c5ee73223bf78eb4ca7a839",
- "generated_utc": "2026-09-10T18:41:24+00:00",
+ "git_commit": "c0f91f1f5d72ab3d7b4ed4be9294e7b023a372aa",
+ "generated_utc": "2026-09-13T16:54:15+00:00",
  "python": "3.12.13",
  "numpy": "2.4.3",
  "matplotlib": "3.10.9",
@@ -52,6 +52,14 @@
   "ax1/line2/x",
   "ax1/line2/y",
   "ax1/line3/x",
-  "ax1/line3/y"
- ]
+  "ax1/line3/y",
+  "data/snr0.5/median_cc",
+  "data/snr0.5/rms_pct",
+  "data/snr0.5/signed_step_error_pct",
+  "data/snr4/median_cc",
+  "data/snr4/rms_pct",
+  "data/snr4/signed_step_error_pct",
+  "data/stack_days"
+ ],
+ "step_error": "recovered minus true step amplitude, each the median of the 120 days after the event minus the median of the 120 days before (synthetic_demo.step_amplitude); negative = recovered drop deeper than the truth"
 }

paper/build_calibration_table.py

@@ -20,6 +20,17 @@
 OUT = HERE / "calibration_table.tex"
 
 
+def _mean_se(values: list[float]) -> dict:
+    n = len(values)
+    mean = sum(values) / n
+    se = (
+        (sum((v - mean) ** 2 for v in values) / (n - 1)) ** 0.5 / n**0.5
+        if n > 1
+        else None
+    )
+    return {"mean": mean, "se": se, "n": n}
+
+
 def _pm(m: dict, scale: float = 1.0, digits: int = 3) -> str:
     if m is None or m.get("mean") is None:
         return "--"
@@ -46,19 +57,35 @@ def main() -> int:
         + f" ($n={r['summary']['n_realizations']}$)"
         for r in runs
     ]
+    for r in runs:  # per-realisation max of the finite split R-hats, then mean, se
+        vals = []
+        for x in r["results"]:
+            if not x.get("ok"):
+                continue
+            finite = [
+                float(x[k])
+                for k in ("rhat_tau2", "rhat_s2", "rhat_lambda")
+                if isinstance(x.get(k), int | float) and x[k] == x[k]
+            ]
+            if finite:
+                vals.append(max(finite))
+        r["summary"]["rhat_max_realisation"] = _mean_se(vals) if vals else None
     metrics = [
         ("Member 68\\%", "member_coverage68", 1, 3),
         ("Member 95\\%", "member_coverage95", 1, 3),
+        ("Member 95\\%, held-out half", "heldout_member_coverage95", 1, 3),
         ("Posterior 95\\%", "coverage95_posterior", 1, 3),
         (r"Median $\sigma_{C_d}$ (\%)", "median_sd", 100, 3),
         ("Bias (\\%)", "shared_bias", 100, 3),
         (r"RMSE($\mu$) (\%)", "rmse", 100, 3),
         (r"Rescale $s$", "s", 1, 2),
+        (r"Correlation length $L$ (d)", "corr_length_days", 1, 1),
+        (r"Split $\hat R$, largest of three", "rhat_max_realisation", 1, 3),
         (r"Prior scale share $\tau^2$", "prior_weight_tau2", 1, 2),
     ]
     rows = []
     for label, key, scale, digits in metrics:
-        cells = [_pm(r["summary"][key], scale, digits) for r in runs]
+        cells = [_pm(r["summary"].get(key), scale, digits) for r in runs]
         rows.append(label + " & " + " & ".join(cells) + r" \\")
     body = "\n".join(rows)
     header = "Quantity & " + " & ".join(headings) + r" \\"
@@ -71,17 +98,23 @@ def main() -> int:
         "target of the proposed single-member error scale; $z=1$ for the 68\\% "
         "level and $z=1.96$ for the 95\\% level. "
         "Posterior: fraction of epochs whose 95\\% credible band on $\\mu$ contains "
-        "the truth. Rescale $s$: the fitted factor on the within-method floor of "
-        "eq.~\\ref{eq:weaver}. "
+        "the truth. Held-out half: $C_d$ fitted on a random six of the twelve "
+        "members and scored on the other six. Rescale $s$: the fitted factor on the "
+        "within-method floor of eq.~\\ref{eq:weaver}. $L$: fitted temporal "
+        "correlation length of the residuals. Split $\\hat R$: two Gibbs chains "
+        "from different seeds, each split in halves, largest of the three scale "
+        "hyper-parameters per realisation. A realisation redraws the additive "
+        "noise on one fixed generating coda; the truth and the twelve "
+        "configurations are the same in every realisation. "
         "Means with standard errors across realisations; no realisation failed; "
         "$\\sigma_{C_d}$, bias and RMSE in percent; prior $\\tau^2$ is the fraction of "
         "conditional posterior rate supplied by the prior scale term (the $\\lambda$ ratio is "
         "below 0.01 throughout). "
         "Clock drift: $4\\times10^{-5}$\\,s/day from 40\\% of the record. "
         "Shared source: a seasonal source "
         "effect warping the coda beyond 6\\,s lapse with a spurious 0.2\\% "
-        "seasonal \\dvv, seen by every configuration. Generated from "
-        "\\texttt{paper/data/calibration/} by \\texttt{paper/build\\_calibration\\_table.py}.}\n"
+        "seasonal \\dvv, seen by every configuration. Generated from the archived "
+        "runs distributed with the paper.}\n"
         "\\label{tab:calibration}\n"
         "\\begin{tabularx}{\\textwidth}{@{}L"
         + "r" * len(runs)

paper/calibration_table.tex

@@ -1,18 +1,21 @@
 \begin{table}
 \footnotesize
-\caption{Coverage calibration of the Bayesian measurement model over independent synthetic realisations of the volcano scenario (2.5 years, SNR 7, 4-day cadence, 12-member ensemble). Member coverage: fraction of member epochs with $|m_k(t)-\mathrm{truth}(t)|\le z\,\sigma_{C_d}(t)$, the target of the proposed single-member error scale; $z=1$ for the 68\% level and $z=1.96$ for the 95\% level. Posterior: fraction of epochs whose 95\% credible band on $\mu$ contains the truth. Rescale $s$: the fitted factor on the within-method floor of eq.~\ref{eq:weaver}. Means with standard errors across realisations; no realisation failed; $\sigma_{C_d}$, bias and RMSE in percent; prior $\tau^2$ is the fraction of conditional posterior rate supplied by the prior scale term (the $\lambda$ ratio is below 0.01 throughout). Clock drift: $4\times10^{-5}$\,s/day from 40\% of the record. Shared source: a seasonal source effect warping the coda beyond 6\,s lapse with a spurious 0.2\% seasonal \dvv, seen by every configuration. Generated from \texttt{paper/data/calibration/} by \texttt{paper/build\_calibration\_table.py}.}
+\caption{Coverage calibration of the Bayesian measurement model over independent synthetic realisations of the volcano scenario (2.5 years, SNR 7, 4-day cadence, 12-member ensemble). Member coverage: fraction of member epochs with $|m_k(t)-\mathrm{truth}(t)|\le z\,\sigma_{C_d}(t)$, the target of the proposed single-member error scale; $z=1$ for the 68\% level and $z=1.96$ for the 95\% level. Posterior: fraction of epochs whose 95\% credible band on $\mu$ contains the truth. Held-out half: $C_d$ fitted on a random six of the twelve members and scored on the other six. Rescale $s$: the fitted factor on the within-method floor of eq.~\ref{eq:weaver}. $L$: fitted temporal correlation length of the residuals. Split $\hat R$: two Gibbs chains from different seeds, each split in halves, largest of the three scale hyper-parameters per realisation. A realisation redraws the additive noise on one fixed generating coda; the truth and the twelve configurations are the same in every realisation. Means with standard errors across realisations; no realisation failed; $\sigma_{C_d}$, bias and RMSE in percent; prior $\tau^2$ is the fraction of conditional posterior rate supplied by the prior scale term (the $\lambda$ ratio is below 0.01 throughout). Clock drift: $4\times10^{-5}$\,s/day from 40\% of the record. Shared source: a seasonal source effect warping the coda beyond 6\,s lapse with a spurious 0.2\% seasonal \dvv, seen by every configuration. Generated from the archived runs distributed with the paper.}
 \label{tab:calibration}
 \begin{tabularx}{\textwidth}{@{}Lrrr@{}}
 \toprule
 Quantity & clean ($n=200$) & shared source ($n=200$) & clock drift ($n=200$) \\
 \midrule
 Member 68\% & 0.808 $\pm$ 0.001 & 0.795 $\pm$ 0.001 & 0.804 $\pm$ 0.001 \\
 Member 95\% & 0.9561 $\pm$ 0.0003 & 0.9496 $\pm$ 0.0004 & 0.9543 $\pm$ 0.0003 \\
+Member 95\%, held-out half & 0.937 $\pm$ 0.004 & 0.930 $\pm$ 0.004 & 0.936 $\pm$ 0.004 \\
 Posterior 95\% & 0.590 $\pm$ 0.004 & 0.340 $\pm$ 0.004 & 0.604 $\pm$ 0.004 \\
 Median $\sigma_{C_d}$ (\%) & 0.225 $\pm$ 0.001 & 0.330 $\pm$ 0.001 & 0.222 $\pm$ 0.001 \\
 Bias (\%) & -0.017 $\pm$ 0.001 & -0.045 $\pm$ 0.001 & -0.019 $\pm$ 0.001 \\
 RMSE($\mu$) (\%) & 0.0669 $\pm$ 0.0005 & 0.161 $\pm$ 0.001 & 0.0656 $\pm$ 0.0005 \\
 Rescale $s$ & 11.67 $\pm$ 0.03 & 12.70 $\pm$ 0.04 & 10.88 $\pm$ 0.03 \\
+Correlation length $L$ (d) & 6.11 $\pm$ 0.03 & 6.18 $\pm$ 0.04 & 6.16 $\pm$ 0.03 \\
+Split $\hat R$, largest of three & 1.024 $\pm$ 0.002 & 1.036 $\pm$ 0.005 & 1.030 $\pm$ 0.004 \\
 Prior scale share $\tau^2$ & 0.09 $\pm$ 0.01 & 0.025 $\pm$ 0.005 & 0.08 $\pm$ 0.01 \\
 \bottomrule
 \end{tabularx}

paper/data/bayes/README.md

@@ -0,0 +1,15 @@
+# Checks of the Bayesian measurement model (audits UQ-03 and UQ-04)
+
+Both files are produced from the manuscript's Fig. 14 realisation
+(`codameter.uq_bayes._build_bayes`: seed 55, 2.5 years, SNR 7, 4-day cadence,
+12-member ensemble).
+
+- `duplication_check.json`, from `scripts/check_ensemble_duplication.py`: the
+  credible band on mu, the C_d scale, s, tau and the share of the precision of
+  mu supplied by the smoothness prior, for the ensemble as measured and with
+  every member duplicated and quadrupled.
+- `cd_whitening.json`, from `scripts/check_cd_whitening.py`: raw and whitened
+  lag autocorrelations of the member residuals about mu, the whitened variance
+  under the full C_d and under its diagonal alone.
+
+Rerun the two scripts to regenerate; each takes a few minutes.