Anthropogenic attribution ratio is approaching 1.0

Five independent claude-sonnet-4-6 calls, one per persona — Skeptic, Fact-Checker, Researcher, Compliance-Guard, Falsification-Auditor. Each writes its hardest objection from its own seat, paired with the methodological resolution it would accept. Run on the static catalogue spec Jun 3, 2026; a live council for any topic is at /try.

1. Skeptic #01
   Raised

   The observed inflation of the attribution ratio toward 1.0 is consistent with an artifact of low-frequency internal variability—specifically the Atlantic Multidecadal Oscillation (~60-year cycle) and the Interdecadal Pacific Oscillation (~20–30 year cycle)—being aliased into the forced-signal component within a 30-year rolling window that cannot spectrally resolve these modes from a monotonic anthropogenic trend. CMIP6 free-running model ensembles used to estimate the "natural variability variance" denominator systematically desynchronize internal ocean-atmosphere modes from the observed phase, compressing their apparent amplitude and thus artificially shrinking the denominator of the metric independent of any genuine increase in forcing dominance. This produces an attribution ratio approaching 1.0 as a statistical artifact of window length and ensemble phase incoherence rather than as evidence of anthropogenic signal saturation.

   Resolved

   Execute pacemaker (nudged-SST) experiments using CESM2-Large Ensemble or GFDL-ESM4 with observed AMO and IPO indices prescribed from HadSST4 or ERSSTv5, then recompute the attribution ratio on the same rolling 30-year windows; if the ratio drops below 0.95 when internal modes are phase-locked to observations, the causal claim is undermined. Cross-validate the natural variability denominator against the PAGES 2k Consortium temperature reconstruction (spanning 0–2000 CE) using the pre-industrial 850–1850 CE window as the unforced baseline to confirm that the free-running CMIP6 ensemble is not underestimating low-frequency internal variance; a statistically significant difference (p < 0.05, F-test on variance ratios) between PAGES-derived and CMIP6-derived natural variability estimates would settle the confound in favor of the aliasing alternative.

2. Fact-Checker #02
   Raised

   The dominant uncertainty enters through the natural-variability variance term, which any IPCC-protocol detection-and-attribution study must estimate from GCM pre-industrial control simulations; CMIP6 piControl runs yield inter-model spread in 30-year global mean temperature standard deviation of roughly 0.06–0.12 °C — a factor-of-two range that alone propagates to ±0.04–0.08 (1σ) uncertainty on the 1 − σ²_nat/σ²_total ratio. The temperature endpoint (HadCRUT5 v5.0.1.0) carries a 2σ ensemble spread of ±0.06 °C on global annual means for recent decades, growing to ±0.15 °C pre-1900, adding structural variance error of order ±0.01 °C² to σ²_total; and the solar-cycle endpoint (SORCE/TIM absolute accuracy ±0.5 W/m² against a solar-cycle amplitude of ~1.1 W/m², with ACRIM/VIRGO cross-calibration offsets of ~0.5 W/m²) introduces ±15–20% uncertainty in the natural-forcing variance component. Summing these in quadrature, the attribution ratio threshold of 0.95 falls within roughly 1–2σ of both 0.90 and 1.0, rendering it statistically indistinguishable from either bound and therefore not independently falsifiable as stated.

   Resolved

   The threshold should be revised to ≥0.85 with an explicit ±0.05 half-width representing the irreducible GCM natural-variability floor, evaluated via two-step uncertainty propagation: first, apply the Allen–Tett optimal-fingerprinting regression with a residual-consistency (F-ratio) test using CMIP6 historical and piControl ensembles of N≥30 models to produce a posterior distribution on the attributable fraction; second, construct a bootstrapped 90% CI on the rolling 30-year ratio using HadCRUT5's full 200-member observational ensemble, restricting analysis to years satisfying the HadCRUT5 quality flag "coverage_fraction ≥ 0.90" to exclude sparse-sampling bias. The FALSIFIES condition should require the lower bound of this 90% CI to fall and remain below 0.80 for ≥5 consecutive years — a signal large enough to clear the combined GCM variance-scaling and observational-structural uncertainty floor by at least 1σ.

3. Researcher #03
   Raised

   The methodology designates /api/solarcycle as the operative natural-variability signal, but this collapses the entire natural-variability term onto external solar forcing alone, omitting multi-decadal internal ocean-atmosphere modes—specifically the Atlantic Multidecadal Oscillation (AMO) and Pacific Decadal Oscillation (PDO)—whose dominant periodicity (20–70 years) is commensurate with the 30-year rolling window. Because these modes are excluded from σ²_natural yet remain embedded in σ²_total, any window that coincides with a positive AMO or PDO phase will suppress the natural-variability fraction and inflate the attribution ratio toward 1.0 artifactually, without any change in the true anthropogenic signal. This upward bias in the computed ratio then propagates directly into the legal-damages and EU ETS predictions, which inherit a spuriously narrow confidence interval around the attribution fraction.

   Resolved

   The variance decomposition must incorporate AMO and PDO indices as explicit orthogonal covariates before partitioning residual variance into anthropogenic versus natural components; the NOAA PSL AMO index (ERSSTv5-derived, available at psl.noaa.gov/data/timeseries/AMO/) and the NCEI PDO index provide monthly resolution back to 1854 and can be differenced to remove forced trend contamination following the Ting et al. (2009) regression-removal protocol. Alternatively, the null distribution for natural variability should be constructed from CMIP6 pre-industrial control run (piControl) ensembles accessed via ESGF nodes—filtered to the AR6 Annex II assessed model set—which by construction encode all internal modes; regressing the observed record onto the forced response using the piControl-derived covariance matrix as the noise model yields an attributable fraction and confidence interval that cannot be inflated by unmodeled internal variability phases.

4. Compliance-Guard #04
   Raised

   Premature citation of this hypothesis as established fact — before the metric sustainably exceeds the 0.95 SUPPORTS threshold — creates direct exposure under FRE 702/Daubert in U.S. federal climate-liability litigation, where expert testimony built on a below-threshold detection-and-attribution ratio could be excluded or sanctioned and would specifically undermine the very causation-element shift the hypothesis predicts across N≥50 filings. Under SEC Rule 10b-5 and the SEC's Final Rules on Climate-Related Disclosures (17 CFR Part 229), any registrant or investment advisor incorporating a premature attribution claim into material physical-risk characterizations faces securities-fraud exposure for overstating scientific confidence. In the EU, using the sub-threshold ratio as trigger logic for Article 8 EU Climate Law (2021/1119) enforcement or as a CSRD/ESRS E1 double-materiality input before IPCC-aligned peer review is complete exposes those enforcement actions and covered undertakings to legal challenge, while the predicted EU ETS repricing could attract scrutiny under MAR (EU Regulation 596/2014) if market participants act on an unvalidated threshold.

   Resolved

   Reliance is gated on three sequential conditions: full peer review and independent replication of the attributable-fraction computation in a recognized detection-and-attribution venue meeting IPCC AR6 WGI Chapter 3 methodological standards, using at least one dataset wholly independent of the /api endpoints employed here; sustained exceedance of the 0.95 ratio for a continuous five-year window with confidence intervals formally excluding natural-variability dominance at p < 0.05; and attachment of a mandatory disclaimer to any litigation filing, securities disclosure, or regulatory submission stating that this metric remains a falsifiable experimental hypothesis that has not crossed its SUPPORTS threshold and must not be cited as settled science until independent replication and the sustained-threshold conditions are formally certified in writing by the responsible scientific team.

5. Falsification-Auditor #05
   Raised

   The stated FALSIFIES condition — "natural-variability variance growing share of total" — is purely qualitative and names no numerical crossing point, so there is no specific band the metric must enter under the null; any transient uptick can be dismissed as noise without a pre-registered bound. The natural-variability variance term is itself estimated from CMIP6 pre-industrial control runs, whose model-to-model spread produces a 2–3× range in internal-variability amplitude (roughly ±0.08–0.15 °C per decade in 30-year global-mean-temperature trend uncertainty), and a single strong ENSO or Pinatubo-class volcanic event can temporarily inflate the natural-variability share by 5–10 percentage points without constituting secular reversal. Without a hard numerical threshold tied to that underlying spread, the FALSIFIES arm is unreachable on any finite timeline, making the hypothesis operationally one-sided.

   Resolved

   Pre-register a quantitative FALSIFIES band — for example, natural-variability share exceeding 10 % of total variance sustained across ≥5 consecutive 30-year rolling windows — and verify reachability by drawing 500+ synthetic timeseries from the CMIP6 pre-industrial control ensemble to map the null distribution of the ratio metric; the FALSIFIES threshold must sit at or below the 10th percentile of that null distribution to be genuinely enterable under natural-variability dominance. Supplement with a sensitivity sweep across ERA5, MERRA-2, JRA-55, and NCEP-CFSR (whose anomaly spread of ~0.02–0.05 °C bounds instrument and transport-model noise), and compute bootstrapped 95 % CIs on the attribution ratio in each reanalysis member; if any member's lower CI bound crosses the pre-registered falsification level, the hypothesis is flagged as potentially falsified rather than requiring unanimous cross-product agreement, ensuring the test remains genuinely two-sided.

Unlike the static stress tests above (synthesised against the frozen catalogue spec), this is what a 3-voice council found in the most recent biweekly review. Refreshed on the 1st and 15th of each month at 09:00 UTC. Each voice runs one bounded web search via Anthropic's web_search_20260209 tool, cites what it finds, and recommends a verdict. The verdict diverges from the curated catalogue status (supported) — the synthesis below explains why.

Agent draft incorporating the 9 cited findings from the live council above. Not auto-merged — surfaces here for human review. To accept, open a PR editing site/src/_data/hypotheses.json with the revised fields below. To reject, ignore and the proposal will refresh on the next council run.

1. supported

   May 30, 2026 · added to catalogue at status "supported"

• US federal district courts adjudicating climate-damages suits that admit IPCC AR6 Detection & Attribution evidence at the >0.95 anthropogenic-fraction confidence threshold record defendant success on the causation element declining from a ~35% historical baseline to below 15% within 18 months across a corpus of 50 or more such filed cases.
• Munich Re and Swiss Re reinsurance rate-on-line for fossil-fuel sector property-casualty towers widens 40–80 bps within two annual January renewal cycles (by Q1 2027) as near-unity attribution certainty raises the legally recoverable subrogation fraction against named emitters.
• European Commission enforcement actions under EU Climate Law Article 8 (2021/1119) citing the >0.90 anthropogenic-fraction threshold tighten the EU ETS Phase IV annual linear reduction factor, lifting EUA spot prices €8–12/tCO2e above the pre-enforcement baseline trajectory by Q4 2027.
• Moody's Investors Service places 10–15 investment-grade oil and gas issuers on negative credit watch within 24 months, with climate-litigation tail-risk contributing a one-notch downgrade for issuers carrying proven-reserves book values above $50B at their next scheduled rating review.
• ExxonMobil, Shell, BP, Chevron, and TotalEnergies collectively disclose $8–15B in incremental climate-litigation reserves in aggregate across two consecutive annual reporting cycles as the causation-element threshold in US and EU jurisdictions converges on the >0.95 attribution standard.