01
Parameter recovery fails
Verification · Validation · Falsification · Uncertainty
Kill or validate: every claim needs a way to fail.
BenchEWS separates publication, mathematical specification, correct software computation, and empirical validation. None of these levels substitutes for another.
PUBLISHED THEORY ≠ EMPIRICAL VALIDATION ≠ IMPLEMENTED SOFTWARE
01 / Evidence Path
No signal without a named measurement chain.
- 01System and research question
- 02Boundary and observables
- 03Data and provenance
- 04Model and estimator
- 05Null and competing models
- 06Preregistered criteria
- 07Reproducible run
- 08Evidence, uncertainty, and abstention
02 / Published Framework
BenchEWS Individual v0.2: ten hard falsification criteria
The published theoretical framework specifies ten explicit conditions under which its temporal-lead hypothesis should be considered unsupported or falsified. It reports no empirical data and is not a validated diagnostic instrument.
02A / Programme-level falsifiability
The positioning paper states how the architecture must fail.
The published v1.0 study narrows the BenchEWS claim to a non-scalar O/F/D/R/V profile and specifies conditions that would weaken, replace, or reject it. It is a positioning, formalization, and falsifiability study—not an empirical validation of a working diagnostic instrument.
Read and cite the positioning study ↗03 / Temporal Validity
Estimator latency is part of the claim
Window length, smoothing, sampling, and detection delay can create an apparent temporal lead. Observed lead and estimator-latency-corrected lead must therefore be tested separately.
02
Null or competing models explain the data equally well or better
03
Temporal lead disappears after latency correction
04
The result is not robust to plausible windows and preprocessing
05
The persistence measure changes direction or meaning across regimes
06
Signal–response coupling is not identifiable
07
Perturbation recovery is not reproducible
08
Out-of-sample performance does not exceed the preregistered null
09
The result depends on undisclosed settings
10
Independent replication contradicts the central relationship
04 / Decision States
Bounded outcomes instead of artificial certainty
- VERIFIEDComputation agrees with its specification and reference cases.
- SUPPORTED IN CASEA hypothesis passes named preregistered tests in a bounded case.
- FALSIFIED / UNSUPPORTEDA hard criterion fails or a competing model performs better.
- INDETERMINATEData, measurement quality, or identifiability are insufficient.
- NOT TESTEDMissing evaluation must not be interpreted as indirect success.
05 / Self-Correction Validation
Earlier only counts when it is also reliable.
Evaluation must cover discrimination, calibration, lead time, incremental information, robustness, cross-domain transfer, and reproducibility.
Δtlead = tbaseline − tBenchEWSΔtlead > 0
A positive lead is not an improvement if false positives, instability, or estimator bias increase.
Inspect falsification and recovery criteria →06 / Software Boundary
Theory, implementation, and evidence remain separate.
Studio 1.4.0 is released. Studio 2.0 remains functionally separate and in release preparation. BenchEWS Individual, Quality-Driven Propagation, and Adaptive Reopening belong to the planned Studio 3.0 research horizon as published theory—not retroactively to Studio 2.0.