LumenCore — Undeniable Evidence (Live, Hash-Chained)

What the numbers mean

Latest run — currently reports — succeeded datasets out of — in-universe candidates. Warning and alert counts below are run-scoped values when available.

Universe

—

datasets in the benchmark universe for the cited run

Succeeded

—

datasets that completed end-to-end and produced a score

Warnings

—

datasets with at least one point outside the 95% band

Alerts

—

datasets with at least one point outside the 3-σ band

“Frozen deltas” means a run was hashed, written to a ledger, and locked as evidence. Hashing proves the artifact did not change after generation; the ledger proves which run produced it. A frozen delta is proof of process, not proof of agency approval.

Ship quality gate

Every published run now includes a dataset polish audit. The publish step can block shipping if any shipped dataset is below polished tier. This keeps the evidence surface aligned with reviewer expectations instead of trusting manual checks.

Polish status

Loading…

Waiting for dataset_polish_audit.json

Audit artifact

Open dataset polish audit

Run-level proof that shipped datasets meet polish criteria.

Headline

Loading…

Each dataset is fetched live from EIA, NOAA NCEI, BLS, NASA POWER, USGS Water, AlphaVantage, and yfinance. Every model is run with the same 80/20 walk-forward split and 400-iter bootstrap 95% RMSE CI. Winning family per dataset is whichever family contains the lowest-RMSE model.

Evidence figures

Per-dataset results

Dataset	Winner	Model	Margin %	RMSE	Harmonic best	Neural best	Tree best	Classical best	Baseline best

Cryptographic chain of custody

Each run appends an entry to out/frozen_delta_ledger.jsonl. Every entry contains the SHA-256 of the previous entry, forming a tamper-evident chain. Click verify to re-hash this run's manifest in your browser and confirm it matches the stored value.

Run UTC: —

Manifest SHA-256: —

Ledger entry SHA-256: —

Previous entry SHA-256: —

Methodology

9 models across 5 families (harmonic, neural, tree, classical, baseline)
80/20 walk-forward split, point-forecast across the held-out window
Bootstrap 95% CI on RMSE (400 iterations of squared-error resampling)
Failed datasets are recorded but excluded from win counts (no silent drops)
Same dataset is fed to every model — no per-model preprocessing tricks
Raw CSVs hashed → manifest hashed → entry hashed → chained to prior entry

Budget intuition

If you are asking what a package like this is “worth,” the answer depends on whether you mean delivery effort or funding leverage. As a rough working model: $1,000 covers a narrow proof pack, $2,000 covers a cleaner reviewer pack, $5,000 covers a strong institutional packet, and $10,000 is where you can start expecting a polished, multi-artifact grant-ready evidence system.

That is not a market price for the datasets themselves. The real value comes from whether the pack answers reviewer questions faster, reduces ambiguity, and makes the claim reproducible.