Evidence

We try to catch ourselves lying.

Evidence, not adjectives. Every number below renders from metrics.json and carries its ratification status.

The yardstick

KID-HP
Does our fine texture look statistically drawn from real aerial photography? The primary realism gate.
hf_ratio / hf_frac
Is there the right amount of fine detail — not too soft, and no fake grain? A floor guards blur; a 1.5× ceiling guards faked texture.
Radial PSD distance
Is detail present in the right amounts at every scale?
LPIPS
Perceptual distance to a ground truth photographed on a different date — recorded, but demoted from decisions for exactly that reason.
DISTS
A second perceptual opinion, more tolerant of texture resampling than LPIPS.
2AFC
Can a human tell ours from real? 0.5 = coin flip = the goal. The actual success criterion.
Round-trip
Every output is shrunk back to 10 m and compared with the source pass — the anti-hallucination check.

Current gate table — 2.5 m plane

ModelKID-HP ↓ DISTS ↓LPIPS ↓ PSD ↓P16Status
Bicubic ×4, radiometry-normalized 18 297 0.342 0.72 1.594 0.9 Control
SEAN v5 @ ~21k steps 4 685 0.15 0.236 0.1 1.01 Proxy
SEAN v5 3 262 0.145 0.225 0.1 1 Ratified
SEAN + high-pass-masked spectral fine-tune λ=100 2 878 0.146 0.224 0.122 1.01 Pending sign-off
PiSA-SR zero-shot 7 777 0.238 0.426 0.337 0.98 External
OSEDiff zero-shot 9 302 0.243 0.465 0.425 0.98 External
SEN2SRLite 19 117 0.371 0.736 1.433 1.1 External
LDSR-S2 19 175 0.343 0.586 1.033 1.02 External
DiffFuSR 22 757 0.296 0.447 0.316 0.99 External

Real-input arms: LPIPS/PSD are content-confounded (different-date ground truth) — KID-HP and DISTS are the meaningful columns there. The 0.5 m research track uses a different fixture and scale and is deliberately not shown beside these numbers.

Texture realism per version

Against the public frontier (same real inputs)

The credible externals lose by ≥1.7×; the S2-native public models sit at the bicubic floor.

Can you tell which is real?

Twenty pairs. One patch is real aerial photography; one is generated from a Sentinel-2 pass. Founder benchmark (session 1, 2026-07-06): accuracy 0.90 — the goal is 0.50, a coin flip.

1 / 20

Your running accuracy — dashed line = 0.50 goal

no answers yet

Reports

Report PDFs (baseline table, metrology, validity annex) are pending the founder's public-release selection — nothing is published here until that call is made. The numbers above are the same ones those reports pin.

The ones we lost

Corrections & caveats we publish on ourselves (5)
  • The LLM judge is UNVALIDATED (repeat-agreement failed both prompt versions) — its numbers are recorded facts, never decision inputs.
  • Two historical hf-fraction definitions exist; only the pinned def-A values with a stated tile set are citable.
  • The honest bicubic floor is LPIPS ≈ 0.720, not the 0.841 recorded before radiometric normalization.
  • A 2026-06-21 claim that real-input output HF was identical to synthetic input was withdrawn — the deficit is ~15% and was masked by rounding.
  • Indistinguishability is a goal, not an achievement: the founder still tells our output from real aerial 90% of the time.
  • LLM-judge repeat agreement: 0.562 (threshold 0.8) — UNVALIDATED — recorded facts, not decision inputs.