标签：#calibration

TL;DR In generative video pipelines, running cheap low-step sketches to pick parameters sounds like free optimization. But when prompts go out-of-distribution, surrogate scorers return noise, turning a \$0.002 check into a bad decision that triggers a \$15 compounding failure. Here's why skipping the cheap step is sometimes the cheapest option. Three numbers run Scenematic's generation loop. A think-frame costs \$0.002. A full render costs \$0.50. A bad scene that slips through and gets built on costs about \$15.50, because the scene chain compounds it before anyone looks. The constant in lib/generation-loop.ts carries the arithmetic in a comment: 15.502, // CALIBRATION_TARGET: 0.002 + 0.50 + 15.00 . Most of the pipeline exists to keep spend at the cheap end of that ladder. One module decides when the cheap step should be skipped entirely. A hundred-contract baseline then put numbers on how often that decision was wrong. 1. The rehearsal lib/think-frames.ts generates quick, low-inference-step sketches before committing to a full-quality keyframe. The file header credits DeepGen's think tokens as the inspiration. Each sketch tries a different preservation focus, character, environment, mood, composition, or atmosphere, with its own image-to-image strength and seed. The reward mixer scores the batch and the winner's parameters go to the full render. The economics only work if those scores mean something. That assumption fails quietly, and it fails hardest on the prompts where a rehearsal looks most useful. 2. Where the scores stop meaning anything Scoring a sketch of A detective leans forward across a metal table, interrogating a nervous suspect under fluorescent lights works fine. The scorer has seen a thousand shots like it. Scoring A sentient equation writes itself across a blackboard that extends infinitely in all dimensions does not fail loudly. It returns a number, and the number is noise. Both prompts are verbatim from the baseline harness. lib/ood-detector.ts

The last post was about making a number trustworthy. Leakage geometry, purge widths, de-overlap, a baseline that could not cheat. It ended with a minute-scale ceiling that held at 52% across seven configurations and a model family swap. This one is about what happens after you trust the number. Because a probability you are going to bet on is a different object from a probability you are going to report. The probabilities are not decorative The path-passage classifier is a three-class LightGBM. It returns p_up , p_down , p_none . Those go straight into the expected-value score that decides whether to take a trade and how big: long_score = p_up * ( B - C ) + p_down * ( - B - C ) + p_none * ( - C ) short_score = p_up * ( - B - C ) + p_down * ( B - C ) + p_none * ( - C ) B is the barrier, C the cost. Read the arithmetic. Every term is linear in a probability. Scale p_up by 1.2 and you scale the long score by very nearly 1.2. So miscalibration does not stay in the model. It becomes a bet-sizing error, in proportion, in the bins where the gate actually fires. A classifier that is right 70% of the time while claiming 90% is not 20 points wrong. It is sizing every position in that bin as though the edge were far larger than it is. Boosted trees are known for uncalibrated softmax output. I had been consuming it as if it were a probability. The audit Seven live assets. For each one, fit an Inductive Venn-Abers wrapper on the time-ordered older 80% of that model's training data, 6,988 rows, and evaluate against a 500-row uniform-random sample of the newer 20%, seed 42. The LightGBM models are reloaded from disk and left alone. Only the wrapper is fit. Measure Expected Calibration Error and log-loss, before and after. Asset ECE before → after ECE Δ Log-loss Δ BTC 0.1272 → 0.0621 -51.2% -5.5% ETH 0.1795 → 0.0298 -83.4% -11.5% SOL 0.1680 → 0.0386 -77.0% -10.6% XRP 0.2219 → 0.0645 -70.9% -17.7% ADA 0.1419 → 0.0369 -74.0% -8.0% LINK 0.1260 → 0.0737 -41.5% -1.2% LTC 0.1508 → 0.0603