Store the law, discard the data
For a law-governed signal, keep only the generator (trend + dominant periodicity) and a residual certificate, and regenerate on demand. The certificate is issued only if its bound COVERS held-out data — and it refuses signals that aren’t law-governed. Honest lossy-in-bytes, bounded-in-information.
- 1,261×
- reduction on a certified trend+periodic signal
- holdout-tested
- certificate must cover held-out data before issuance
- refuses noise
- a random walk fails the gate, a cert that can't fail is no cert
- bounded
- reconstruction lies within the q99.9 residual bound
at1 ruletier store signal.csv -o signal.at1rule # CERTIFIED (explained_var=0.992, holdout_coverage=0.999, # q99.9 bound=2.67); 1261.8x reduction at1 ruletier regenerate signal.at1rule -o out.csv # within the bound # a random walk? -> REFUSED to certify (not law-governed)
Keep the law, drop the readings it implies
Some signals are governed by physics rather than by events. A cooling curve, a decay, a pressure response following a known law: the readings are not independent facts, they are consequences of a small number of parameters. Storing every sample stores the same law thousands of times over.
Where the law holds, keeping the parameters and regenerating the series is enormously smaller. The risk is obvious and it is the reason most tools do not offer this: if the law stops holding partway through and nobody notices, you have replaced your data with a confident fabrication.
The holdout is the whole product
So part of the series is withheld from the fit and used to test it afterwards. If the regenerated values do not match the held out readings, the law is rejected and the data is stored normally. The generator is only kept when it has already predicted values it was never shown, and a file that falls back to ordinary storage is the system working rather than failing.