Plumb
Reading stable · switch to edge
How-to guide

Tracking a parameter

For anyone carrying constants nobody has validated.

A parameter is a decision with a model and a measurement. Everything about recording one — the condition, informs, aspects, what the board will not do — is on Recording a decision, and none of it is repeated here. Set it up there first. This page is only what differs.

A threshold, a window, a limit — min_oi: 100, a 20-day lookback. Where you can state a model connecting the value to something measurable, and you have the measurement, it derives exactly as a decision does and reads differently:

id,lexicon,reduce,aspects
min_oi,parameter,worst,attribution|beats-control|holds-out-of-sample
word what happened
supported the measurement backs the value
unvalidated nothing has tested it — the ordinary case for most constants
refuted it was tested and lost
unmeasured cited, and no measurement ran

Aspect names are yours. The core never reads them, so name them after what your evidence actually distinguishes.

Is it a parameter, or just a number?

One question decides it: is there a model — something that would validate this number?

Not whether you have measured it yet. The model is the half you can write down today, and writing it down is what forces you to say what would confirm the value. The measurement is the half that arrives later, and it only says how far along the parameter is: unvalidated before anything tests it, unmeasured once something cites it and no measurement has run, supported or refuted after. Gating on the measurement would send a value you fully intend to tune off to be recorded as a decision, where the vocabulary does not fit a number and the intent is lost.

Is it a parameter, or just a number? One question decides it: is there a model — something that would validate this number? Without a model it is not a parameter, whether or not you can measure anything: with no measurement it is a decision with a number in it, and with a measurement it is observable but not tunable. With a model it is a parameter already, and the measurement only says where in its life it sits — unvalidated or unmeasured before one has run, supported or refuted after. a value you chose IS THERE A MODEL? something that would validate this number no yes MEASUREMENT? MEASURED YET? no yes not yet yes not a parameter a decision with a number in it — record it as one not a parameter observable, not tunable — a chart says nothing to do PARAMETER unvalidated, then unmeasured once a test cites it PARAMETER supported, or refuted — tunable, error derived the model decides. the measurement only says how far along it is most values that look like parameters have no model — nothing would ever confirm them