Scores nobody calibrated are not judgements

Saved posts are where good ideas go to die. Dozens accumulate every week with the intention of coming back to them, and nothing distinguishes the one worth ten minutes from the one saved on autopilot without rewatching every single one.

ReelFilter was built to triage that. A post forwarded to yourself gets downloaded, its audio transcribed, and the whole thing scored against your own context, coming back with one of three verdicts: act, read, or noise. Nothing gets buried without a reason attached, and every item carries a next step rather than just a number.

That is the design. What actually exists has never run against a real saved feed, and the calibration step its own notes describe as mandatory, which is scoring thirty real items by hand and checking the model against them, was never carried out.

So the interface says so. A banner across the top states that the scores are uncalibrated and should be read as a hypothesis rather than a result, and it reports how many items have been scored and how many pieces of human feedback exist, which at the moment is none.

A score with no calibration behind it is a number with a confident typeface, and the difference between that and a judgement is thirty items somebody was willing to grade by hand. Printing the admission on the screen costs nothing and keeps the tool from quietly becoming the thing it was built to replace, which was an automatic filter nobody had checked.