ExpMan — Insertion Evidence

Several systems each know a little about when this sample went in. We take what each one says, weight it by how well it knows, and combine them. The more sources agree, the higher the confidence we publish. Everything on this screen is computed from records we already have, so there is nothing here to fill in.

Established insertion

2026‑07‑10 08:14:32.6

Rotary specimen rack (RSR) · σ_t ±0.92 s

Combined uncertainty

σ_t = 0.92 s · reduced χ² = 0.65 · ν = 3

each source weighted by 1/σ², after correcting for its known lag

Class
Confirmed eligible for model training

Reading the three numbers. σ_i is each source’s standard error: the camera is good to about 1.5 s, a person tapping a button to about 6 s. Inverse-variance weighting gives each source a vote of 1/σ², so the camera outvotes the kiosk tap 16 to 1. σ_t = 0.92 s is the standard error of that weighted mean, which is always tighter than any single source. Reduced χ² is the goodness-of-fit check, χ² per degree of freedom, with ν = 3 here (four estimates, one fitted mean). Around 1 means the scatter matches the quoted σ; 0.65 means slightly tighter than expected; above 3 we stop and ask a person.

Time: a number, so sources can be combined

Source	t_i	σ_i	bias b_i	corrected	residual	weight	Standing
Intercom tapanalog line, digitised + transcribed	08:14:32.1	1.2 s	0	08:14:32.1	−0.52 s	0.59	Machine time
Camerain-core view, machine witness	08:14:33.6	1.5 s	0	08:14:33.6	+0.98 s	0.38	Observed
Experimenter tapkiosk, human-mediated	08:14:24.0	6 s	−3.0 s	08:14:27.0	−5.62 s	0.02	Asserted
AcknowledgementRO ↔ exp, human-mediated	08:14:40.5	8 s	+2.5 s	08:14:38.0	+5.38 s	0.01	Agreed
Calendar approvalwindow 08:00–13:30	–	5716 swindow/√12	–	–	consistent	check only	Intent
Ex-core detectorfuture, would rank above the camera	–	–	–	–	–	–	Not installed

What the intercom tap is good for. The timestamp comes off the audio sample clock, so there is no human lag to correct for. What gets said is usually “okay, it’s in”, which pins the time and leaves the sample and the facility to other sources. It carries most of the weight here and none of it in the location panel.

Two human sources are not two independent sources. The tap and the acknowledgement both come from people talking, and people run late in the same direction. A machine source agreeing with a human source counts for more than two human sources agreeing, and the weights are set from that.

The console log for this run reads 08:12:00. An operator typed it when the sequence started, and it is the number the old workflow would have kept. The independent sources put the sample in the facility 2 minutes 33 seconds later. That gap is the whole reason for this screen.

Location: one right answer

A facility is a name, so there is nothing to average. We take the answer from the highest-ranked source and ask the others to match it.

ex-core detector > camera > acknowledgement > experimenter claim > schedule

Camera	RSR	authoritative
Acknowledgement	RSR	agrees
Schedule	RSR	prior only

All three say RSR. If one said something else we would leave the location open and flag it.

Mass: the scale and the barcode

Two questions, kept apart: how good is the scale, and is this the same object we weighed.

Measured: 14.802 g
Instrument σ: ±0.5 mg balance cal 2026-06-02
Identity link: Unbroken barcode on canister, scanned at weigh and at insert

A precise weight on the wrong canister beats nothing and loses to a rough weight on the right one, so we publish both numbers.

Why this is the golden input

The ROM learns how core state maps to flux, and every training example has to say what was in the core and when. An insertion time off by a few minutes teaches the model a wrong pairing at full confidence. An insertion nobody recorded turns that run into a clean baseline, and baselines are what the rest get measured against.

We can filter on confidence. Every event carries a class, so the training set can be selected on it. Only confirmed events go into training. Asserted and intent-only events still drive the live views, where a person sees the label and corrects it.

When sources disagree

We downgrade the event, keep both numbers, and name the two sources that disagreed.

Worked example. The camera reads 08:14:53.0 while the kiosk tap still reads 08:14:24.0. Over the same four sources, after lag correction, the reduced χ² goes from 0.65 to 41 against a threshold of 3, so we refuse to combine them: averaging sources that disagree produces a confident wrong answer, and nothing downstream can catch it. The event drops to Asserted, we keep the residuals, and it comes up on the day-of check while someone still remembers which one was right.

Marking a facility empty

With no insertion record, a facility could have been empty, or something could have gone in unwritten. Baselines are what the other runs get compared against, so we want a facility marked empty because sources say so. The camera and the scheduled-slot state give us that, and the facility gets marked Known empty on the strength of them.