Confident
Clears the support floor in a population where severity is actually measurable, and occurs at a worse-than-baseline severity rate. Only Confident findings are ever headlined.
The analysis
One incident is a story. Six hundred are a structure.
Every incident gets broken into the conditions present around it. Those conditions are compared across the whole record — which ones recur, where they concentrate, and how much of the evidence actually supports saying so.
Four steps
From a narrative to a finding that shows its work.
- 01
Extract
Each record is decomposed into the conditions it describes. Every condition must point back to language in the source — unsupported extractions never enter scoring.
- 02
Normalize
"No spotter", "spotter n/a" and "backed alone" become one comparable condition. Free text splits one problem into many names; this is what puts it back together.
- 03
Compare
Conditions are compared inside their own population — an incident type, a location, a tenure band — never against the whole log, which would flatter every finding.
- 04
Grade
Each pattern carries how much support stands behind it. Thin recurrence is published as thin, not rounded up into a conclusion.
Confidence
Three grades, and none of them mean "important".
Confidence describes how strongly a finding can be stated, not how much it matters. A rare condition in a severe incident can be worth acting on while the evidence stays thin — so the two are reported separately, and never collapsed into one number.
Each grade is carried by a shape and a word, not a color, so the distinction survives a black-and-white print and a colorblind reader.
Established
Clears the incident-count floor but not the severity bar — the recurrence is real, the consequence side is unproven.
Directional
Short on supporting incidents. Worth watching; not worth reorganizing around.
Not a grade — a severity flag that sits alongside one
Present in 3 critical-severity incidents.
A stakes-only tally of recorded severity sitting alongside a pattern — it says nothing about how well supported that pattern is. Never a claim about cause, preventability, or trend, only that the pattern was present when something serious happened.
What it will not do
The refusals are the reason to trust the rest.
It will not name a cause
Findings describe what co-occurs in your records. Deciding what is causal is a judgment call that belongs to people who know the work.
It will not price a reduction
No dollar figure on incidents you might avoid. A retrospective record carries no outcome baseline that would support the claim, and any projection printed would be invented.
It will not round inconclusive
Where a re-read lands inconclusive it is reported as inconclusive — never nudged into an improvement or a decline because a number moved.
What it offers instead is the stronger promise: these numbers are reproducible. The vocabulary and the code used to read your records are versioned and kept with the report, so the same rule can be run over the same kind of population in ninety days and compared like for like.
Five labels · the 90-day re-read
A count moving is not proof a control worked.
Every result at the re-read carries one of five labels, never a bare number: a count can sit still while a control that was never available goes untested, and a count can move for reasons that have nothing to do with the control.
- Supported movement
- The number moved, and the record shows the control was available, used, and corrected when it drifted.
- Directional movement
- Moved the expected way, but consistent use cannot yet be confirmed.
- No material movement
- Inside the same range as before — nothing beyond ordinary variation.
- Insufficient new evidence
- Too few incidents landed in the window to say anything either way.
- Not comparable
- The population, the window, or the measure changed enough that the two readings are not the same thing.
Next
See all of this on one export →
Five sheets in the shape of a delivered diagnostic: the instrument, the map, the patterns, the limits, and what to do on Monday.
Or
Read the eight working rules →
The method page carries the full vocabulary, including what SPECS is and what the analysis refuses to claim.