Choosing perspectives

Choosing perspectives

The test for a perspective set is not "are these real job titles". It is would these two disagree. If not, one of them is redundant and you are paying twice for one read.

What makes a perspective produce different output

Three things, in decreasing order of effect:

1. What they are exposed to. A compliance reviewer notices retention and deletion because they will be asked to attest to them. A support engineer notices what a customer will quote back at them. Exposure drives attention more than expertise does.
2. Whether they are subject to the thing. A reader who would be recorded by a system reads it differently from one who would read it. This perspective is the one most often missing, and it produces findings the others structurally cannot.
3. Time horizon. Someone answering for this in three years asks different questions from someone shipping on Friday.

Job title alone is the weakest of these signals and the one people reach for first.

A set that disagrees

For most artefacts, four to six perspectives is enough, chosen so their interests conflict:

Perspective	Positioned to see
The inheritor	Why is this like this, and what did we already rule out
The dissenter	Where this is used against the person it records
The operator	What happens at 2am when it misbehaves
The attester	What can actually be evidenced under scrutiny
The outsider	What is unintelligible to someone who was not here

The dissenter is not optional. Without a perspective that stands to lose from the artefact, a fan-out is a hype loop with citations — every reader wants it to work, so every reader finds ways it could.

Dials

Modifiers change where a perspective looks, never what it treats as true. That distinction is what keeps them safe: aim divergence at the search space, not at the evidence rules.

Dial	Range	Changes
Divergence	conventional → out-of-the-box	How far from the obvious case they hunt
Rigour	practitioner → academic	Lived friction versus named general classes
Charity	adversarial → generous	Whether premises are granted or attacked
Horizon	immediate / seasonal / multi-year	Today's friction versus the long-lived artefact
Disposition	burned / neutral / invested	What feels salient right now

Honest status: unvalidated. The one measurement of these held a perspective constant and varied four dials at once, at n=3. It cannot attribute any effect to any single dial, and it credits the perspective with nothing since the perspective never varied. Treat the dials as advisory vocabulary for shaping a prompt, not as a mechanism with demonstrated effects.

Two tensions follow from the contract rather than from measurement. Both are predictions — worth acting on because the countermeasures cost nothing, and worth labelling because neither has been tested:

• Divergence should fight specificity. The genuine edges of a search space are the cases with no artefact to name — so a high-divergence reader is steered toward material it must then label constructed, which gets discounted. The stable strategy becomes mid-divergence content with divergent adjectives: the appearance of exploration without the substance. This is an argument from the incentives the contract creates, not an observed outcome.
• Incoherent combinations should produce noise. An attester at maximum charity, or a marketer at minimum, is a reader who cannot hold their own position. Refuse the combination rather than obeying it — cheap insurance against an untested failure, not a documented one.

Fan-out discipline

If you run several:

• Do not cross-product. Perspectives times settings times dials explodes. Assign one combination per reader and spread them so no dial level is confounded with a perspective.
• Keep them blind to each other. Independence is what makes agreement mean anything.
• Vary something. Duplicates give several draws from one distribution at full price.
• Build in a control. Hold one variable constant across two readers. If their output is near-identical, the variable is doing nothing, and you learn that for the cost of nothing extra.
• Stage the first batch. Run one, check the grounding is real, then commit to the rest. One round trip against a whole wasted fan-out.

Reading the results

Not by counting votes.

Pattern	Reading
Agreement across opposed interests	The strongest signal available — but check whether the agreed frame came from the briefing first
A finding from one perspective only	Often the sharpest. The one reader positioned to see it was the only one who could. Do not discount by count
Anti-scenarios clustering	Where the artefact is genuinely weak
A perspective reporting "this adds nothing for me"	A scope signal, and the one result shape a briefing alone cannot produce
A checkable structural claim	Verify it directly. This is where the value concentrates

Verify before relaying. A fan-out produces leads. Claims that a section contradicts another, that a term appears nowhere, that a function cannot do what it promises — those are checkable in seconds, and checking them is the difference between a review and a pile of opinions.