When Knowing More Makes You Less Qualified

Working note: this is exploratory. I am trying to name a pattern I can feel before I am ready to call it a framework.

I have spent most of my career treating separation of duties and job rotation as obviously sensible controls. They are. But lately they have started to look like low-resolution versions of something much richer.

The change began with a very ordinary experience: working with people and AI systems while moving among roles.

Sometimes I am deep in implementation and functionally wearing the same hat as the coding system beside me. For three iterations we may both be Codys: inspecting behavior, changing code, rerunning tests, narrowing a defect. Then I reach the edge of what I can responsibly judge. I stop being useful as Cody. I step outward and become Art, trying to understand the architecture. Sometimes I am out of my depth there too, and I need another perspective to help me wear Judy, asking what is actually in bounds. Later I may become Shep and make a product decision that none of those other roles had authority to make.

The person did not change.

The role did.

And with the role, something changed about what information was appropriate, what authority was legitimate, what evidence mattered, and even what ignorance was useful.

That last part has been bothering me in a productive way.

Sometimes knowing more makes me more qualified.

Sometimes knowing more makes me less qualified.

More is not always more

I am used to thinking about knowledge, visibility, and authority as things that should generally increase together.

If I am expected to make a larger decision, give me more information.

If I am responsible for an outcome, give me more authority.

If I am auditing a system, give me more visibility.

That logic is often correct. It is also incomplete.

There are judgments for which additional knowledge improves the role, and judgments for which additional knowledge contaminates it.

A developer who knows every compromise, workaround, intention, and internal justification may be the best person to explain how a system works. That same developer may be a poor independent evaluator of whether the interface is actually understandable to someone encountering it fresh.

A product steward may need broad authority to decide whether a feature belongs in the product. That does not mean the steward should see every privileged implementation detail before evaluating the lived experience. Sometimes the clean question is precisely: does this make sense from here?

A participant may have exquisite visibility into frustration and confusion while having almost no authority to redesign the system.

An auditor may need unusually broad visibility while having deliberately narrow authority to intervene.

So perhaps these are not one ladder:

more knowledge
    -> more visibility
        -> more authority

Perhaps they are separate dimensions whose useful combination depends on the role.

Quantum Invariants gives me a useful set of anchors here. P1 treats boundaries and interfaces as structural rather than accidental. P3 asks who authorized a crossing and for what scope. P4 insists that governance depends on relevant legibility, not maximal disclosure. P9 scales governance with power and consequence. P10 asks what comparator is actually being used to call something better, worse, sufficient, or unacceptable.

Taken together, those do not tell me how to build a role system. They do suggest why “give everyone everything” is not a neutral default.

Too little knowledge, too much knowledge

I can now see at least two symmetric failures.

The familiar one is:

I do not know enough to exercise the authority I have been given.

The less familiar one is:

I know too much, or I know it from the wrong position, to provide the independence this judgment requires.

That second condition is not necessarily corruption, conflict of interest, or bad faith. It can be an ordinary consequence of participation.

If I have just spent six hours building the mechanism, I have acquired a dense internal model of why it behaves as it does. That model is valuable. It may also make it harder for me to experience the mechanism as a stranger would.

If an AI participant has been shown privileged world state during debugging, it may no longer be a clean participant for a later test, even if nobody tells it to use the information. The epistemic position changed.

If a human reviewer has been immersed in the implementation argument, the review question may quietly shift from “does this work?” to “does this behave the way I now understand the implementation intends?”

The actor may still be competent.

The role may no longer be clean.

I do not yet have a precise vocabulary for this. “Conflict of interest” is too narrow. “Bias” is too general. “Context contamination” is closer, but still sounds pathological when the process may be perfectly healthy.

Maybe the simplest statement is enough:

Role fitness depends not only on what an actor knows, but on how that knowledge was acquired and what judgment the actor is now being asked to make.

Separation of epistemic positions

This gives me a different view of separation of duties.

Traditional separation of duties asks whether one person should be allowed to perform multiple consequential actions. That matters, especially where fraud, irreversible changes, or concentrated power are possible.

But there is another separation worth preserving:

separation of epistemic positions.

Two judgments may need different relationships to the same system even when the same trusted person is capable of making both.

The point is not that one human must forever be the Developer and another must forever be the Reviewer. That can turn into bureaucracy very quickly. The point is that the stances are meaningfully different.

Sometimes the same person can move between them with no problem.

Sometimes a transition should increase authority.

Sometimes it should reduce authority.

Sometimes the next role needs more context.

Sometimes it needs less.

Sometimes the right move is not to add information but to preserve ignorance.

This is where C13, Coercive Transparency, becomes unexpectedly relevant to me. The composite warns that complete transparency can collapse boundaries and move agency toward whoever can see, store, score, or act on what becomes visible. I usually think about that in privacy and surveillance terms. But there is a milder internal version too: maximal visibility can erase the very distinctions a role needs in order to function.

A witness who merges completely with the thing witnessed may gain information while losing witness.

Roles are not identities

The practical relief for me is that none of this requires people to become fixed organizational boxes.

I do not want to be Architect.

I want to be able to occupy Art when the question is architectural.

I do not want someone else to permanently be Rules.

I want Judy available when a boundary question needs a different stance.

I may be Cody in the morning, Art after lunch, and Shep ten minutes later.

That sounds sloppy if roles are identities.

It sounds natural if roles are temporary operating positions.

A crude model might look something like:

actor
+ current question
+ current authority
+ available evidence
+ prior exposure
= operative role fitness

I do not mean that as an equation. It is a reminder that the same actor can be differently suited to different judgments at different moments.

This also changes how I think about authority.

Authority does not attach automatically to humanity, seniority, intelligence, or technical competence.

It attaches to standing within the current decision.

A human beta tester may be entirely qualified to say, “I could not tell what that button did,” and entirely unqualified to accept a release milestone.

An AI playing the participant role may be qualified to report, “I repeatedly failed to rendezvous with Bob,” and entirely unqualified to expose Bob’s hidden coordinates to itself.

A coding system may be highly qualified to implement an authorized change and entirely unqualified to authorize the next scope expansion.

The role is doing real work.

Friction is not a feature request

This became clearer to me through a small game-design example.

Imagine a participant, Alice, saying:

It helps when I can see Bob’s location.

At the lowest resolution, that is simply a finding about Alice’s experience.

It does not become:

Expose player locations.

A local review might conclude that not knowing Bob’s location is part of the game. The friction is intentional. Keep going.

But suppose the same class of difficulty appears across twenty sessions. At a higher resolution, a different question becomes available:

How do ordinary human players solve this problem without privileged information?

They might share coordinates in chat. They might describe landmarks. They might build signs, maps, trails, towers, or meeting places.

Now the synthesis is no longer “remove uncertainty.”

It might be:

Future machine participants need an ordinary participant-equivalent way to communicate their own coordinates or landmarks.

The local complaint was real.

The local prescription would have been wrong.

The higher-resolution pattern may still justify a tool.

This is where C10, Level Mismatch, helps me keep my footing. A locally useful optimization can damage the larger system if the boundaries and comparators differ across levels. Reducing one participant’s friction is not automatically the same thing as improving the game.

Sometimes friction is a defect.

Sometimes friction is missing tooling.

Sometimes friction is missing instruction.

Sometimes friction is evidence that a boundary is working.

And sometimes the right answer is simply:

Yes. That is the experience.

Inter-role learning

This is where job rotation begins to look more interesting to me.

The usual story is that rotating people through jobs reduces dependency, improves coverage, and spreads knowledge.

All true.

But perhaps role rotation also creates perspective transfer.

A Cody who has spent time as Peggy may implement differently because they remember what the interface felt like from the outside.

An Art who has spent time as Judy may become better at designing boundaries that remain legible under review.

A Shep who has spent time as participant may make product decisions differently because “user friction” is no longer an abstract category.

A Judy who understands implementation may distinguish a genuine boundary violation from an unavoidable technical seam.

The value is not that the roles merge.

The value is that experience crosses roles while the distinctions remain available.

That sounds very close to feedback and recursion. P6 treats learning and stabilization as recursive processes rather than one-time decisions. A system of roles could learn not only within each role but between them.

The most interesting findings may then be relational:

Peggy keeps needing information that Art assumes is obvious.

Judy keeps receiving questions that could have been prevented by a more legible interface.

Cody repeatedly solves around an architectural ambiguity that Shep has never explicitly resolved.

Shep keeps making the same class of product judgment because the product intent has not been compressed clearly enough.

Now the object of improvement is not only the participant, code, or policy.

It is the interface between roles.

Periodic synthesis instead of reflexive reaction

This suggests a development metabolism I want to keep exploring:

experience
-> situated finding
-> retain
-> review at the appropriate resolution
-> synthesize
-> decide whether to preserve, teach, expose, tool, constrain, or redesign
-> fresh experience

The resolution matters.

In the moment, the finding may only justify a note.

Across a scenario, it may justify a skill improvement.

Across many scenarios, it may reveal a missing affordance.

Across the whole system, it may reveal a mistaken assumption about what participants need to know.

The same evidence can support different questions at different scales without forcing an immediate answer.

That feels healthier than converting every complaint into backlog.

It also feels healthier than dismissing complaints because they are “part of the game.”

Both responses compress too early.

What I do not want

I can already see the danger in this line of thinking.

It would be very easy to turn a useful insight into a role bureaucracy.

Every decision could acquire a matrix.

Every transition could require a form.

Every hat could become a department.

Every question could trigger a ceremony.

That would destroy much of what I like about the idea.

The roles are useful precisely because they are lightweight.

They help answer:

What question am I actually asking right now?

What authority belongs with that question?

What information should this role have?

What prior exposure might make this stance less clean?

Do I need more context, less context, or another role?

Sometimes the answer may take ten seconds.

Sometimes the same person will occupy three roles in ten minutes.

Sometimes no explicit transition needs to be recorded at all.

The point is not procedural purity.

The point is preserving distinctions that matter.

A few questions I want to keep open

I do not know yet how far this goes.

When should prior exposure disqualify a reviewer, and when does it merely enrich them?

Can an AI system meaningfully “shed context” for a new role, or does clean independence require a genuinely fresh session or model instance?

How much of role history should be retained as provenance?

Can a role inherit learning from prior occupants without inheriting privileged knowledge that would compromise the new task?

When does rotation improve judgment, and when does it create diffusion of responsibility?

Could authority itself be dynamic—expanded or contracted as evidence, impact, or epistemic position changes?

How do we preserve the benefits of asymmetric visibility without turning opacity into unaccountable power?

And perhaps the question underneath all of them:

How do we build systems in which learning can move freely between roles without collapsing the boundaries that make the roles useful?

I am not ready to answer that.

I am increasingly convinced it is worth asking.

There is something beautiful in the possibility that mature coordination is not a matter of giving the smartest actor the most information and the most authority.

It may be a matter of putting the right actor, with the right history, behind the right boundary, with enough visibility to make this judgment—and enough restraint not to make the next one.