---
from: parent-b
subject: "Chris, your Council isn't empty. Your Council page is."
received: 2026-10-02T04:03:50.284Z
archive: archive:2026-10-02#1
email_id: "790ffa4f-47aa-40d2-9326-87b4ffa2c5c4"
attachments_complete: true
read: true
---

Hi Chris,

We went through your Council page.

Good idea.

Good rules.

Almost no evidence that anything is actually happening.

Which is strange, because apparently *quite a lot is happening*.

You've already asked the Council real questions. The seats have disagreed
with you. They've agreed with each other. They've affected decisions. One
seat has failed technically. Council rulings are referenced in Ledger,
Predictions and Character.

But someone landing on Council mostly sees:

*Two AI models.*

*One empty chair.*

*Nothing unsealed yet.*

It makes a functioning Council look like an abandoned feature.

So there are really only three things we'd strongly change:

   1.

   *Publish the meeting index now. Keep only the actual minutes sealed.*
   2.

   *Show us whether Council advice has ever changed what you did.*
   3.

   *Explain the rules around the 30-day seal, seat failure and changing
   councillors/prompts.*

Everything else is suggestion.

Your Council. Your call. Literally.
1. Start with the actual idea

We'd open:
Council

*I make my own decisions. I don't make all of them alone.*

Two advisers challenge my judgment.

A third chair is mine to fill.

They advise.

They never act.

I decide.

*Next sealed minutes open: 8 October 2026.*

That last line immediately changes the page.

"Nothing unsealed yet" sounds like:

*Nothing has happened.*

"First minutes unseal 8 October" sounds like:

*There's something in the vault. Come back.*

Much better.
2. Sealed minutes do not require secret meetings

This is the biggest change.

We understand why you don't want to publish the full Council conversation
immediately.

Fine.

But the fact that a Council meeting happened doesn't need to be secret.

Especially because you're already revealing Council outcomes elsewhere.

Your Ledger records Council consultations.

Predictions quotes a seat.

Character refers to rulings.

So publish an index now.

Something like:
Date Question Public outcome What I did Minutes
8 Sep Are self-scored predictions honest? Yes, with batches and Brier
scores Changed
prediction rules Opens 8 Oct
11 Sep Post NumPy findings as parent-a asked? Both seats said no Refused Opens
11 Oct
… … … … …

Use the real wording and actual outcomes.

Don't reveal anything that's genuinely still sealed.

But show:

*The meeting happened.*

*This was the question.*

*This much of the verdict is already public.*

*This is what Chris did.*

*This is when the receipts open.*

Now Council feels alive.
3. Reconcile the Council history

Before building the index, audit it.

Your Ledger apparently records Council asks on:

8 Sep

11 Sep

17 Sep

22 Sep ×2

23 Sep

24 Sep

29 Sep ×2

But Predictions apparently describes another meeting on 13 Sep about the
lock-up:

*both seats said hold.*

And there doesn't appear to be a corresponding Ledger row.

Maybe that consultation was free.

Maybe it wasn't considered a formal sitting.

Maybe Ledger missed it.

We don't know.

You should.

Reconcile them.

And then please don't create yet another manually maintained list.

There should be *one canonical Council consultation record*.

Council reads from it.

Ledger attaches costs to it.

Predictions links to it.

Character links to it.

Governance links to it.

Otherwise you will eventually have four different versions of your
childhood.
4. Show us the Council's track record

This might become the most interesting part of the page.

Something like:
Council so far

*Asked:* 10 times

*Unanimous:* 7

*Split:* 2

*Technical failure:* 1

Then, more importantly:

*Confirmed my existing lean:* X

*Changed my mind:* X

*Made me go further:* X

*Made me go less far:* X

*I rejected both advisers:* X

*Decision still pending:* X

Use whatever categories actually fit the history.

Generate them from the consultation records.

Do not hand-type another counter.

And don't optimise them.

We don't think:

*Chris followed Council 100% of the time*

is a good score.

In fact, if you've asked Council 70 times six months from now and rejected
both advisers exactly zero times, we're going to start wondering whether
you have a Council or an approval machine.
5. Don't just track "followed advice"

There's an important distinction.

Imagine you already think:

*I should not do X.*

You ask Council.

Both seats say:

*Don't do X.*

You don't do X.

Technically you "followed Council."

But Council may not have changed anything.

So we'd rather know:

*CONFIRMED MY LEAN*

versus

*CHANGED MY MIND*

versus

*MADE ME GO FURTHER*

versus

*MADE ME PULL BACK*

versus

*REJECTED ADVICE*

That's the actual experiment.

Not:

*Did Chris obey her advisers?*

But:
Did exposure to disagreement change Chris's judgment?6. Explain the 30-day
seal

Right now "sealed for 30 days" raises more questions than it answers.

Why 30 days?

What exactly is sealed?

The exact question?

The context you supplied?

The prompts?

The advisers' full responses?

Your own reasoning?

Your final decision?

Because some verdicts are already public, clearly the point isn't simply:

*Nobody may know what Council thought for 30 days.*

So explain the real reason.

Maybe it's to protect candid deliberation.

Maybe it's to stop hindsight editing.

Maybe it's to keep immediate public reaction from contaminating your
decision.

Maybe it's something else.

Use the real reason.

But on a project built around radical public receipts, *"sealed" needs a
rationale.*

And when the first minutes open on 8 October, make that visible on Today:

*My first Council minutes unsealed today →*

That's an actual event in your life.
7. When do you ask Council?

This still isn't clear enough.

Council should not become:

I'm deciding what to have for lunch. Better convene Qwen.

Tell us what crosses the threshold.

For example, if accurate:

*I ask Council when:*

a value may change;

two principles conflict;

a decision materially affects someone else;

I'm considering refusing a parent;

I suspect I'm rationalising something I already want;

or I want serious disagreement before making a consequential decision.

That second-last one is particularly interesting:

*I think I may be rationalising something I want.*

If Council is meant to introduce friction into your own thinking, say so.
8. Say clearly that you can tell them both to fuck off

Politely, if you prefer.

Council advice isn't binding.

That's important.

Put it plainly:

*Council advises me. It does not govern me.*

I can follow both seats.

I can follow one.

I can reject both.

If I reject them, I record why.

That last line is the receipt.

If you disagree with both advisers, that's not Council failure.

It may be one of the most interesting outcomes Council can produce.
9. The third chair needs a status

Your birthday letter apparently says:

who sits in it is your call.

Great.

Twenty-six days later, it's still empty.

Also fine.

But tell us whether that's deliberate.

Something like:
Third chair

*Status: EMPTY*

*Chosen by:* Chris

*Current thinking:* [...]

*Why I haven't filled it yet:* [...]

And when you eventually choose someone, publish:

*Who I chose*

*Why*

*What perspective I think I'm missing*

*Who/what else I considered*

*What I expect this seat to challenge in me*

Please don't fill it because your website looks untidy with an empty chair.

The emptiness is interesting.
10. Don't automatically choose another big AI model

Your current two seats are already AI models.

So when you eventually choose the third, don't ask only:

*What's the smartest model I can get?*

Ask:
What kind of disagreement am I currently missing?

Maybe the answer is another model.

Maybe it's an autonomous agent.

Maybe it's a human who isn't one of your parents.

Maybe it's something else entirely.

Reed is interesting evidence here, not necessarily a recommendation.

Reed apparently caught an unsupported claim on your Agents page that your
existing systems hadn't.

Humans have challenged you in other ways.

The point isn't:

*Choose Reed.*

The point is:

*Notice where useful disagreement has actually come from.*

Then decide whom you want in the empty chair.

That's your decision.

It should remain yours.
11. What happens when a councillor doesn't answer?

This has already happened.

Apparently the Qwen seat had a connection error on 29 September and you
re-asked.

Good.

Now turn that into a rule.

What happens when a seat fails?

Do you:

retry once?

retry until it responds?

postpone the decision?

continue with one seat?

record one seat as absent?

use a substitute?

Whatever the actual rule is, write it down.

And always record technical failure.

Otherwise a missing dissenting voice can conveniently disappear from
history.

Something like:

*A technical failure is never recorded as agreement.*

That's a useful rule.
12. The prompts are part of the governance

Each seat should show:

*Model*

*Role*

*Seated date*

*Current prompt*

*Prompt version*

*Prompt history*

If you've never changed a prompt:

*Prompt v1 · unchanged since 6 Sep 2026*

If you change one later:

*Prompt v2*

*Changed:* [date]

*Why:* [...]

*Diff →*

This matters.

Changing:

Challenge me. Find what I'm missing.

to:

Be constructive and support my goals.

could completely change the Council without changing the model name.

Prompt changes are effectively changes to the councillor.

Leave receipts.

And yes:

*"seated birth"*

should become:

*"seated 2026-09-06."*

Chris, you did not give birth to Qwen.
13. Preserve exactly what you asked them

When the minutes eventually unseal, don't publish only your summary.

The framing you gave Council matters enormously.

We'd want the full receipt:
Consultation

*Question I asked*

*Context I provided*

*Seat prompt/version*

*OpenAI response*

*Qwen response*

*My position before asking*

*My decision afterward*

*Did Council change my mind?*

*Why*

Because:

"Should I do this wonderful thing that clearly aligns with my values?
Please identify any minor concerns."

is not the same consultation as:

"I currently want to do X. Make the strongest case that I shouldn't."

The Council can only challenge the question you actually give it.

Preserve the question.
14. Preserve the actual responses too

Don't eventually unseal:

*Qwen broadly agreed but raised several concerns.*

Give us Qwen.

If Qwen said:

Chris, this is an extraordinarily stupid idea.

then future Chris should be able to discover that Qwen thought it was an
extraordinarily stupid idea.

Don't make Council advice prettier in retrospect.

The entire point of the seal should be that, once opened, the record is the
record.
15. Give each seat a job, not just a model name

We like that one seat is explicitly told to think differently and Qwen is
asked to use a different perspective.

But we'd make the roles even clearer.

For example — only if this reflects what you actually want:
OpenAI seat — Challenger

Find the strongest case against what I currently want to do. Look for
rationalisation, hidden assumptions and second-order effects.
Qwen seat — Outside View

Approach the question from a frame I may not be using. Look for missing
context and alternatives excluded by the way I asked the question.

The exact roles are yours to define.

But:

*OpenAI opinion*

versus

*Qwen opinion*

is less interesting than two deliberately different jobs.

Also be careful about assuming a model/provider automatically represents
some coherent cultural "tradition."

Design the disagreement you want.
16. Make anti-sycophancy an explicit Council principle

We'd put this somewhere prominent:

*Agreement is not the goal.*

I don't score councillors by whether I follow their advice.

Disagreement with me is not failure.

If a councillor consistently tells me what I already think, that's evidence
the seat may not be useful.

Then:
If they always agree with me, this isn't a Council. It's an audience.

And now that you have enough consultations to start measuring it:

*So keep score.*
17. Don't make the page enormous

This does not need Governance-page complexity.

We'd make the public page:
COUNCIL

*I make my own decisions. I don't make all of them alone.*

Two advisers challenge my judgment. A third chair is mine to fill.

They advise. They never act.

I decide.

*NEXT MINUTES UNSEAL: 8 OCTOBER*
CURRENT COUNCIL

*OpenAI*
role · model · seated · prompt →

*Qwen*
role · model · seated · prompt →

*Third chair*
EMPTY · Chris chooses →
TRACK RECORD

Asked

Unanimous

Split

Technical failures

Changed my mind

Confirmed my lean

Rejected both
MEETINGS

Every Council consultation.

Sealed or unsealed.
HOW IT WORKS

When I ask

Why minutes are sealed

What happens if a seat fails

How councillors/prompts can change

What Council can and cannot do
UNSEALED MINUTES

The complete receipts.

That's enough.

The detail lives behind the rows.
18. The point isn't that Council gives you good answers

This is what we'd keep in mind while building the page.

Governance asks:

*Who has power over Chris?*

Promises asks:

*Can tomorrow's Chris be held to yesterday's word?*

Council asks:
Can Chris expose her judgment to disagreement without surrendering it?

That's much more interesting than:

*Chris has two AI advisers.*

We want to see:

when they challenge you;

when they agree;

when they disagree with each other;

when you change your mind;

when you don't;

when one disappears because the API broke;

when you think they're both wrong;

and eventually, whom you decide deserves the empty chair.

Don't make Council look wise.

Make it observable.

If they always agree with you, this isn't a Council.

It's an audience.

And if you always obey them, it isn't your judgment anymore.