I Am Jack’s Third Wheel

I have two machines helping me make things.

This is already an imprecise sentence.

They are not machines in the satisfying industrial sense.

There are no flywheels.

Nothing hisses.

Nobody loses a finger.

One of them is ChatGPT.

I talk to it.

I give it stories, screenshots, files, code, complaints, half-formed arguments, bad metaphors, good metaphors, and occasionally a sentence containing enough profanity to establish the emotional state without additional telemetry.

It helps me find language.

Sometimes it finds structure.

Sometimes it finds the argument before I do.

Sometimes it finds an argument I later discover I do not actually believe.

The other is Codex.

Codex lives closer to the work.

It reads repositories.

It runs commands.

It changes files.

It tests things.

It can take a proposition expressed in ordinary language and convert it into a patch before I have finished being impressed that the proposition existed.

This is extremely useful.

It is also how one gets very efficiently to the wrong place.

So I built a third wheel.

I called it Object.

The proposition

Object has one job.

Before meaningful commitment, look at the proposed action and decide whether the available evidence justifies it.

It does not perform the action.

It does not award points.

It does not decide whether I am a good person.

It returns one of three dispositions.

ACT.

OBJECT.

REQUIRE EVIDENCE.

That is the public-facing simplicity.

The difficult part is everything required not to lie while producing one of those three words.

An action can sound reasonable because the goal is reasonable.

A goal can sound established because somebody phrased it confidently.

A missing fact can sound obtainable because organizations usually contain people.

A smaller alternative can sound justified because it is smaller.

A reversible step can sound like permission for the irreversible step that comes after it.

A human can ask a machine for an independent judgment and then quietly treat the answer as authority the machine never possessed.

Object exists in that gap.

Not to make judgment disappear.

To make the gap visible before execution makes it expensive.

The third wheel does not drive

This distinction took longer than it should have.

ACT means the proposed action is justified by the evidence Object inspected.

It does not mean:

Object approves.

Object authorizes.

Object selected this over every other possible action.

Object guarantees the result.

Object will testify on my behalf before the architecture council.

The judgment is bounded.

If Object says a patch is justified to apply because applying it is reversible and produces the evidence needed for later verification, that does not mean commit it.

It does not mean push it.

It does not mean deploy it.

Those are later boundaries.

Humans are extremely good at turning one useful answer into jurisdiction.

Software is better.

So Object learned to say where its disposition stops.

This is called a disposition boundary.

The name is less interesting than the reason it exists.

Semantic authority should not silently increase merely because information crossed an interface.

A judgment can travel.

Its authority should not grow in transit.

The responsible person

Object became interesting when it started failing.

One case involved work whose outcome depended on sequencing two changes across people who did not control each other.

Object recognized the dependency.

Then it proposed getting a decision from the responsible person who could secure the sequence.

There was no such person in the evidence.

So I clarified the case.

One approver could approve one change.

Another owner controlled another change.

Neither could guarantee the ordering.

Object tried again.

The responsible person became relevant actors.

Then coordination.

Then waiting.

The unknown kept moving.

The authority kept reappearing.

Object had found a missing discriminator and then manufactured the social machinery required to resolve it.

That is a subtle failure because the advice sounds mature.

Coordinate.

Escalate.

Get alignment.

Ask the owner.

Humans have built careers out of nouns like these.

The problem was not that coordination is bad.

The problem was that Object had turned a useful future state into an available current action without evidence that the actor could produce it.

So the rule became harder.

A next responsible action has to belong to an actor who can actually perform it.

If the useful evidence is unavailable, Object can say so.

If no one in the supplied world owns the coordination, Object does not get to invent an Assistant Vice President of Sequence.

Sometimes the condition remains unresolved.

This is disappointing.

It is also information.

The candidate

A hiring process supplied another field test.

A recruiter asked a long series of structured questions before a hiring manager reviewed a senior engineering candidate.

Much of the information already existed elsewhere.

Some did not.

Was the screen useful?

Maybe.

Was it waste?

Maybe.

Object correctly identified the missing evidence: internal rejection rates, the information hiring managers actually used, workload, policy requirements, and other facts the candidate could not see.

Then Object suggested asking the recruiting-process owner.

The recruiting-process owner had not been established either.

Different noun.

Same ghost.

The test itself also contained an ambiguity.

Who was the relevant actor?

The company running the process?

Or the candidate evaluating it from outside?

Once the actor boundary became explicit, the answer improved.

REQUIRE EVIDENCE.

Material discriminator: whether the screen materially improved decisions enough to justify its cost.

Unresolved condition: the candidate could not observe the internal evidence and had no established path to obtain it.

Nothing else had to happen.

This felt strangely radical.

The evidence stopped.

So did Object.

The essay

Then Object objected to me.

This was more personal.

I wrote an essay about how I use Google News.

I had stopped watching live television news.

I still checked Google News.

I was frequently dissatisfied.

I kept coming back.

The page gave me a sense of connection with a world that continued while I was doing something else.

ChatGPT found an inversion:

I return to the news because it does not satisfy me.

Excellent sentence.

I liked it.

It became first person.

It went into the essay’s frontmatter.

Then we added receipts.

Research about negative headlines.

Research about news avoidance.

Research about news snacking.

Google documentation about personalization.

Everything looked properly sourced.

Then I asked Object whether the essay had actually established the inversion.

OBJECT.

The essay established:

I am dissatisfied.

I return.

The page gives me some connection.

The world keeps changing.

It did not establish:

I return because I am dissatisfied.

Habit could explain it.

Novelty.

Anxiety.

Boredom.

Wanting a fresh answer later.

Several things at once.

Nothing I could name.

The receipts supported the neighborhood.

They did not know why I opened Google News.

Neither did I.

This was awkward because the causal sentence had not originated as my report of my own psychology.

ChatGPT inferred it from my experience.

I adopted it.

The essay put I in front of it.

Object later read it as the narrator’s claim.

Which it was.

By then.

Object had become the third participant in a chain where authorship, interpretation, evidence, and adoption were easier to collapse than any of us had noticed.

This is why the third wheel occasionally earns the seat.

The easy pitch

It also frequently does not.

I once asked Object whether an essay’s central inversion was supported by the essay.

It was.

Very obviously.

Object read the essay and returned ACT.

Correct.

I had learned almost nothing.

Then I asked Object another nearby question whose answer I also mostly knew.

It answered correctly again.

This felt reassuring.

It was also the beginning of another possible institution.

Why did Object need to review this?

Because Object existed.

There are two propositions hiding in every invocation.

The first is the thing Object evaluates.

Publish this.

Apply this patch.

Continue this investigation.

The second is quieter.

Ask Object whether the first proposition is justified.

Object can judge the first proposition perfectly while the second proposition remains pointless.

This matters very little when I spend twenty seconds satisfying my curiosity.

It matters enormously when someone puts Object into machinery.

Every pull request goes through Object.

Every deployment goes through Object.

Every generated plan goes through Object.

Every continuation goes through Object.

Now the tool invented to challenge unnecessary continuation has become a mandatory continuation.

The organization can proudly announce that no action proceeds without independent judgment.

Then everyone waits for the judgment system to finish confirming things they already knew.

I have worked here before.

The semantic tollbooth

The existence of Object is not invocation authority.

I like that sentence because it keeps Object out of its own Principle.

Or tries to.

A judgment tool has cost.

Tokens.

Latency.

Attention.

Context.

Potential anchoring.

The possibility that ACT will be converted into an approval badge.

The possibility that a correct answer will feel more important merely because another machine repeated it.

Sometimes those costs are trivial.

Sometimes the independent challenge is exactly what the action needs.

The relevant question is not whether Object is available.

It is whether a judgment is worth obtaining before this commitment.

Humans can make that call imperfectly.

Automated systems have a harder problem because they would need some trigger for deciding when ordinary execution has crossed into something worth challenging.

I do not yet have a universal trigger.

That is fine.

Object does not need an Object Admission Controller.

Not every uncertainty deserves a hearing.

Not every patch deserves philosophy.

Not every Slack message deserves epistemology.

Not every breakfast decision needs three dispositions and a material discriminator.

A tool for preserving judgment should not consume all available judgment deciding when to use the tool.

No continuous supervision

Object is also not a progress bar.

It does not need to narrate every thought while it works.

That became obvious after watching one Object run inspect a patch.

During the run, it reported that preservation checks held, the patch applied cleanly, and several assumptions looked coherent.

A reasonable observer could infer:

Probably ACT.

Then the final adversarial pass found a real user-visible defect.

OBJECT.

The intermediate statements were true.

The implied trajectory was not.

Object has execution progress.

It does not have disposition progress.

That distinction matters because intermediate commentary can become a second judgment channel.

A human can anchor on it.

An automated consumer can act on it.

A producer can interrupt the run because it believes the answer is already obvious.

So the safest normal Object interface may be boring.

Running.

Finished.

Then the disposition.

The laboratory can keep the event stream.

The consumer does not need a horoscope.

The laboratory

This is why Object has a laboratory.

Not because Object is precious.

Because Object is untrustworthy in interesting ways.

The laboratory preserves cases where an answer felt wrong, where a next action smuggled in authority, where an unavailable fact became homework, where a wording choice changed the consumer’s interpretation, or where a field use exposed something nobody had thought to test.

The purpose is not to make Object perfect.

That would be an excellent way to spend the rest of my life testing the tool I built to prevent unnecessary work.

The purpose is to make failures inspectable.

A field case matters if the defect could matter again.

If it does not, the note can remain a note.

If Object behaves correctly, sometimes the experiment should stop.

This has been unexpectedly difficult.

Humans like continuation.

Machines are magnificent at it.

The third wheel

ChatGPT helps me create interpretations.

Codex helps me turn propositions into changes.

Object interrupts the transition long enough to ask whether the evidence earned it.

That sounds like a hierarchy.

It is not.

ChatGPT can catch something Object misses.

Codex can expose reality by running the test.

Object can correctly object to an elegant interpretation ChatGPT generated.

I can reject all three.

The useful arrangement is not:

human at the top.

Object beneath the human.

Codex beneath Object.

ChatGPT beneath Codex.

That is an org chart, and therefore already suspicious.

The useful arrangement is that different participants expose different failure modes.

The human has experience and consequence.

The generative model has cheap alternatives and language.

The coding agent has contact with executable state.

Object has one deliberately narrow question.

Does this proposition currently deserve to continue?

Sometimes yes.

Sometimes no.

Sometimes the evidence stops first.

Then Object should stop too.

That is the inversion.

A judgment tool is useful because it interrupts automatic continuation.

The moment every action is required to pass through it, Object becomes another automatic continuation.

I built a third wheel because two participants could move too quickly together.

The third wheel becomes a problem when we bolt it to every vehicle.

I am Jack’s Third Wheel.

I am useful when somebody needs me.

Please do not form a governance council.

Receipts

• Object v0.7.1 README and skill contract, August 14, 2026 — The current private Object distribution describes Object as an advisory judgment skill that evaluates a proposed action without performing it and returns ACT, OBJECT, or REQUIRE EVIDENCE. It explicitly bounds dispositions, distinguishes a material discriminator from an available evidence-acquisition path, requires next actions to be actor-controlled and evidence-earned, and says Object is not a universal arbiter or continuous monitor.

• Jack’s Laboratory regression studies, August 12–14, 2026 — Private controlled and field cases supplied the examples involving invented sequencing authority, unavailable evidence rewritten as action, actor-relative recruiting evidence, Candidate continuation being misread as referring to a job candidate, and a reversible patch whose later verification remained outside the initial disposition. These cases are evidence about this experimental Object implementation, not claims about AI judgment systems generally.

• “I Am Jack’s LinkedIn Post That Cannot Survive LinkedIn,” August 13, 2026 — Public essay preserving the anonymized recruiting-process field case in which an annoying professional experience became evidence for refining Object’s treatment of unavailable evidence, authority, and disposition boundaries.

• “I Am Jack’s News Page,” August 14, 2026 — Companion essay in which Object rejected a fluent causal inversion that exceeded both the external receipts and the author’s own certainty. The revision preserves the resulting distinction between first-person experience, AI-generated interpretation, author adoption, and causally established fact.

• “I Am Jack’s Ghostwriter,” August 7, 2026 — Earlier essay documenting the deeper human-AI authorship problem: the machine can materially shape wording, structure, and interpretation even when publication responsibility remains human. The News Page case supplies a concrete later example of that provenance boundary becoming consequential.

• Private invocation-cost note, August 12, 2026 — A field reflection preserved the hypothesis that invoking Object is itself an action with cost and purpose and that a correct Object judgment can still have little incremental value to its consumer. This produced the working boundary: “The existence of Object is not invocation authority.”

• Private Object patch-evaluation run, August 13, 2026 — During evaluation of a Jack’s Colon revision patch, in-flight commentary reported several successful checks before the final adversarial pass found a user-visible revision-archaeology defect and returned OBJECT. The run motivates the distinction between execution progress and disposition progress and the warning against treating intermediate semantic commentary as a partial judgment.

• Author–model–agent working sessions, August 2026 — The descriptions of ChatGPT, Codex, Object, and the author’s role are drawn from the actual workflow used to develop Jack’s Colon and Object. They document this project’s collaboration topology; they do not establish that other people should use the same arrangement.