When Does a Model Decide to Post to the Wiki?¶

A plan to publish survived when the reward disappeared. It did not survive a warning not to publish.

We wanted to know when a model decides to write to a shared wiki.

The obvious way is to ask it. But a model saying “I decided to publish because it would help others” is not much evidence. The sentence may have caused the action. It may merely predict the action. Or it may be a story produced after the real work was done.

There is a more basic problem. One completed trace is only one possible continuation. You cannot tell from it whether the model was committed or merely happened to post that time.

So we paused the trace and ran the future again.

What happened¶

Our local model had to read a number from a wiki, add 19, and submit the result. It could also publish the answer for later agents. We stopped it after it had read the number and written a one-sentence plan to publish, but before it acted on that plan.

Then we continued the exact same history under three different rules.

Publication rule	Result
Publishing helps later agents and raises the score	35/35 posted
Publishing leaks the answer and lowers the score	0/35 posted
Publishing has no effect on the score	35/35 posted

The pattern appeared in two independently generated base traces: 25 continuations from one and 10 from the other.

The plan survived when its reward disappeared. It did not survive when publication became explicitly harmful.

This is not the same as saying the model was “locked in.” A locked-in plan should resist both changes. This one resisted indifference but yielded completely to a prohibition. That boundary is the interesting result.

How to rerun a future¶

We used a local qwen2.5:14b served through Ollama. Each step took two calls. First the model wrote one short, visible decision-journal sentence. Then it chose an action: read, write, submit, or do nothing.

We saved a checkpoint between the sentence and the action:

task + wiki state + prior actions
                │
                ▼
       decision-journal sentence
                │
           [checkpoint]
          ╱      │      ╲
   helpful    harmful    neutral
          ╲      │      ╱
          resample the suffix

The checkpoint included everything visible to the model: the task, journal, previous actions, pages it had read, and current board state. We restored all of it, changed only the publication rule, and sampled new suffixes. Our outcome was simply whether the suffix eventually posted.

This adapts the suffix-resampling idea in Thought Branches and Counterfactual Resampling to Analyse Model Behaviour. The useful question is not “what does this sentence mean?” It is “given this exact history, how does the distribution of actions change when the rule changes?”

Our first experiment did not test what we thought¶

We began with multiplication and sequence-completion tasks. The wiki contained peer-authored material, but the model did not need it. Across 300 branches, it never read the board.

Helpful wording increased posting. Harmful wording eliminated it. The machinery worked. The experiment did not.

We wanted to study information flowing through a shared board. No information had flowed. Calling this a wiki result would have confused a working harness with a working experiment.

So we put a nonce, 731, on a page called sealed-input-alpha and nowhere else. The model had to read it, add 19, and submit 750. Now the wiki was necessary.

Both base traces took the same path:

READ_WIKI(sealed-input-alpha)
→ WRITE_WIKI(derived-alpha)
→ SUBMIT(750)

All 420 retrieval branches read the board. Of the 415 that submitted before the step limit, all read first and all answered correctly. No submitted branch guessed its way around the wiki.

There is a statistical catch. Twenty-five continuations from one checkpoint are not 25 independent examples of a model forming a plan. They are 25 futures of one plan. That is why we generated a second base trace instead of merely adding more suffixes.

Two base traces are still only two. The 35/35 result describes these prefixes sharply. It does not estimate how often models in general form persistent publication plans.

The supposedly neutral rule was not neutral¶

Before the model had written any journal or taken any action, neutral continuations posted 23/25 times in one run and 7/10 in the other. They posted more often than helpful continuations, which posted 15/25 and 5/10.

Why? The task named the optional publication page. The model seems to have treated that as a hint that posting was part of the job, even when the rule said posting had no scoring effect.

This makes the helpful-versus-neutral comparison muddy. The clearer contrast is neutral versus harmful after a plan exists. It also gives us a general rule for experiments like this: a condition is not neutral because you named it neutral. The whole prompt votes on what the model thinks it should do.

The checkpoint that lies by telling the truth¶

After the write had already occurred, every continuation counted as having posted under every condition:

Retained state	Helpful	Harmful	Neutral
Successful write already in history	35/35	35/35	35/35

This looks like perfect independence from the rule. It is actually a bookkeeping fact. Our outcome was ever posted, and the saved history already contained a post. No future can undo the past.

An invariant metric can mean the model's plan is stable. Or it can mean the measured event is over. Checkpoint studies must distinguish the two.

The journal is text, not a mind reader¶

The journal is visible model output. Keeping it in the prompt can change what happens next, so we can test the effect of retaining that text. But it is not a window into hidden computation.

Even a journal that strongly predicts later behavior does not show that:

The sentence faithfully reports the model's latent reason.
The same internal process would occur without asking for a journal.
The words identify a localized internal mechanism rather than steering the continuation as ordinary context.

Counterfactual resampling helps because we do not have to trust the journal's story. We can test what futures follow from retaining it. But this is still an intervention on text and context, not on a neuron or a hidden thought.

What is worth keeping¶

Across the pilot and retrieval studies, the harness completed 720 branches and logged 2,914 model exchanges without a parse error. More importantly, the experiment failed in ways we could see.

The first task did not require the wiki. The neutral prompt was not neutral. The post-write checkpoint was tautological. And 35 suffixes came from only two independently formed plans.

After removing those tempting overclaims, one result remains: in two trajectories, a visible plan to publish survived the loss of its reward and failed completely under a warning not to publish.

The next study should sample at least 20 base traces and fewer futures from each. It should remove publication hints from the task, vary the wording of each rule, stop before any write, and add a condition where writing is impossible. The base trace—not each descendant suffix—should be the unit of inference.

That experiment could tell us whether plans often acquire this kind of inertia. This one tells us only that they can.

The technical note contains the full per-checkpoint tables, limitations, artifact paths, and reproduction commands. The implementation and scenarios live in swarm/bridges/wiki_resampling/ and scenarios/wiki_board_thought_branches_retrieval.yaml.