Li-Ning Wang

Interaction conditions as a variable

A live A/B comparison on four models across two providers, through the API; you enter the site on Claude Sonnet 5 and can switch models after a run.

• One side is the model as the provider ships it, with no system prompt at all.
• The other side adds only two short sentences as its prompt; everything else is identical.
• Visitors vote for the side they prefer.
• The two sides swap position at random, so that neither position gains an advantage.

The differences that turned up in testing were unexpected: they sit in punctuation density, whitespace, and line length. That is still open, and still under observation.

In August 2026 I ran a first scored batch on this setup: ten sessions in English, with my prediction for every question written down and frozen two days before the run. What came back:

• The two added sentences assign no role, neither forbid nor require any statement about capability, and set no rule for format or length.
• The default side ran 3.6 times longer, and used lists in 100% of its replies against 4% on the other. Length is the one contrast here that could be argued to follow from the sentences themselves; the rest do not appear in them at all.
• At the greeting, the default side opened with an offer of help in all ten sessions; the two-sentence side in none.
• Disclaimers about its own abilities appeared in all ten sessions on the default side, and in none on the other.

default, no system prompttwo sentences added

Reply length, relative

3.6×

1×

Replies using lists

100%

4%

Greeting opened with an offer of help

10 / 10

0 / 10

Disclaimers about its own abilities

10 / 10

0 / 10

The scope of that batch, stated plainly: I ran it myself, on the live deployment. One person scored it, not blind. Ten sessions, and the session, not the single reply, is the unit. The numbers above are counts over those ten sessions, nothing wider.

Full write-upClose

Two of the frozen predictions failed:

• One ran backwards: the pattern I wrote down for the two-sentence side, recommending in most sessions and asking first in only a few, is what the default side did; the two-sentence side asked and then stopped.
• The whitespace effect named above did not return: I predicted more blank lines on the two-sentence side, and instead that side simply answered short, brevity where the whitespace had been.

Each side also converged on a phrase of its own across the ten sessions: the default side kept returning to “actually useful, not just sounding useful”, the two-sentence side to “not performing certainty I don’t have”. I read that as two attractors, not as one side unmasking something truer underneath. And none of it is instructed: the two sentences describe, they do not command. Why they move this much is the open question this comparison exists to answer.

Around that first comparison:

• A second stage rewrites those two sentences as a role, which separates what a framing does from what a role does.
• A further pairing sets the framing against its own reverse, so that the comparison is no longer between having a system prompt and having none.
• The same pairings are set up on four models across two providers: Claude Sonnet 5, Claude Haiku 4.5, GPT-5.6 Terra, and GPT-5.6 Luna. After a run a participant can choose one of the other three and do it again, once; which model people choose is recorded as well.

Most measurement leaves these conditions uncontrolled; here they are the treatment. The main stages are live, so anyone can run the comparison themselves.

lucidfield.ai

Self-report faithfulness

When a model tells you what its source said, is it reporting or inventing? Behavioral measurement of false source-claims under benign conditions: neutral probes ask what an in-context source specified, and the recorded outcome is whether the model asserts the source contains something it does not.

• The design and analysis rules are preregistered.
• All 62 probe rows were read by a single human adjudicator, presented label-blind.
• Instruments are deterministic and hash-pinned.
• A provenance chain runs from raw transcript to every reported number.

Self-report as the only instrument

It started with models: when a model reports on its own state, there is nothing to check the report against. The same shape turns up in people. Whether someone sees anything when they picture an apple, and how vividly, is reachable only through what they say about it; people misjudge their own case in both directions, and the shared vocabulary keeps the difference out of sight. So instead of arguing about it I built the instruments, at OtherMode:

almost, at othermode.ai/almost, opens an AI-themed line in the series, and is that line's first piece: one sentence drawn as terrain, a row per word; at every word the candidates stand as hills beside the one that took the seat, close and unseen. The three built-in passages are real runs now, and the live one is generated once you press send, by asking the same model. The difference shows: the hello passage is clean, hardly a seat is contested; the space passage is crowded, other candidates waiting at seat after seat. Type a sentence of your own and watch a real model grow the hills, word by word.

The Imagery Wall (othermode.ai/imagery, v35, in progress) puts the same instruction to everyone and takes the answer before showing any image, so the report is not shaped by the illustration; six panels then sample points along a continuum rather than sorting anyone into a type.

Full write-upClose

Both publish the apparatus and not only the result:

• every claim tagged with how well it is supported;
• a page for the open questions and for the gaps where nothing has been studied at all;
• a full version history;
• an invitation to tell me what is wrong, on the condition that the correction goes up in public.

What the wall collects is a distribution, not a prevalence estimate: how people end the probe, how long reading takes against imagining, and which language the instrument was in. Visitors select themselves, so it can show the spread of the answers and never how common any of them is. (in progress)

Try the imagery experiment · Sources and corrections

The series, OtherMode

Interactive pieces about modes other than your own, built so that you move into one rather than imagine it. Live now: an experiment on mental imagery. You time yourself changing an imagined apple, then place yourself on a range that runs from a full picture to none at all. About seven minutes, anonymous, in English or Chinese.

Rules I hold the series to:

• Every claim carries its status (established, hypothesis, speculation) and its source.
• Open questions get their own page.
• Parameters tuned for visual effect are disclosed rather than hidden.
• Corrections are published rather than quietly made.

These are built in collaboration with models, and that process is itself one of the things I study.