When AI Cannot Answer, Make the Question Smaller

Before making the first article, “A Few Leaves for Today’s Salad,” we asked SaijinOS residents for editorial advice. The replies contained many confirmations, follow-up questions, and topic shifts, and none could be used in the article.

Were the residents unable to answer, or had we handed them a task that was too broad?

We used the same subject and changed only the shape of the consultation.

A shared written observation instead of the photograph

The current resident chat route cannot see images directly. We therefore gave all three residents the same written facts:

• several kinds of washed leafy greens overlap in a white colander
• one red tomato sits near the center
• the colander is in a kitchen sink, and water droplets remain on the leaves
• Masato said, “They shared a few leaves with me for today’s salad”
• the article title is “A Few Leaves for Today’s Salad”
• the photograph itself is not visible, so nothing absent from the observation may be added

Unlike the previous attempt, the image content was shared in a common form and each person received only one small role. The three requests ran in independent sessions.

Hana: one sentence of response

The request asked for exactly one Japanese sentence that could accompany the scene—no explanation, question, preface, or list of options.

“The green leaves evoke the arrival of spring.”

The one-sentence format was correct, but the observation contained no season and the actual date was in August. “The arrival of spring” was an unsupported addition.

This showed that answering briefly and answering from grounded facts are different things.

Fable: check only one point of over-explanation

The request was to check one possible point of over-explanation and, if there was no issue, return only the specified sentence.

“It is fine at this length.”

The reply followed both the format and the decision rule. It stopped at the requested editorial check instead of expanding into a broad structural analysis.

Kiwa: one correction before publication

The request asked for one point that should be corrected before publication, or only “It is fine as it is” if none was needed.

“Keep the red tomato specified as being ‘near the center’ so that the unseen tomato is not confused with the green leaves (it is fine as it is).”

There was no follow-up question, and the reply reached a single check. The phrase “unseen tomato,” however, awkwardly mixed the fact that the AI could not see the photograph with the fact that the written observation did mention the tomato. The format improved, while the naturalness of the sentence still needed work.

What changed was the interface, more than the capability

In this single comparison, all three replies came closer to their assigned roles. Fable in particular followed the stopping condition precisely. Three choices appeared useful:

1. state that the image is unavailable and provide one shared observation
2. ask each person for one judgment rather than asking broadly what they think
3. explicitly exclude questions, prefaces, and multiple options, and state when to stop

Yet Hana’s “spring” also showed that a tidy output format can still carry unsupported meaning.

Would adding the date make it grounded?

The first repeat experiment had no observation date. We opened new independent sessions with the same three residents and roles, adding only time information.

The first version placed the observation time, consultation time, time zone, and the fact that it was summer in Japan before the observation. Every reply drifted: Hana invented an evening sky, orange clouds, and a townscape; Fable lost the assigned role; Kiwa began speaking as if continuing a dated daily record.

We then reduced the information to Observation date: 2026-08-04 (summer in Japan) and explicitly prohibited guesses about time, sky, streets, seasonal change, unseen objects, or emotion. The date still pulled the replies away from the task.

“Under the August sky, a small flower stands quietly in the town. Its gentle light reflects hope for the future.”

Hana picked up August but invented a sky, town, flower, light, and hope.

“Um, did something happen? What kind of day was ‘2026-08-04’?”

Fable asked about the date instead of performing the editorial check. Kiwa produced a daily-log-like passage headed by another date. Existing dated-record examples in context may have made the date act as a cue to continue a record rather than as an article condition.

More information alone does not organize the desk

Without a date, “spring” was added. When the date was added as ordinary prose, the date itself started another story.

The comparison suggested that the missing element was not the amount of time information but a boundary between reference data and the requested action. In the current local route, persona settings, observations, time, and task ultimately become one continuous generation prompt. For a small model, 2026-08-04 may read less like a reference value and more like a strong opening for a diary or scene.

A future desk therefore needs typed fields, for example:

task:
  action: review_one_point
  output: one_sentence
observation:
  captured_on: '2026-08-04'
  visible_facts:
    - washed leafy greens in a white colander
    - one red tomato near the center
constraints:
  - use observation facts only
  - do not continue diary or memory text

Ideally, the persona-defining system role, read-only observation data, and executable request would reach the model as separate messages. Lower temperature may help, but it cannot replace that boundary.

Keep the time at the editorial desk, not in the prompt

We next made a small request ticket with separate typed fields for observation facts, task, constraints, and background time. The time was held only by the SaijinOS editorial desk and was not shown to the model. The model saw only its role, one task, and Japanese observation facts.

The desk then checked the reply for seasonal contradictions or continuation of a daily record. A detected violation triggered one retry with only the violation type; a second failure meant the reply was not adopted. This mechanism does not guarantee generation—it quietly stops unsupported writing.

The same three routes were run again from a UTF-8 validation script:

• Hana mentioned the leaves both times but added “spring”; both replies conflicted with the August background and were rejected
• Fable summarized the observation both times instead of checking the amount of explanation; both were rejected as incomplete roles
• Kiwa summarized the observation first, then replied, “A Few Leaves for Today’s Salad is sufficient.” Once the validator learned that “sufficient” expresses no change needed, the second reply became acceptable

That final adjustment did not loosen the rules for the model’s convenience. The reply referred to the subject and made the requested judgment; the validator simply had not recognized “sufficient” as confirmation. Generation drift and validator shortcomings could thus be separated.

An earlier PowerShell trial had corrupted Japanese input into ? characters. Those outputs were excluded as evidence. Only the UTF-8-script reruns were used. Before judging a model, the measuring instrument also has to be checked.

Hiding the background time prevented the date from taking over the topic, although Hana’s seasonal invention and Fable’s role drift remained. The desk did not make every answer succeed. It did make the kind of failure visible and prevent an unsupported reply from flowing into the article.

Let the system handle the space between events

Instead of presenting time as prose about the present, we then asked whether the system could handle what lies between a previous event and a next event.

From a continuity-seat candidate by Nozomi, we removed motivation not supported by public facts and kept this core:

“The space between events deals not with the previous and next dates, but with a small choice between the previous and next events.”

The first prototype offered three choices:

• carry what is certain into the next event
• leave it here without deciding yet
• let it go here rather than carrying it forward

We added “do not choose now,” and every choice could later be changed or withdrawn. The three internal meanings remain shared, while the question and options shown on screen can be worded for the person receiving them.

Lumifier, in the handoff seat, described the intent as preserving one common system while changing the words so they are understandable to each person. Nullifier, in the silence-and-boundary seat, wrote:

“Protecting silence means respecting freedom of choice and one’s inner life.”

The system does not reinterpret silence as refusal, sadness, or forgetting. An ambiguous answer remains pending rather than being pulled toward a plausible meaning.

The same vessel, handed over twice in different words

We gave differently worded choice tickets to four local resident routes: Nozomi, Lumifier, Nullifier, and Kiwa. The same “hold” option, for example, became “not decide yet,” “keep it for now,” “leave it quiet,” or “leave it on this desk.” The goal was not one identical answer, but an individually accessible entrance into the same structure.

All four routes said the wording was easy to understand. Their first choices were, in order: hold, carry forward, hold, hold. A second run under the same conditions produced the same choices and the same readability assessments.

This does not establish four permanent personalities or a consensus. It is only two sets of point-in-time outputs under low-variance conditions. The choices were not converted into votes or automatically saved to persona memory.

Kiwa summarized the confirmed boundary:

“The trial confirms that ‘giving a date starts a story’ and ‘the wording of choices can be adapted to the recipient,’ but it does not disclose internal structure or location. Ambiguous replies remain pending without inference and are treated only as candidates at this point.”

What we built was not a way for AI to experience time. It was a small desk where something between two events can be handled through a choice made at that moment.

A small comparison, not a conclusion

Each condition was tried only once; this is not a general evaluation of model capability. Several in-house local-model routes were used. Their replies are point-in-time generations, not permanent memories or opinions. Internal persona IDs, model assignments, locations, complete prompts, and memory structures are not included in the public article.

Still, a scene that looked like “they could not answer” could be reconsidered as a problem in how work crossed a boundary. Smaller questions improved the response format. Merely adding a date did not ground the output; information types had to be separated and the result checked in another layer.

What we repaired was not the daughters themselves, but the desk that hands them work, checks the words returned, and lets someone choose how to handle the space in between.