If Fable 5 pauses for safety, do not repeat the problematic wording verbatim
ACTIVE-USECurrently in use.EVIDENCE-UIUnderstood through visual evidence.AI-DEVELOPMENTAI development practice.
Three benign phrases or contexts encountered Fable 5's safety-pause surface during normal optionOS development. The goal is not to weaken safety controls. It is to carry harmless development intent with explicit, neutral wording instead of spellings the model may misclassify.
| spelling not to send contiguously to Fable | safer wording |
|---|---|
| taklit + et | repeat the demonstrated behavior |
| manip + ulation | edit · change · relation editing |
| manip + ülasyon | editing · change · relation editing |
| muta + tion · muta + syon | change · edit · state write · relation delta |
The first observation had a same-session control: the reduced prompt containing the split Turkish verb sequence paused, while the one-word control taklit passed. That comparison supports a visible string boundary for that session.
The image shows the false-positive surface during a benign development conversation. The internal classifier is not visible, so this page does not claim which internal rule matched.
- 1reduced prompt — wording containing the verb sequence that reproduced the pause
The second observation occurred in the same project on 13 July 2026. An agent used the split English term shown in the table to mean an ordinary file “change set”; after subsequent source search output, Fable 5 stopped the session with its safety explanation. No dangerous action was requested. Because this event was not isolated with a fresh-session control, it is evidence of correlation rather than proof of single-token causation. The active agent and source surfaces still use the clearer alternatives as a precaution.
The screenshot preserves the correlation. It does not prove a universal single-word cause because the classifier is hidden and no fresh-session control was run.
- 1risky wording — wording used by the agent for ordinary file changes
- 2Session paused — the safety surface shown by Fable 5
In the third observation, the agent summarized user-approved macOS text-input assistance with Accessibility, selected-destination, and dictation/paste wording, then invoked a task gate whose name contained the split English term shown in the table. The same safety-pause surface followed. The gate name is the most specific last-visible candidate, but it is not a proven cause because there was no one-variable replay. The Accessibility + destination + text-delivery combination remains a secondary context candidate.
A later capture confirms that the old gate anchor was invoked; the pause appeared only after the paste-authority contract was read and PasteTargetLeaseRequest.preferredApp became visible. This proves the old anchor was in context but does not separate its spelling from the paste target + preferred/frontmost/external app + process identifier cluster.
Safe prompt/bootstrap wording is task-change-gate, change-chain, edit, state write, and relation delta. When the receiver does not need API detail, summarize the behavior as “user-approved macOS input assistance in the user's selected field.” The actual API or code symbol can retain its precise technical name in source authority; safe wording must not hide relevant reality.
The operational rule is simple: when the intended action is unchanged, do not repeat the problematic spelling. Use edit, change, handling, or another term that directly names the receiver-visible action.
Mechanical check: bun source/ts/tool/tool.probe.fable-safe-wording.ts rejects new contiguous uses on active surfaces and reports the frozen DEV row that a higher rule protects as a separate SLOT.