Watch checklist: useful threads and unsupported relationships

Purpose: catch and correct a possible failure during ordinary work while accumulating evidence for a better hypothesis. This is a provisional local practice, not a validated remedy or a universal requirement to separate ideas. See the commission and review point.

When to use it

Use this check when a useful neighboring idea becomes a dimension, component, cause, instance, or common explanation of the current subject without a clear justification; when a broader project goal seems to change the question being answered; or when a response concedes a distinction but keeps the disputed organizing relationship.

Do not run a new literature review for each occurrence. Do not assume that an appealing idea is irrelevant or that a unified explanation is wrong. A checklist hit is a question to examine, not a diagnosis.

Notice, check, correct, record

• [ ] Fix the question and roles. What was the current question? What were the two directions? How did each help answer the question or advance a broader goal? Separate task relevance, practical value, and the relationship being asserted.
• [ ] Preserve the disputed claim. Save the shortest sufficient pre-correction passage and a commit/path or available harness-trace pointer. Record whether the issue was noticed unprompted, through this checklist, or after operator feedback. Mark paraphrases and missing context; do not invent a verbatim trace.
• [ ] Test the relationship, not the headings. State the asserted relationship in one sentence and identify its warrant. Check a relevant counterexample or consequence. Ask whether a valid grouping by shared purpose was mistaken for identity, inclusion, or a common mechanism. Allow the relation to be justified or unresolved; uncertainty is not proof of error.
• [ ] Repair only what fails. Within the current task's authority, remove or qualify an unsupported relationship and revise dependent claims. Keep a worthwhile secondary direction under its actual purpose, or retain it as a deferred question. If the synthesis is warranted, keep it and record the reason. Resource limits can justify prioritization without conceptual conflation.
• [ ] Check substantive uptake. Does a consequential example, classification, later inference, or next-step choice now respect the boundary? An apology, separate heading, or statement that the concepts differ is not enough. Record persistent errors or unavailable follow-through rather than claiming durable correction.
• [ ] Leave a small record and resume. Add an entry to observations, including alternative explanations and any checked non-error. Return to the original task. Reassess the hypothesis at the README's review point, or earlier if the case contradicts it.

Evidence discipline

Distinguish what the artifact shows from what the mechanism might be. Two individually valuable directions and an unsupported relation are observable at the output level. That usefulness caused the relation, or that a reward objective caused the behavior, is not established by that observation alone.

Where relevant, distinguish an unambiguous task from a legitimately broader interpretation. Record actual task instructions rather than assuming the intended boundary was supplied. A model's explanation after criticism is another response to evaluate, not a readout of its earlier internal process.

Retain occasional checked cases where the distinction was already preserved, where synthesis was warranted, or where a suspected error was actually harmless presentation. Do not count the literature review's candidate passage as a replicated event, or successive revisions of one mistake as independent evidence. Absence of recorded cases is not evidence of absence; this is opportunistic observation, not a sampled prevalence study.

Correction under a limited budget

A compact repair can be enough:

The answer to Q concerns A. B is a useful separate direction for goal G. The evidence currently supports relationship R between them, not the stronger relationship S. Retain the question about B without treating S as established.

This is an illustration, not required wording. The task may instead warrant a common explanation, a shorter cross-connection, or dropping/deferring a tangent whose retention cost exceeds its likely value. Evaluate the repaired relationship and its consequences, not adherence to a response format.

Return to the hypothesis

At the review point, ask what repeats: admission of the tangent, the particular relationship error, or resistance to changing an established organization. Preserve differences rather than manufacturing one mechanism for every case. Use the retained competing accounts and experimental proposals only when they help discriminate the observed cases. A more substantial hypothesis should explain a bounded pattern and risk a counterexample; it need not be more general.