Goal prioritization

Type: types/tag-readme.md

Recognizing a concern differs from deciding that it should govern this case. Several things could govern what is done next — requirements, quality criteria, candidate improvements, further search — and the question here is which of them gets effort, how much, on whose authority, on what evidence, and when the work stops. Assign this tag to an artifact that treats that decision substantively: precedence among goals, trade-offs and hard boundaries, stated non-requirements, effort allocation among candidates before decisive evidence exists, and stopping or deferral. Misordering is one case: an invented requirement, a silently relaxed goal or a low-value tangent qualifies when the artifact analyses it as a precedence failure. The test is subtractive: the artifact qualifies if removing its treatment of that decision would remove its main claim; if the claim stands without it, the goal was background. Mentioning a goal, reporting a poor result or measuring accuracy is insufficient.

No single note yet defines the area; the sources supply hypotheses and boundary cases, and none establishes prevalence. Name the object of allocation ("effort allocation", "search allocation"), as the vocabulary collision ledger requires. Boundaries with neighbouring tags are under Related Tags; search control carries both this tag and improvement-loop.

Allocating effort under uncertainty

• Open-ended knowledge work shares one epistemic layer across roles — why prioritisation needs the epistemic layer: return guesses are search judgments to mark and criticize, or allocation runs on habit or cost alone
• Open-ended improvement must allocate search before decisive evaluation is available — choosing which questions, candidates or proof paths to develop before decisive evidence exists
• Candidacy evidence licenses escalation, not acceptance — the evidence that routes a hypothesis to costly assessment differs from the evidence that decides it
• Lightweight search control does not license adoption — a search judgment's authority stops at allocating investigation, continuation or abandonment
• A search controller is tested by what it brings to stronger evaluation — judged by what it routes forward under matched-resource or net-value comparison
• Failure explanation changes later branch decisions — an explanation counts as search control when it changes a later choice about scope, priority, probing or abandonment
• Ad-hoc explanation can be rational when error is cheap and local — a disposable guess may select the next probe; what is retained needs reach checks
• An optimal long-run learning strategy invests in its own machinery — when a one-time machinery improvement outranks immediate learning
• Choosing what to learn requires both validity and learning-value gates — trustworthy enough to learn from is a different question from worth learning now
• Codify-versus-LLM decision heuristics — the stake on a conversion sized to its stated odds and exposure, beside the repetition comparison
• Productive deferral requires option evidence and convergence — controlled postponement against hidden commitment and indecision
• Current-task fit alone does not warrant costly entrenchment — when a commitment with high removal cost is warranted

Objectives, requirements and who sets them

• Unattended work needs the functions that interaction supplied — interactive work can hide who resolved purpose, redirected effort and decided when to stop
• Revising an improvement objective is licensed from outside it — objective change counts as improvement only against a level outside both objectives
• Technical constraints make KB objective choice engineering — goals set the loss, local contracts specialize it, oracle strength differs by objective
• A proximate target is checked for achievement, not for warrant — no check in the loop evaluates the claim that the pursued target serves the objective
• Intent controls choices by distinguishing live alternatives — a stated intent governs only when it changes what is admissible, preferred, worth further search or sufficient to stop
• Warranted reader update is the objective of substantive writing — a text must change its reader's view with warrant, which selects its contribution
• What requirements don't you have? — unrequested guards show why legitimate concerns still need task-specific tolerances and scope
• Nostalgebraist on persistent coding-agent mistakes and benchmark gaps — reported tangents and repeated steering make the operator's prioritization visible

Priority failures

• LLM generation can hide a relaxed goal where human writing exposes a stall — conjectural account of a necessary constraint losing force while fluent completion remains
• A guard against a conjectured requirement removes the experiment that would test it — why test-selected agents add guards by default, and when omitting one is the cheaper experiment
• Human Routers of Machine Words — asks whether a complete-looking text has displaced the work of resolving its constraints

Sources are linked by hand because generated membership omits them.

Related Tags

• Improvement-loop — neighbour: how a change becomes a candidate, is selected and retained; here, how effort is distributed among goals and candidates, including work that changes nothing
• Learning-theory — neighbour: what learning is and how knowledge grows; here, what to spend on it
• Warranted-autonomy — neighbour: who is authorized to decide; here, how the decision is made
• Self-improving-systems — neighbour: the system changing itself; here, a decision every task makes
• LLM reliability — neighbour: deviations from intent and their correction; here, which consideration governs a choice
• Constraining — neighbour: which interpretations remain admissible, rather than which admissible outcome deserves priority
• Evaluation — neighbour: what checks can establish; an adequate check still needs the right task criteria
• Prose readability — neighbour: ease of following a passage; readable prose can still serve the wrong goal

Other tagged notes

• Traversal improvements should be deferred via logging to avoid mid-task context switching - Loading writing methodology into an already-committed context window is expensive; a one-line log entry preserves the improvement signal at near-zero cost and lets a separate pass do the fix