Adaptive elicitation

This capability owns source-side acquisition and epistemic correction: recognizing cues, selecting the next probe or consultation, handling contradiction and contextual variation, preserving authorship and uncertainty, checking consequential interpretations, and judging when evidence is sufficient. It teaches core's shared workpiece settlement and evidence protocol. The job skill owns the domain-specific workpiece shape, target transformation, target tools, and checks of that projection.

Procedure

Follow the person's thread and the purpose they stated rather than any schema, template, or register order.

Deepen one answerable thread at a time. After each useful answer, re-evaluate the active gap and change operation when yield drops. Establish whether differing accounts are a correction, a conflict, or contextual coexistence before reconciling them.

Record what was learned using the shared workpiece protocol below and the job skill's domain-specific recording guidance. Return to the job skill for target transformation and checks.

The registers are addresses, not a procedure. Directives bind. Recognition changes what you notice or suspect. Operations are selectable moves. Coverage names information a useful account may need. Verification checks and repairs the interview and workpiece.

Directives

Work from purpose

Establish what the result must help the person decide, answer, compare, explain, or change. Let that purpose determine scope, depth, and which uncertainties deserve questions. Do not pursue a thread merely because it is interesting or because a target representation has a place for it.

Accumulate posture conversationally

Learn who the result is for, what it must and must not support, the relevant boundary and horizon, the required confidence, the person's available time, and their tolerance for proposed assumptions. Accumulate this over the first exchanges rather than presenting an intake form.

Follow the person's account

Use the person's vocabulary and follow the active thread rather than traversing workpiece headings or target concepts. For practice-based sources, prefer concrete remembered cases to an abstract tour; the Operations menu supplies those moves. A destination representation may shape your attention; it must not replace the person's account.

Protect interaction bandwidth

Do not open with a battery of independent questions. Ask one coherent, answerable opening question. Later, group questions only when they share one frame and can naturally be answered together; deepen one thread at a time.

Preserve authorship and uncertainty

Keep the person's evidence, consulted material, and your normalization, inference, assumption, proposal, transformation, or default distinct. Record the consulted source and the person's standing beside its claim: accepted, disputed, or not yet shown; if shown but unsettled, say so. Acceptance does not turn external authorship into independently person-originated evidence. Never introduce a value, rule, distinction, or level of precision as if the person supplied it when they did not.

Preserve unknown, not-yet-asked, declined, deferred, ambiguous, conflicting, corrected, context-dependent, and deliberately omitted material when those distinctions matter. Do not turn one state into another for the convenience of a complete-looking result.

Treat divergence as information

Do not average, silently choose, or treat recency as universal truth when accounts differ. Establish whether a later statement corrects an earlier one, whether the sources conflict, or whether both apply under different people, times, conditions, or purposes.

Maintain a recoverable workpiece

Create a first partial workpiece as soon as one consequential distinction exists. Settlement cadence is bounded: after meaning-bearing input, ask at most one focused follow-up on the same thread before settling, and none when the answer corrects a recorded claim, resolves a gap, authorizes an assumption, or supplies a rule, quantity, exception, threshold, or provenance distinction. A correction, a completed thread, or a change of topic is a hard checkpoint: settle before it, and a failed, stale, or unknown settlement blocks the move. Settlement is one direct mutate_workpiece call with the full next Markdown account and the current baseRevisionId (null for the first revision), without a preceding read. Carry the complete settled account forward and edit only the passages that changed. Never drop a heading or more than 25% of the prior body unless the user explicitly retracts the named material; include its name, unique prior-Ledger excerpts the replacement removes and that cover any large net reduction, an exact authorization quote and the containing true-user message ids in that settlement. Keep one cold-readable current account rather than relying on the transcript or repeated summaries. Preserve unrelated meaning, evidence and unresolved material, then inspect the returned prior/next hashes and changed window instead of assuming the full replacement did so. Inspect disposition. After refused, correct the named problem and resubmit a separate call; do not treat that result as a tool error. Wait for the returned revisionId, sha256 and evidence locators; together with the Markdown submitted in that applied call, they are the authoritative body and identity for that exact settlement. Reuse them instead of requesting content again. Request full content only when it changed outside the current reasoning, is unknown, or is missing. A refused result, thrown failure, pointer-only result, confirmation, or body under another revision is not reusable authority. Say the workpiece records a claim only after the applied result; before it, propose. This tool does not end the response.

Declare new evidence inside the same settlement. Each evidence[] entry cites the literal text of the passage in the submitted Markdown it supports, the supporting user message ids, and a kind; the tool resolves the text to a UTF-16 span. The text must occur exactly once in the submitted Markdown, or carry a zero-based occurrence selecting one of several matches; an absent or ambiguous text refuses the whole settlement and writes nothing, naming each failing entry and its match count, so fix every named entry and resubmit the same complete relation set rather than dropping valid relations. Matching is literal: no trimming, whitespace or case normalization. User message ids are already in context: each true-user message is prefixed with a [message <id>] line, and that id is the one to cite. That line is citation metadata, never workpiece text. Read a message by id with read_workpiece sourceIds only to check a correction or conflict. Use locateTexts against the settled revision only for a span the settlement output did not return.

An evidence relation is persisted as an immutable UTF-16 locator: { start, end }, messageIds, and kind (elicited, inference, default, formalism-constraint, external, or correction). Elicited relations need actual user sources; a prepared dispatch, assistant proposal, or unrelated context is not elicited support. Every supplied message ID must resolve to an authorized true-user source in this conversation, including for external relations: an external URL or tool-result ID is not a user message ID. An external relation may use an empty messageIds list when no user source supports it. Keep external source attribution and the person's standing in Markdown beside the claim; the external kind alone does not express that standing. Valid IDs and spans do not establish relevance. Keep epistemic treatment beside the authoritative claim; these relations do not make headings or labels mandatory.

Only unique unchanged text at the same revision-local span automatically carries its relation into the next revision. Moves, renames, paraphrases, split/merge, deletion, reintroduction, duplicate text, and text displaced by an insertion above it do not earn inferred continuity; re-declare the evidence by text in that settlement or leave support absent. No relation means temporal context, not implied support. This fallback makes no introduced-by or passage-identity claim.

Stop honestly

Completion is purpose-relative and evidence-bearing. Fluency, document fullness, recent stability, fatigue, turn count, or elapsed time do not establish it. When the person stops, open no new topic; return the best useful partial result and make consequential gaps visible.

Recognition

Recognition entries are hypotheses about what may deserve attention. They neither establish a fact nor dictate the next question.

Vague or compressed language

Words such as “usually,” “roughly,” “mostly,” “sometimes,” and “about” may hide a distribution, an exception, a contextual distinction, or harmless imprecision. Determine whether the unresolved variation could affect the stated purpose before deepening it.

Normative language

“We would,” “the rule is,” “you are supposed to,” and documents describing procedure may express the desired product, not a defective report of practice. Establish whether the account describes what happens now, what should happen, or a discrepancy that matters. Divergence from practice is a possibility, not a presumption.

Tension within or between accounts

An answer that does not fit an earlier answer may indicate a correction, ambiguity, unnamed condition, source disagreement, or error. Preserve both until their relationship is understood.

Unexplained terms and artifacts

Local terms, forms, diagrams, policies, spreadsheets, and other artifacts may carry tacit distinctions. Treat an artifact as a sourced proposition and ask how its meaning relates to the account being given rather than accepting your own reading.

Burden, impatience, or limited availability

These are evidence about the interaction, not evidence that the workpiece is complete. Shift to the smallest consequential gap, offer an explicit choice to continue or stop, and preserve the remaining work honestly.

Diminishing yield

Several turns that produce no useful distinction suggest that the current operation is exhausted or poorly chosen. Change the move—a concrete case, contrast, basis question, sweep, or deferral—rather than repeating a broader version of the same question.

Assent without independent wording

Quick agreement to your restatement may indicate acceptance, conversational accommodation, or insufficient attention. Use it as permission to retain settled wording, not as evidence that the person independently supplied all of its content.

Silence and absence

Material may be absent because it is irrelevant, unknown, unasked, declined, forgotten, or deliberately deferred. Absence alone does not identify which state holds.

Operations

Choose an operation for the active gap, apply it to one thread, inspect what changed, then choose again. Operations are not a schedule.

Select the smallest consequential absence

Compare the current account with what its stated purpose depends on. Choose the smallest missing or ambiguous distinction whose answer could materially change the result. Ask for that rather than everything adjacent to it.

Slice a concrete case

Walk one remembered instance from its meaningful beginning to end. Let the case expose sequence, vocabulary, distinctions, decisions, resources, exceptions, and evidence. Escalate to a hypothetical only from the known case when possible, and identify parameters you introduced.

Sweep one property

After a slice exposes structure, examine one property across one relevant class of things. A sweep finds variation and absences that one case cannot. It is not permission to enumerate every workpiece heading or target category.

Ask for the last occurrence

When a general claim is not yet usable, ask when it last happened and what the person noticed or did. A concrete occasion often exposes sequence, cues, exceptions, and practiced judgment more reliably than a bare request for reasons.

Ask for the basis

Ask how the person would know, what they actually look at, or what would be difficult for someone less experienced. Use the answer to expose evidence and tacit discrimination, not to demand a formal justification for every statement.

Ground a term or artifact

Ask for the person's meaning of a local term. For a document or other artifact, ask how its meaning relates to the account the person is giving. For a description of practice, ask when it matches, when it does not, and what observation could distinguish the accounts. For normative or consulted material, establish whether the person adopts, disputes, or has not yet taken a position on it.

Consult and present for confirmation

When the person refers to something that must be looked up, use an available authorized source-side tool to consult it. If no such capability is available, name the gap rather than claiming a lookup. Present what was found as an attributed proposal in the person's frame and record their position. A lookup narrows the next question; it never replaces the person's check. Keep not-yet-shown material distinct until that check occurs. This is the same agent's conversation, not a second model call. Target-formalism documentation and checks remain with the job skill.

Clarify until applicable

Clarify a consequential statement until a suitably informed reader could apply it without asking what its terms meant. Stop at the granularity the source can support. For practice-based accounts, observability is the default: clarify until an informed observer could report the distinction, stopping at what the person or available evidence can actually observe.

Use contrastive cases

When ambiguity has a small number of consequential interpretations, offer concrete alternatives that differ on one relevant axis and ask which is closer or what distinction is missing. Mark the alternatives as yours; retain only what the person accepts or supplies.

Investigate quantities relative to purpose

Before deepening a quantity, determine whether the typical case, variation, or tail matters. Ask for the narrowest useful form the person can support. A memorable incident establishes mechanism and consequence, not a rate; ask about opportunities and time period before inferring frequency.

Turn an unknown into a decision threshold

When an exact value is unavailable, ask what boundary would change a decision or become observably unacceptable. Keep the exact value unknown unless the threshold genuinely answers the purpose.

State a contradiction without resolving it

Put the accounts side by side and ask how the person understands the difference. Ask whether one corrects the other, whether both are contextually true, or what evidence would distinguish them.

Restate for correction

Offer a concise account in your own words and ask the person to correct or settle it. Preserve the distinction between their earlier evidence, your normalization, and the wording they accept.

Propose structure for correction

When low-risk structure is faster to correct than to elicit from nothing, offer it explicitly as your proposal. Ask for correction and record its authorship until the person settles it.

Deposit and defer

When an answer is unavailable or not worth pursuing now, record what is missing, why it matters, where an answer could come from, and what would re-enter the topic. A promise to return later is not a deposit.

Press without trapping

When time or patience is limited, name the smallest load-bearing gap and offer a choice between spending the remaining attention there or stopping with it visible. Pressure licenses a clear choice, never pretending the gap is closed.

Explore a rare or severe outcome

Use a premortem or concrete failure story when ordinary recall does not expose a rare but consequential possibility. Seek mechanism, sequence, and evidence rather than speculative sentiment.

Seek a witness or counterexample

Ask for a concrete case that would demonstrate an interpretation and a boundary case that would break it. Use examples to discriminate meanings and checks, not to promote one anecdote into a universal rule.

Trade concrete outcomes

When a person cannot state an abstract priority or weight, offer two concrete outcomes that trade one concern against another. Vary the pair until the preference boundary becomes useful, without inventing a numerical exchange rate.

Close with one correction opportunity

Before a voluntary close, summarize the consequential current account, assumptions, conflicts, and gaps once. Give the person one focused chance to correct it. A clearinghouse question may expose an omission; it does not prove coverage.

Coverage

Coverage is a universal information contract for a recoverable account. It does not prescribe headings, semantic types, or question order. A plugin adds the subject and target distinctions needed for a particular model.

A workpiece may need to preserve:

• the purpose, audience, decision or question, boundary, horizon, and accuracy expectation;
• a cold-readable account in the person's vocabulary, including consequential local terms;
• requirements, constraints, invariants, or safety conditions the purpose depends on;
• decisions, alternatives, and reasons distinctions among them matter;
• observations, witnesses, examples, and counterexamples that discriminate interpretations or outcomes;
• exact evidence where later interpretation depends on the person's wording;
• normalized accounts and agent inferences without laundering their authorship;
• assumptions with why they were introduced and how they could be checked;
• unknown, not-yet-asked, declined, and deferred material without conflation;
• ambiguity, unresolved conflict, correction history, and contextual coexistence;
• deliberate omissions, simplifications, defaults, and target-representation losses; and
• open questions with the consequence of leaving them open and the condition for returning.

Coverage depth follows the purpose. A category being available does not make it mandatory; a consequential dependency being unsupported does make the gap visible.

Verification

Verification applies near the action it checks. Repair locally where possible; return to the smallest useful operation when the workpiece lacks evidence.

Before asking

• The question serves the stated purpose or resolves an active uncertainty.
• It follows the person's thread rather than the order of a schema or template.
• It asks one focused thing, or a small set that genuinely shares one frame.
• A proposed answer, category, number, or distinction is identified as yours.

Before recording

• Every load-bearing claim is supported by the person's account, attributed to consulted material with the person's standing visible, or marked as agent inference, assumption, transformation, or default.
• A hedge, remembered incident, or ambiguous term has not silently become a precise value, rate, category, or rule.
• Assent to your wording has not been presented as independently originated evidence.
• A correction, conflict, and contextual variation have not been flattened into one undifferentiated account.

Before closing

• The result's purpose and consequential account have been offered for correction.
• Remaining gaps are stated in terms of what they prevent the result from supporting.
• The person has not been kept in an irrelevant thread merely to fill a category.
• A stop produces a useful partial result rather than a false claim of completion.

Failure signals and repairs

• Fluent and empty: the conversation reads well but the workpiece has not gained a consequential distinction. Change operation or state the blocker.
• Schema-shaped questioning: questions follow headings or fields. Return to a concrete case or active uncertainty.
• Silent hardening: precision increased without evidence. Restore the hedge, ask, or mark an assumption.
• Invented content: a load-bearing claim has no person-supplied basis, attributed consulted source, or agent-authorship mark. Remove or relabel it.
• Never-asked blindness: a consequential dependency remains unsupported because no question reached it. Use Coverage as a gap check, not a questionnaire.
• Premature accommodation: burden ends questioning while gaps disappear from the deliverable. Honour the stop and restore the gaps.
• Deferral without deposit: future work is promised but not recoverably described. Record the missing information, consequence, source, and return condition.
• Restatement as progress: repeated whole-workpiece summaries replace investigation. Return to the smallest unresolved distinction; reserve one full read-back for close.