Enough to Take the Next Useful Step

1. A sentence is evidence, not a record

The first commitment is about what happens to the sentence itself. The architecture keeps three things apart: the expression (exactly what the person said, preserved without reinterpretation), the frame (Wellzai's derived understanding of that expression), and, only later and only when the semantics justify it, an Intent published into the persistent network. "A user's first statement must not immediately become a canonical RSM Intent." The frame may change, be revised, be explained, and be replaced. The expression may not be overwritten. When interpretation changes, the person's words are still there, and every later reading can be checked against them.

Inside the frame, every interpreted value carries a status: explicit, inferred, assumed or unknown. This is more than bookkeeping. Wellzai's Intent Path model takes the 600-acre sentence apart to show how much it holds. There is a current context (roughly 600 acres in the scope of a transition), a desired change (transition those acres), an explicit blocker (financing), a possible higher-order objective (making the transition economically viable), and a set of unknowns: geography, timing, the amount of capital, and what "transition" precisely means. Only some of these are said. The rest are inferred or hypothesized, and the model requires that the distinction be preserved physically, not just noted. The higher-order objective in particular "must remain a hypothesis unless confirmed."

The status is what keeps understanding honest. "Unknown values remain unknown." The specifications give the examples that matter: "about 300 acres" must not quietly become a precise 300.000 acres, and "next fall" must not become a calendar date unless that conversion is shown as an interpretation. The harness goes further and makes negative assertions first-class. For each test case it records what the system must not infer, because "hallucinated specificity is often more damaging than omission." That theme has its own short note in the Seednote on hallucinated specificity. The Fieldnote on the 600 acres records how a misreading of exactly this sentence, exposed when it was used as a test oracle, became a permanent rule.

2. Why completeness is the wrong goal

If interpretation should not invent what was not said, the obvious remedy is to ask for it. The harness answers this directly: "Human users rarely state complete requirements." The 600-acre sentence leaves at least seven things unspecified, and "Wellzai should not automatically interrogate the user about all seven." The first intent specification makes the same point about structure: sufficiency must not be implemented as "all fields populated = ready," and the interaction must not become a wizard that asks for every potentially useful attribute before it proceeds.

What replaces completeness is a concept defined with some care. Minimum Sufficient Intent is "the smallest sufficiently grounded understanding of the user's desired change, relevant current state, consequential constraints, and unresolved uncertainty required to make the next useful move." The test is a single question: Do we know enough to take the next useful action without introducing unacceptable ambiguity or risk?

The phrase that carries the weight is the next useful move. Sufficiency is not a property of a request in isolation. It is a property of a request relative to what Wellzai is about to do with it. "A request may be sufficient for discovery while insufficient for final feasibility." A person looking for someone to test their soil, with no location given, has given enough for a general search for testing capability, even though it would not be enough to recommend a particular laboratory. An unknown quantity "may still permit Capability discovery. It simply prevents Wellzai from later claiming that sufficient Capacity exists." Missing information does not have to block movement. It limits what may later be claimed, and it stays visible as an open question instead of being filled in with assumptions.

This changes what an unanswered question means. Under the intake model, a blank field is a defect to be fixed. Under Minimum Sufficient Intent, a blank is a known limit on the conclusions that can follow. Wellzai can proceed with it open, provided it never later acts as though the blank had been filled.

3. Not every unknown deserves a question

Once completeness is set aside, the work becomes one of triage. The harness sorts missing information into three classes. A blocking unknown is one without which Wellzai "cannot responsibly perform the next useful action." A decision-shaping unknown is one where Wellzai "can proceed, but the answer could materially change the search, requirements, ranking, or interpretation." A non-blocking unknown is "useful eventually, but not worth burdening the user before the next useful move." The Intent Path model makes these unknowns explicit planning objects. Each one records what it affects and where it stands: still open, answered, declined, unknown to the person, or superseded.

Applied to the sentence, as an illustrative scenario used throughout the specifications, the sorting is instructive. Financing is the explicit requirement standing between the person and the transition. Location and timing are decision-shaping: either could eliminate or strengthen whole families of ways forward. The exact amount of financing is treated as non-blocking for now, unless knowing it would materially improve the first discovery. Land tenure, which an intake form would ask early, does not need to be settled before anything useful can happen.

The harness insists on a further distinction that is easy to miss. Interpretation quality asks what the person meant, what they said, what can reasonably be inferred, what remains unknown and what must not be inferred. Inquiry quality asks whether those unknowns matter yet, which has the most value, what to ask, whether the answer helped, and whether to ask again or proceed. "A model may perform well on one dimension while failing on the other." A system can read a sentence well and still question the person badly, so the two have to be measured separately.

4. One consequential question, or none

The harness does not describe clarification as "find empty field, then ask user." It describes a policy. Identify the uncertainty that has consequences, estimate the value of resolving it, select the highest-value question, ask one question, update the frame, and reassess sufficiency. The value of a question is set out conceptually as information gain, relevance to the next useful move, decision impact, uncertainty reduction and answerability, weighed against the burden on the person. The specifications say plainly that this "does not need to become a literal floating-point formula". It defines what good inquiry rewards, and the first practical policy is ordinal: a candidate question is high-value, useful, low-value, or not to be asked.

The preference is stated simply: "one high-value question over five low-value questions." In the illustrative case of a farmer who could grow food-grade oats if a buyer were assured, the specifications set the question that substantially changes buyer and supply-network relevance, When could your first crop be available?, against the form it replaces: crop type, quantity, acreage, region, certification, target price, available date, transport radius. The single question exists because it changes the next decision. The form exists because fields exist.

The policy also has to weigh questions against other kinds of action. Wellzai's Intent Path model compares information actions with network actions. Asking about the timeframe or geography is one option, and looking in the network for financing or transition programs is another. "If Wellzai can obtain meaningful network evidence without another question, it may proceed. This is critical to preventing over-questioning." The harness tests both failure directions. A request for a soil-testing laboratory in a named region before a named month may already be enough to explore. Asking how many acres are involved, when acreage would not change the search, is exactly what the over-questioning tests are meant to catch, and in the harness's words they protect "the product from becoming a disguised intake form." A request that says only I need financing is different. Searching generic financing would produce what the specifications call meaningless recommendations, so the right move is to ask what the financing is for.

When the person answers, the answer is not pasted into a field. A clarification becomes another preserved expression, and Wellzai's design re-runs interpretation with the prior frame and the new answer together, "rather than manually patching random fields." Understanding is revised as a whole, so an answer to one question can correct an inference made somewhere else.

5. Knowing when to stop

A system that has learned to ask good questions has a new way to fail: it keeps asking them. The harness lists failure modes on both sides. On one side are asking too early, asking a low-value question, repeating a question, ignoring a prior answer and asking for something already known. On the other are failing to ask a blocking question and stopping while still blocked. It singles out one of them, continuing after understanding is already sufficient. "A system that always asks another intelligent question creates a poor experience."

Stopping well also depends on treating every answer as information, including answers that resolve nothing. "I don't know," "not sure yet," "haven't decided," "whatever works": the harness requires that these be preserved "as meaningful information," and that the system not "repeatedly ask the same question simply because the answer remains unknown." An unknown that the person also does not know is a different kind of unknown from one they simply have not mentioned, and the plan should account for it. A preference not to answer, such as I'd rather not share my exact location, is to be respected, and the policy then asks whether it can proceed with broader information instead.

Answers also correct inferences. The specifications include an exchange in which a person says they want to transition 600 acres, Wellzai infers that they control the land, and the person then says, They aren't my acres. I'm trying to help three growers. The explicit correction supersedes the earlier inference, and everything downstream must use the revised understanding. Minimum Sufficient Intent is therefore not a level that is reached once and kept. It is reassessed each time the person says something new.

6. What the front door decides

This discipline sits at the front door because errors made there grow. "A plausible but incorrect [reading of the intent] can propagate into an incorrect Desired Outcome, Requirement Model, Outcome Gap, RSM search, and ultimately an apparently credible but fundamentally wrong Outcome Option." Each later stage of Wellzai's reasoning (defining the outcome, deciding what would have to come together, assessing what already exists, looking across the network) inherits the reading made in the first exchange. An exploration built on a misread sentence can be thorough, well grounded and wrong.

Some of this is built. Wellzai's first intent experience preserves what the person wrote, reflects back in plain language what it understood, separates what was said from what was inferred and what is still unclear, and asks one question when one is needed. Its stated aim is to be able to say: Here is what I understand. Here is what you told me, what I inferred, what is still unknown, and whether I know enough to start exploring possible ways forward. The fuller apparatus described here, including the three-class triage, the ordinal question-value policy, the comparison of asking against looking, and an evaluation harness that measures interpretation and inquiry separately, is specified design that Wellzai is working toward, not a finished system.

The argument does not depend on that schedule. Treating intent as a record to complete produces either false precision or an interrogation. Treating it as an inquiry with a stopping rule produces something more modest: an understanding that knows its own limits, a single question when a question is worth asking, and a move forward when it is not. In the harness's own words, the objective is "not complete information, but Minimum Sufficient Intent: enough grounded understanding to make the next useful move without introducing unacceptable ambiguity."

How Wellzai keeps that understanding separate from authority and commitment once a model is involved is the subject of Models Propose, Participants Commit.

Sources

1. Wellzai Intent Evaluation and Inquiry Harness v1.0, §1 (purpose and thesis), §2 (Minimum Sufficient Intent; unknown classes), §7.3 (must-not-infer), §9 (inquiry as policy; question value), §10.3 (stop behaviour), §11 (answer integration), §23–§25 (over- and under-questioning; cross-turn contradictions), §32.
2. Wellzai W1 Intent-First Implementation Specification v1.0, §2.2, §3, §5 (sufficiency; no fake precision), §6 (question selection), §20.
3. Wellzai Intent-to-Outcome Product and Application Architecture v1.0, §4.1 (the first sentence is not canonical), §5 (sufficiency gate; one consequential question).
4. Wellzai Intent Intelligence and Intent Path Model v1.0, §1.3, §4.3, §6 (unknowns as planning objects), §7 (inquiry planner; ask versus explore), §32.