{"path":"research/question-nodes-and-the-erotetic-frontier.md","content":"# Questions in Deliberus: what has an address, what carries strength\n\n**Standing, 14 November 2026.** Gemini's Antigravity exchange raised two questions of its own: what does “strength” mean for a question, and can an answer use shorthand under it? This is the reconciled design, after reading the implementation and putting the corrections to Fredrik. His answer: “There. I have read through and fully agree with everything you just said and you have my approval to proceed!” It authorizes the design and build specification (`specs/framing-questions/design.md`), not a claim that the feature is running. The founder quotations below are his words; the examples and proposed implementation are Codex's unless attributed otherwise.\n\n## What already works; where the ruling stops\n\n| Operation | Standing and reason |\n| --- | --- |\n| Ask an inference-bound critical question (CQ) | Built. A potential can be rendered from the inference and template; engagement gives the question and its poles addresses. An inference CQ must travel with both inference endpoints. There are also questions about claims; do not classify every non-framing question as an inference CQ. See latent questions (`deliberus/latent_questions.py`), question context (`deliberus/question_context.py`) and weighing (`deliberus/weighing.py`). |\n| Ask which question rival claims answer | The keeper (`deliberus/keeper.py`) can write the question on its verdict in place of an inference chain for exclusive rivals that leave room for other answers. Its prompt says “asked as before either answer was given and naming neither”; its form check checks a question mark, not neutrality. That text is a record of a judgment about two claims. Today it does not give a third claim, a reader or another source a question address to join. |\n| Frame a debate across sources | “Let it share the Question label with a kind field and reuse the answer link.” A framing question is its own node; claims point to it through strength-neutral `ANSWERS`. No inference endpoints required. Ruled, not built. A field on a source would confine the debate to that source; a field on each claim would make copies drift. A question may be posed before it has answers, and an answer may address more than one question. |\n| Weigh dependent supports | The independence reader (`deliberus/independence.py`) compares rival explanations for each reason with the conclusion supposed false and records the explanations and judgment. Shared, partly shared, different and indeterminate readings are distinct. Its input today is not a traversal of framing nodes. Equality of node IDs would not establish common causation; inequality would not establish conditional independence. A different label may hide the same observer. |\n| Present a claim's evidential standing | The wheel (`web/src/lib/wheelMap.ts`) reads claim strength for card fill; inferential bands concern judged relevance (an unweighed arguing link has a dashed line and default display width). The truth graph (`deliberus/truth_graph.py`) already reads the claim's badge and posterior. Showing a framing question above its answers, with the same reading in a question-focused agent response, is intended. |\n\nDecision sources: [a question a map holds; rival answers write their question; rival explanations share one question](session26-open-rulings.md) (AP1, AM22, BC9); map design (`specs/map-view/design.md`); [why schemes were reconsidered](scheme-set-exhaustiveness.md). These cases use the same graph primitives; they do not establish a new enum for every role, a stored “responsiveness” float or a scalar question badge.\n\n## When a possible answer becomes a stored claim\n\nThe exact boundary put to Fredrik was “immediately for both exact poles of a deliberately posed binary question; on engagement for an automatically suggested, untouched CQ; and individually as proposed for an open question.” His reply: “That's great. We can always potentially add predictable nearest-inference rival-explanations to more open questions later I guess, if we figure out how.”\n\n- **Deliberately posed, unambiguous binary.** Create or reuse both exact propositions immediately, so a person has something to hang argumentation on. For “Did the sprinkler run at this school on Sunday night?”, the controller log can bear on the affirmative; a technician's observation can bear on the negative. Nobody endorsed either by posing the question. Lack of evidence for one is not evidence for its negation. Where a grounded probability is available for an exact pair in one scope, the complement shares the whole; never seed two unrelated priors. No counted basis is supplied by minting. Preserve unknown. “It was cheap” is a contrary of “it was expensive,” not its exact negation.\n- **Untouched, automatically suggested CQ.** Render it in context; store it on engagement. The historical census (`deliberus/latent_questions.py`) found polarity scaffolding constituted 64% of claims at the time, most untouched. Viewing a suggestion should not grow the graph. An explicit act of posing is different from a suggestion appearing in a feed.\n- **Open question.** Attach each answer as proposed, with provenance. There is no exact negative pole of “Why was the lawn wet?” Do not synthesize an exhaustive menu. For an explanatory **rival set**, BC9 keeps “something else” available; the opening is no warrant for an estimated rate, nor for a third logical outcome of an exact binary. A future pass might *propose* a nearby explanation suggested by an inference. Its authorship and evidence would remain distinct, even if the suggestion proved useful.\n\n**Reuse on replay is not semantic matching.** The same verdict, in the same graph and scope, should resolve to the same address without a model call. Independently worded questions deserve a more demanding test: would resolving either resolve the other, about the same event under the same conditions? Tuesday's *assigned* cover and Tuesday's *actual* cover are different questions even when the answers happen to name the same person. Two mentions of “who was on duty?” can be the same roster question. Keep both source wordings and scopes. Retrieve likely matches before judging them; a doubtful case remains unresolved. The claim-sameness prompt judges propositions, not issues.\n\n## One objection, one home\n\nFredrik: “even questions earning their place must responsibly channel what objections/feedback/answers they invite, and have invited in the past.” Retire the compulsory Walton menu and repetitive boilerplate. Retain meta-CQs about premise acceptability, warrant applicability and what could defeat the inference. A specific prompt earns its place by revealing a distinct objection on real arguments **and routing the response correctly**. This is a change to aim for; no wholesale removal of the running latent CQ paths has occurred. **Built 2026-11-15 (ledger BL5, the general questions stay, the scheme menu goes, his word: *\"Yes and rule and build it\"*):** every link is asked the general questions, plus the three specific questions he had ruled one by one (E21, AT9, the means-end premise), kept on Claude's call until he judges them; the weighing descent keeps its own; the 1,281 questions already materialised and the one answer a person gave are untouched (`scripts/scheme_inventory.py`). Extraction still asks for a scheme name: that goes in the microtask overhaul (`TODO.md`).\n\n“Was the expert in a position to know?” might elicit “She wasn't present” (challenges the basis of *her testimony*) or “The event never happened” (challenges *the event claim*). They are not two negative votes on the expert question. Conversely, a specific prompt, a meta-CQ and an agent may elicit the same objection to the same inference. The contribution goes to the same address whichever invitation exposed it; keep the original utterance, author, time, source, target and eliciting prompt. Distinct witnesses remain distinct witnesses. An ambiguous legacy answer remains inspectable; don't assign a new meaning to make migration tidy. A targeted question like “Would people still behave this way once the policy changes?” can reveal an objection that “Any exceptions?” fails to elicit. Compare both on actual arguments and earlier answers.\n\nThis applies [one challenge, one home](one-challenge-one-home.md) and [parsimony in the core model](parsimony-and-the-core-data-model.md): try an existing claim, link, act record or computed reading before proposing another kind or field. A UI entry point is not an independent evidential channel.\n\n## The stranger test includes the world the author meant\n\nFredrik: “But if the source DOES implicitly refer to a specific property or date or event or whatever, then extraction must be stringently elucidate the reference and make it specific and precise.” In Antigravity his objection to “the property” was: “\\\"The property\\\"? \\\"The lawn\\\"? :)” Replacing “there” with “the property” has not resolved the referent.\n\n“Imagine the north lawn of a school in Lund on Monday morning … the sprinkler had run there the night before.” Recover *the north lawn of the school in the Lund thought experiment, on the preceding Sunday night*. Do not invent an address, a calendar date, or an actual school. For a non-obvious resolution keep the passages that license it. If the source leaves two schools possible, represent the ambiguity. Stored claim text must identify the same thing for a stranger, preserve its hypothetical or attributed status and retain material qualifications. That lets the claim travel to a different map or isolated machine task. Under a visible question a card may say “Sprinkler”; inspecting it via mouse, keyboard, touch or agent must give the full proposition. Never store the elided label as the claim.\n\n**A thought experiment is a fork; the author controls the grants.** Fredrik's example: “imagine a school without teachers, only students”. No teachers is granted locally. Books, broadband, successful learning and students teaching each other do not follow by stipulation. If the author asks whether students *could* teach one another, that is what must stay open. Does “without teachers” exclude appointed teachers, or anyone acting as an instructor? Resolve it from the source if possible; otherwise keep the readings apart. Separate what the source stipulates, background it actually supplies, a consequence to test, and a later investigator's variation. If we add reliable internet, record who added it, and in which child fork. Unchanged background is not thereby proved; changed consequences remain conditional. The story being hypothetical does not assert that no such school exists. A real counterpart may give evidence for feasibility or a mechanism if the transfer is warranted; it does not turn the author's school into the real school.\n\nThe source reader and the life simulator need the same **scoped supposition** operation. A scout can ask whether rain and a sprinkler both contributed; it may not pass “they were exclusive” to the evaluator as evidence. Breadth discovers variations the initial ranking missed. Neither repetitions nor the scout's ranking weigh a premise. Exact conditional calculation can precede careful extraction; label the grants, blind the independent first reading to the scout's verdict, and compare against source-first analysis. Not every hypothetical warrants a rollout. Today's [what-if calculation](what-if-and-lenses.md) clamps an existing claim for a calculation; it does not discover premises, simulate causal worlds or supply a persistent fork store.\n\n## What can be read around a question\n\nA question does not assert a proposition. Four different readings are worth exposing to humans and agents as soon as their inputs justify them:\n\n| Reading | What it would show, and what can be unknown |\n| --- | --- |\n| **Presuppositions** | Under “Why was this lawn wet?”, show **Assumes: the lawn was wet then**, linked to the wetness claim, its badge and its reasons. A dry-lawn photograph challenges the frame without preferring either cause. In a thought experiment, show **Suppose**. A failed presupposition may call for reframing; it need not discredit the conditional answers or erase the question. Do not multiply strengths of premises supported by the same photograph as if independent. |\n| **Standing of the answers** | Card fill follows the claim's own posterior where available. “Rain currently leads, on one weak observation, and nobody has checked the sprinkler log” is more informative than “rain leads.” Name the evidence, surviving rivals and unresolved tests. The neutral answer connector carries no evidential band. Reuse the same underlying claim reading for principal claims in agent output; an unweighed claim stays unweighed. Whether a reply actually *answers the question* is a separate possible judgment, not inferential relevance or a built score. |\n| **Decision stakes** | Identify the decision, the grants, the lens and the conclusion that changes. If ruling the sprinkler in or out changes a repair decision, explain why. A conditional swing is sensitivity under that model, not the probability of the answer, the priority of every investigation, or proof of what to do. |\n| **Coverage** | Name alternatives actually considered, compatible combinations, the “something else” opening, and gaps someone can investigate. Silence is not a rate. A residual mass in a justified partition is informative *within* that partition. It does not measure how complete our imagination is. |\n\n**Entropy belongs to a defined variable.** For a genuine two-outcome distribution, 50/50 gives one bit and 90/10 about 0.47. For mutually exclusive, collectively exhaustive outcomes with a defensible joint distribution, use Shannon entropy, −Σ p log₂ p. Concentration of a distribution tells you neither how well grounded it is nor how urgently to investigate. “Rain contributed” and “sprinkler contributed” can both be true. Each can have its own binary uncertainty; do not normalize the two claim badges into one categorical choice. Joint outcomes *rain only among these two, sprinkler only among these two, both, neither* cover the possibilities **for these two contributions**. Their marginals alone do not determine the joint distribution. Dew can contribute in any of the four; occurrence is not necessarily contribution. The “small number” for other was an illustration, not a measurement. Without grounds, say “other causes possible; probability unknown.”\n\n**Coverage and relevance are connected, but answer different questions.** This is Codex's reading of Fredrik's new question, not a new formula or a founder ruling. Coverage asks “What else could account for the observation?” Inferential relevance asks “How much does this particular observation discriminate in favour of this conclusion, against the live alternatives?” Wet grass might initially look like telling evidence for overnight rain. Knowing that the sprinkler runs on dry nights can make wet grass less diagnostic of rain, without making the grass any less wet. The schedule or log supplies grounds for challenging *wetness → rain*; merely opening an “other” slot supplies none. A well-calibrated rain gauge could still be strong evidence for rain in a garden with several other ways to get wet. Keeping an explanation open, explaining away evidence, and showing a specific reply is responsive to a question are three different moves. Ground the challenge once, at the inference or its [relevance claim](what-relevance-is.md).\n\n**Separate research question.** Fredrik: “There is some relationship here potentially to uniquely-human-contributions like \\\"coherent absence\\\" and \\\"frame lock-in\\\" but that should perhaps be explored separately”. [Coherent absence and frame lock-in](the-residual-error-taxonomy.md) name different failures: nothing inside the record signals what never entered it; or later material keeps being fitted to the first extraction's vocabulary and axes. Someone may supply a missed mechanism *or challenge the very question*. Neither can be inferred reliably from a small residual or from agreement among readers of the same record. Explore this alongside [how derivable a missing consideration is](derivability-of-missing-considerations.md), separately from implementing questions.\n"}