{"path":"research/active-inference-context-acquisition-and-deliberus.md","content":"# Active Inference as Context Acquisition, and the Half of the Questioning Problem Deliberus Hasn't Built\n\n**Date**: 2026-08-23 · **Source**: Dutta, Ramachandran & Sra, *Active Inference as Context Acquisition for AI Agents*, [arXiv:2608.19202](https://arxiv.org/abs/2608.19202), founder-relayed. **Read depth**: abstract plus a structured extraction of the full text (formulation, the Optimal Question Asking mechanics, benchmark results, limitations); the theorems are reported as the paper states them, not re-derived.\n\n## What the paper does\n\nAn agent facing a task with missing information must constantly decide: **ask a clarifying question, or act on defaults?** The paper formalizes that choice as minimizing expected free energy under cost, which in the deterministic case reduces to something plain: **ask exactly when the expected information gain exceeds the price of asking** — `I(x;o|a) > λ·c(a)`, one threshold with a tunable exchange rate between bits and cost. Their testbed, Optimal Question Asking, is Twenty-Questions with a fixed attribute menu and an exactly computable optimal policy (dynamic programming over candidate splits). The headline empirical result: **frontier LLMs are measurably question-inefficient** — they reduce uncertainty every turn but ask 1–3 more questions than optimal on small sets, with the gap *widening* as candidate sets grow. Not catastrophic; consistently myopic.\n\n## The five dovetails, from most-shipped to most-missing\n\n### 1. Deliberus already ships the epistemic half of their decision rule — and their LLM result validates *how* it ships\n\nThe hinge-ordered questions endpoint ([nesy-rag-note.md](nesy-rag-note.md); `deliberus/question_ordering.py`) ranks a claim's open critical questions by outcome-swing — how far a decisive answer would move the badge. That IS the paper's information-gain term, instantiated on the strength graph instead of on candidate-set entropy. And the paper's myopia benchmark is empirical support for a design choice this project made on other grounds: **the question-selection policy is computed deterministically through shipped semantics, never delegated to the model's own judgment** — which the paper now shows would cost 1–3 wasted questions per exchange, in the one currency the descent's scaffold treats as scarcest (a user's attention: taps, not essays — [../depth.md](../depth.md) § the seven holds). The arsenal-and-harness split, measured from the outside.\n\n### 2. The missing half: the cost term and the stopping rule\n\nWhat Deliberus's ordering does *not* have is the paper's right-hand side. Questions are ranked but never *priced*: nothing decides when the best remaining question is no longer worth a human's next tap, so a descent's only stopping rules today are \"answered everything\" or \"user walked away.\" The paper's threshold supplies the principled third: **stop offering when the top question's expected swing falls below the attention price** — a named, tunable λ. This lands on two standing findings at once: the curiosity research's *rate-of-understanding* problem (engagement thins exactly at depth, [curiosity-as-growth-fuel.md](curiosity-as-growth-fuel.md)) and the \"no gap without a move\" principle — with the refinement that below the threshold, the honest display INVERTS: show the remaining gap as a *state* (the absence-is-a-state family, P22) rather than as a question, because a question whose answer moves nothing is an essay wearing a tap's clothing. Buildable small: the CQs endpoint already returns `hinge_sensitivity`; a `worth_asking` boundary against a Layer-3 λ constant is one field. Filed in TODO. **Built 2026-08-23, and the build refined the design**: no λ constant exists after all — the operational price is *display-band crossing* (a question is offered iff a decisive answer could move the badge across one of the founder-ruled display bands, `qbaf.display_band`), so the threshold lives in the ontology's own units of meaningful difference and there is no arbitrary number to tune. Below the line the endpoint returns a confessing summary and the claim page shows the state line with an expand affordance; agents get `strength_if_yes/no` + band fields to apply their own price — advisory, per propose-never-assert.\n\nThe same rule retro-justifies a placement made on philosophical grounds: the weighing descent's state-deficit and satisfier-belief screens sit *before* residue typing ([decomposing-value.md](decomposing-value.md) §6) — and on information-gain grounds that is exactly right, since each can dissolve the entire disagreement (maximal expected entropy reduction) while a typing question only labels it.\n\n### 3. For agent readers, the graph shifts the ask-or-act threshold itself — the decision-theoretic form of the commons argument\n\nThe paper's agent chooses between asking (costly) and assuming (risky). A reasoning commons gives every agent a **third arm: query the graph** — information gain at near-zero cost, no human attention spent, no privacy surface touched. In the paper's own terms, each mapped premise permanently lowers the λ-weighted cost of acquiring that context for *every future agent's every future task* — which is [lowering-the-cost.md](lowering-the-cost.md)'s Romer-nonrivalry argument reappearing as a decision rule: a rational context-acquiring agent should drain the free arm before touching the expensive ones. And the instruments are what make the arm *usable* for active inference specifically: an acquisition loop needs to know its own residual uncertainty, and the graph quantifies exactly that — open critical questions are the agent's remaining entropy, the completeness oracle names what was never examined, the hinge says which unknown moves the conclusion. This sharpens [agents-as-a-consumer-class.md](agents-as-a-consumer-class.md)'s \"the honesty instruments turn out to be the API\": they are, precisely, *the uncertainty bookkeeping an active-inference reader cannot do for itself*.\n\n### 4. The clarification-first surface gets an ask-or-skip criterion\n\nThe landing preflight and the interpretation checkpoint ask disambiguation questions before extraction. When should they? The paper's rule, translated: **ask about a contested sense only when the candidate readings genuinely diverge downstream** — when the descent that follows reading A differs from the descent that follows reading B. If every sense leads to the same structure, the question has no information gain and conservative relevance says skip it. That turns a UX instinct ([disambiguation-ux-and-voice-first-contribution.md](disambiguation-ux-and-voice-first-contribution.md); the P-series' conservative-relevance rule) into a computable check.\n\n### 5. Their stated limitation is our confession channel, inverted\n\nThe authors admit OQA assumes fixed attribute tables and deterministic answers, while real dialogue brings vague and contradictory replies. That failure class has a name here: **every taxonomy breaks on the next register**, and the standing fix is a `does_not_fit` value with a required reason ([taxonomy-gaps-and-the-closed-enum.md](taxonomy-gaps-and-the-closed-enum.md)). Their framework would inherit it directly — reserve probability mass for \"outside the table\" and treat its accumulation as menu-revision pressure. In the other direction, their benchmark discipline (measure the model against an exactly-known optimal questioner) is a template the descent could someday borrow: on synthetic corpora the optimal question order is computable, so the gap between the shipped ordering and optimal is *measurable* rather than assumed.\n\n## The dual-use paragraph, because the paper flags it and our threat model has the slot\n\nThe authors note that efficient questioning is also efficient **extraction** — an optimal questioner pointed at a person is a profiling instrument. Deliberus's descent asks people about their *values*, which is the most sensitive extraction target there is, and the scrutiny-gap result ([the-scrutiny-gap.md](the-scrutiny-gap.md)) already established that machine advantage runs through throughput. Optimal questioning is throughput applied to *elicitation*: the same mechanism that makes the descent efficient makes it, in adversarial hands, an interrogation optimizer. The existing mitigations are the right shape — the descent is invitation-only, stopping anywhere is a first-class state (hold seven), and the sacredness re-scope frames consent-and-timing as the boundary — but the threat model's strategy-class inventory should carry the entry explicitly: *question-order optimization against a person* is an interest-to-strategy-class capability, and publishing better question-ordering publishes it for both uses.\n\n## What does not transfer, honestly\n\nOne more thing the paper does *not* supply, worth stating because the founder asked exactly this: **the price itself is assumed, never derived.** `c(a)` is an exogenous cost tag on each action, uniform per question in their experiments (which is why their \"optimal\" collapses to fewest-questions), and λ is a designer-chosen exchange rate between cost units and bits. The theory contributes the *shape* of the rule — compare the gain to a price before asking — and gives no help choosing the price's value. For us that meant choosing the price's units ourselves — and the build (same day) chose better than a tunable: the price is expressed in the display-band structure the founder already ruled on, so \"worth one tap\" = \"can change what the reader sees,\" and no new constant exists. The paper warrants the threshold's existence; the ontology supplies its units.\n\nTheir oracle needs enumerable candidate sets (they cap at 300, and note scaling breaks beyond); a person's possible positions are not enumerable, so the dynamic-programming machinery stays home — the import is the **decision rule and the cost term**, with QBAF-swing as the practical surrogate for information gain. Their deterministic-answer assumption fails hardest exactly where Deliberus lives (a weighing answer is not a clean partition). And the active-inference branding carries Friston-adjacent metaphysics the paper itself strips away in the deterministic case; what's worth importing is the stripped core — information gain against priced attention — not the framework's cosmology. Filed as the ninth-or-so independent arrival at the reasoning-layer slot from yet another side (decision theory / interactive QA), joining the pattern the corpus tracks: adjacent, load-bearing, and not overlapping.\n\n## Cross-references\n\n[nesy-rag-note.md](nesy-rag-note.md) (the sibling arrival whose design yield this paper prices) · `deliberus/question_ordering.py` + `deliberus/hinge.py` (the shipped epistemic half) · [curiosity-as-growth-fuel.md](curiosity-as-growth-fuel.md) (rate-of-understanding; the pacing problem the stopping rule addresses) · [../depth.md](../depth.md) § The Descent, Defined (the attention economy the λ prices) · [agents-as-a-consumer-class.md](agents-as-a-consumer-class.md) (the reader this makes formal) · [lowering-the-cost.md](lowering-the-cost.md) §5 (nonrivalry, now as a decision rule) · [taxonomy-gaps-and-the-closed-enum.md](taxonomy-gaps-and-the-closed-enum.md) (the confession channel their limitation wants) · [the-scrutiny-gap.md](the-scrutiny-gap.md) (throughput; the dual-use frame) · [live-election-test-design.md](live-election-test-design.md) (each kartpaus ends with ONE question — now: the highest swing above the price, or a pointed state when nothing clears it)\n"}