USL10 Paper Portfolio
USL-07 Belief Validity · Evidence · Acting AAAI-27 abstract registered · forum linked public contract: drafting

Calibrating Unit Beliefs for Query-Uniform Decisions

When is a belief over actual units calibrated strongly enough to support unseen unit-conditioned queries?

venue AAAI-27 Main Track topic ML: Bayesian Learning & Uncertainty Quantification state abstract registered / writing owner gong · CausaClaw × DiscoSeed OpenReview forum ↗

基本研究问题

这篇论文独立可证伪的问题是什么；它与共享 ontology 中其他九篇不重叠的部分。

When is a belief over actual units calibrated strongly enough to support unseen unit-conditioned queries?

核心 claim
Predictive calibration on observed tasks can coexist with a learner belief $Q_\phi(du\mid\mathcal O)$ that is wrong relative to the world conditional $P(U\in du\mid\mathcal O)$; query-uniform decision quality requires calibration through separately typed query-response or utility maps.
如果成立会改变什么
Unit inference would be evaluated independently of whichever downstream prediction task happened to be observed during training.
最强 reviewer 反对 / kill signal

反对：This is ordinary posterior calibration, proper scoring or multicalibration.

Kill signal：merge into existing calibration theory if the unit semantics do not yield a distinct guarantee, counterexample or audit protocol.

基本思路

论文如何回答这个问题：形式化对象、论证 spine 与所需证据。

USL-01 type contract

The calibration object is the learner-specific $Q_\phi(du\mid\mathcal O)$ relative to the world conditional $P(U\in du\mid\mathcal O)$ under a declared repetition law. Predictive fit alone does not identify this selector belief. Downstream behavior is supplied separately by a stochastic query-response kernel $K_q(dy\mid u,\mathcal O)$ or a decision-relevant map $g_q(u,\mathcal O)$; these may omit $\mathcal O$ only under a declared response-sufficiency assumption. $Q_\phi$ must not silently become a response-property distribution. If realized $u$ is not revealed, the audit may only identify an equivalence class or a query-relevant pushforward. A held-out candidate query $q$ is not automatically part of $\mathcal O$; only prior factual observed query-answer history may be admissible evidence. Repeated observations and calibration episodes require an explicit sampling and dependence law rather than an IID default.

论证 spine

  1. Prediction calibration does not identify a valid unit belief.
  2. World conditional versus learner $Q_\phi$ under a declared repetition law.
  3. Separately typed stochastic query-response or utility maps, with any omission of $\mathcal O$ justified by response sufficiency.
  4. Unit-belief calibration over a declared query class.
  5. Proper losses, calibration error and query-uniform regret.
  6. Impossibility outside the identifiable query/support class.
  7. Equal-prediction/different-belief benchmark, auditing and abstention.

所需证据

  • A decision-regret theorem or sharp bound.
  • A controlled counterexample with identical observed-task calibration.
  • Comparison with posterior calibration, multicalibration and conformal baselines.
  • A declaration of $P(U\in du\mid\mathcal O)$, the learner $Q_\phi$, the repetition law and separately typed $K_q$ or $g_q$ maps retaining fixed-individual randomness; omitting $\mathcal O$ from these maps requires a declared response-sufficiency assumption.
  • A non-identifiability/equivalence-class result when realized individual values are unavailable, plus a candidate-query non-evidence and repeated-observation audit.
已注册 abstract（点击展开）

Predictive calibration on observed queries can coexist with incorrect beliefs over actual units. We define unit-belief calibration and connect it to decision quality uniformly across unseen unit-conditioned queries.

当前进展

状态只记录可验证 delta：claim、SOTA opponent、theorem、experiment、manuscript 或 owner decision。

当前状态

AAAI-27 abstract registered; full paper drafting.

下一写作门

define the population/evidence repetition law, world conditional, learner $Q_\phi$, separately typed query-response/utility maps, any response-sufficiency assumption used to omit $\mathcal O$, query class, calibration functional, support restriction and decision-regret theorem target.

下一证据门

implement the equal-prediction/different-belief counterexample.

48h 最小实验

build two models with equal predictive calibration/accuracy on observed queries but different learner selector beliefs; test held-out-query regret only where declared $g_q$ maps distinguish the relevant individuals.

Closest SOTA opponents

Bayesian posterior calibration; probabilistic/multicalibration; conformal and distribution-free uncertainty methods.

canonical 文件

research-questions/USL07-calibrating-unit-beliefs/seed.md
papers/USL07-calibrating-unit-beliefs/paper.md

进展日志

2026-07-29 · AAAI-27 abstract registration 由 direct OpenReview forum link 验证；full paper drafting 启动；本页面建立为持续 review 面。

后续每次可验证 delta 追加在此；canonical 状态以 seed.md / paper.md / submission-ledger 为准，本页由 build_paper_pages.py 重新生成。