E Evidence Press Shop Donate

Press release · 8 October 2026 · version 1.1.0-candidate

Four tests for binary informational substitutes

For two yes-or-no signals, four exact inequalities decide whether every partial disclosure preserves a specified diminishing-returns property.

Listen to this briefingNarrated summary · OpenAI API synthetic voice (fable) · MP3 · download
▶
Watch this briefingVideo summary · YouTube · watch · play on this page

Summary

The question. Imagine two forecasters, each holding a private yes-or-no signal about the same uncertain event. One of them releases a randomised summary of their signal. How much extra value does the other forecaster's complete signal still provide? The strong informational substitutes property studied here is a diminishing-returns requirement: the extra value of the second signal after a partial disclosure must be at least as large as it would be after completely revealing the first signal. It applies in both orders and to every admissible randomised disclosure. Questions of this kind arise in forecasting and information economics.

The difficulty. Two signals can be useful together without behaving well under every partial disclosure, and there are infinitely many possible partial disclosures to check.

What was found. For a binary event and two binary signals under quadratic scoring, this candidate gives an exact way to tell the difference: four inequalities, calculated from the joint probability table, settle the property. Zero-probability cells are included. When an exact rational input fails, the package returns a two-report disclosure and its negative expected-score gap, which can be checked directly from the raw probability masses.

Status. An unrefereed Evidence Press candidate, developed with AI assistance. The universal claim rests on an informal analytic proof; the exact programs provide algebraic and finite verification evidence. No proof-assistant formalisation or external human peer review is claimed.

Summary for specialists

“Value” has a precise meaning: the expected square of the posterior probability of the event. It is not money, accuracy on a particular dataset, or every possible decision-maker's utility. The exact target is Chen–Waggoner's March 2017 Definition 2.2.3, not arbitrary-pair submodularity over the full continuous information lattice. One signal is partially disclosed, and the other is observed whole; the comparison is against full revelation of the first signal.

Exact classification. Let $r_{ib}=\Pr(B=b\mid A=i)$, let $D_A$ be the difference between the event's conditional means in the two $A$ rows, and let $d_b$ be the corresponding difference after also conditioning on $B=b$. On cells shared by the two rows, define

$$M_{Ai}=D_A^2-\sum_{b\in\mathrm{overlap}}r_{ib}d_b^2.$$

For nonconstant $A$, the oriented condition holds exactly when $M_{A0}\ge0$ and $M_{A1}\ge0$. Apply the same calculation with $A$ and $B$ exchanged. A constant source is handled separately, without undefined conditional probabilities. For a binary event and two binary signals, these two endpoint tests in each order are necessary and sufficient.

Further results.

  • Projective substitutes. Proposition 4 proves that these strong conditions imply projective substitutes in this binary quadratic model. The converse fails: its strict rational separator has $M_{A1}=-3/256$ but passes all 32 projective comparisons; the displayed counter-disclosure has score gap $-323/8458240$.
  • One-direction extension. The main proof extends in one direction to a binary source and a finite second signal, and to square-integrable real or finite-dimensional vector targets under quadratic loss. It does not supply a full two-sided classification when both signals have arbitrary finite alphabets.
  • Complements. An appendix treats the reverse, strong-complements condition with endpoint and, when needed, interior-vertex tests.

The broader programme for arbitrary scoring rules, arbitrarily many signals, or arbitrary pairs of partial disclosures is not claimed solved.

Technical account

Condition on the posterior probability of the first signal after its disclosure. A conditional-variance identity expresses the score gap as an average of a residual curve, multiplied by a nonnegative factor. The residual is concave, so its endpoint values decide nonnegativity across the entire interval. Exchanging the signals gives the other two tests.

One detail is easy to get wrong: the proof's residual satisfies $H(0)=M_{A1}$ and $H(1)=M_{A0}$, so the endpoint order is reversed relative to the row labels.

The banner depicts that endpoint mechanism schematically. It is not experimental data or an additional certificate.

Evidence, assurance and limitations

The universal claim rests on the written proof. Internal exact checks cover algebraic identities, all 255 nonempty support patterns on a finite count grid, direct disclosure scores and 9,236 recorded failure witnesses. The 6,560 count vectors include repeated normalised distributions; they are not 6,560 distinct models or a prevalence sample.

Publication revisions add exact appendix/projective identities on 12,800 nonconstant orientations, all 32 projective comparisons for the separator, and rejection tests for invalid inputs and forged witnesses. Normal and optimized Python runs retain the checks. A separately written implementation, which does not import the main classifier, repeats direct channel comparisons and posterior identities. No theorem depends on a solver returning UNSAT or on a numerical tolerance.

These are internal checks, not external reproduction or proof-assistant formalisation. The supplied review recommended minor revisions; its author's identity and independence are not established. The audit records an early sign error in an optional polynomial formula and its correction before release; the main weighted-contrast theorem did not depend on that expression.

This is a reusable mathematical diagnostic for specified probability tables. Applying it to noisy, estimated probabilities would need an additional uncertainty analysis; no improvement to empirical forecasting has been demonstrated here.

Who should care, and why

This page is for readers interested in forecasting, information economics and exact decision criteria.

AudiencePotential useRequired caution
Information-elicitation and forecasting theoristsA finite exact test for strong substitutes, and an explicit counter-disclosure whenever it failsThe classification is for a binary event, two binary signals and quadratic scoring only
Researchers working with specified probability tablesRun the exact checker on their own rational tablesEstimated or noisy tables need a separate uncertainty analysis
Readers of the substitutes literatureA strict separator showing projective substitutes do not imply strong substitutes in this modelThe comparison concerns the exact Chen–Waggoner definition, not other variants

Why the problem matters

The result replaces an optimisation over potentially infinitely many disclosures with a finite exact decision for this model. It also exposes a trap: passing the whole-signal comparison, or a weaker projective-substitutes test, is not enough. A partial disclosure can reveal the failure.

The formulation follows Chen and Waggoner (2017); the generic disclosure-alphabet reduction follows Kong and Schoenebeck (2018), and posterior optimisation was developed in Anunrojwong et al. (2019/2020). These are credited as prior machinery. The recorded search did not locate the present weighted-contrast endpoint criterion. That is a bounded finding, not proof of historical priority, and no claim of a first-ever result is made.

How to inspect or reproduce the recorded checks

Start with AI_INDEX.md in the linked repository, then the manuscript PDF and VERIFICATION.md. The decision tools use exact rational inputs and the Python standard library only. From the package root:

python3 code/exact_checker.py examples/weak_not_strong.json   # four margins and failure disclosures
python3 code/complements.py examples/weak_not_strong.json     # complements certificates
python3 code/check_integrity.py
python3 code/run_all.py --out replay                          # full replay, under a minute

An input file lists the eight probabilities of the $(E,X,Y)$ table in lexicographic order as integers or rational strings; in the code, the paper's $A$ and $B$ are labelled X and Y. Floating-point inputs are rejected so that a tolerance cannot silently change the answer. The executable uses rational inputs; the theorem itself permits real probability tables. run_all.py copies the checks to a temporary directory, compares stable result fields with the frozen records and leaves the recorded evidence unchanged.

The most valuable next projects

Researchers can test the criterion on their own exact tables, scrutinise the proof's disclosure quantifiers, or independently replay the package. Extending the two-sided result beyond binary signals or treating uncertainty in estimated tables are distinct research projects, not completed parts of this release.

What is in the evidence package

Manuscript: complete definitions, proofs, worked examples and complements appendix. Exact checker: rational margins and direct-score counter-disclosures; binary floats are rejected. Replay tools: finite algebra, support and separately implemented checks, with explicit failure handling under optimized Python. Review response and source comparison: every adopted revision, including the reviewer-derived projective result, with prior-work boundaries. * AI index, manifest and licences: navigable claims-to-evidence map and exact payload identities.

Media

The audio briefing is provided in the header above. Download the MP3 briefing · read the transcript.

Four Tests For Binary Informational Substitutes · Watch on YouTube

Verification status

Unrefereed candidate with a written universal proof and internally replayed exact finite checks. No external peer review, independent reproduction, formal verification or historical-priority proof is claimed. The supplied review recommended minor revisions; reviewer identity and external independence are not established.

Cite

Anonymous (2026). Four tests for binary informational substitutes. Version 1.1.0-candidate. Evidence Press. 10.5281/zenodo.23238557.
BibTeX
@misc{binaryquadraticinformationalsubstitutes2026,
  title        = {Four tests for binary informational substitutes},
  author       = {Anonymous},
  year         = {2026},
  doi          = {10.5281/zenodo.23238557},
  url          = {https://doi.org/10.5281/zenodo.23238557},
  version      = {1.1.0-candidate},
  howpublished = {Zenodo},
  note         = {Unrefereed; internally replayed evidence package. Press page: https://evidencepress.org/releases/binary-quadratic-informational-substitutes/}
}

Also: cite.bib · paper.json · this page as Markdown