FlowingFlowing
← Back to all development cases
Development case · blind model judgment pending independent human review; not a formal Benchmark conclusion.
AutocompleteNo-Evidence version preferred63 / 162 · 5da68f88ecb03c7e

A Beginner's Guide to Black Hole Imaging and Associated Tests of General Relativity

Astrophysics · 2402.01290v1

FLOWING EVIDENCE BENCHMARK

How do we tell whether Evidence helps?

At the same writing position, with the same model and task, how does supplying retrieved paper passages change the first output? We compare matched versions and retain ties, unusable outputs and incomplete reviews.

SAME DRAFT · EVIDENCE ON OR OFF

01 · FIXED WRITING POSITION

MANUSCRIPT

Same writing position ▌

Matched autocomplete at the same draft position

02 · TWO MATCHED INPUTS

SHARED BY BOTH

Manuscript context, model, task and prompt

A · With Evidence

Retrieved paper passages supplied

B · Without Evidence

No retrieved passages supplied

LLM

Same model and version

A → first output

B → first output

Blind judge agent

First outputs are anonymized as X and Y

Continuation review: accuracy, fit to the writing task and usability
Returns: X preferred / tie / Y preferred / both unusable

Order check: X / Y → Y / X

HOW DOES BLIND JUDGING WORK?

① Anonymize both outputs
The judge sees the same draft and both first outputs without knowing which received Evidence.

② Compare and swap order
The judge applies task-specific criteria in X/Y and then Y/X order.

③ Review disagreements
A third pass resolves disagreements. Incomplete reviews remain in the denominator.

How are autocomplete positions stratified?

Before seeing generation outcomes, we check whether a retrieved passage contains a specific proposition that directly supports the next writing move. Those positions appear in the left opportunity group; the rest are ordinary positions on the right. We select a balanced sample from admitted papers in each field. The 50/50 split is experimental, not a measure of how often either type occurs in writing.

AUTOCOMPLETE · 110

Biology, statistics and astrophysics

55 positions on each side; statistics uses the ten-paper rerun.

AUTOCOMPLETE · 52

Psychology and climate science

26 positions on each side; psychology includes nine papers and climate science four.

How are the table percentages calculated?

Across five fields there are 81 source-opportunity positions. The original blind review preferred Evidence in 51; a task check moved one empty Evidence continuation to both unusable, leaving 50 in public counts. In the ordinary group, another pair of empty outputs moved from no winner to both unusable. Original verdicts remain visible on case pages.

50Evidence version preferred
÷
81All positions in this group
=
62%Evidence preference in this group

Source contribution is a separate review: 16 of 21 Evidence wins entered into source review directly used retrieved papers; another 29 wins await review.

These are development-stage model judgments pending independent human review. They are not formal Benchmark conclusions and do not, on their own, establish causality.

Manuscript writing position

Text excerpt · not a PDF page

Research manuscript · excerpt

A Beginner's Guide to Black Hole Imaging and Associated Tests of General Relativity

0.1 Introduction

…ic—one that explains the basic theory behind black hole imaging and provides a first look at the most promising ideas for testing GR with VLBI observations, while remaining mindful of their inherent (and often-overlooked) potential pitfalls. Our goal in this Chapter is precisely to provide such a pedagogical tool. As a result, this text is aimed primarily at students (or to novices in the area), and does not purport to give a comprehensive review of the field. Rather, this Chapter is our personal tour through the subject—one necessarily biased by our own experiences, interests, and expertise.

Tests of GR via electromagnetic observations of black holes pose a number of challenges. Perhaps the greatest difficulty arises from the need to disentangle the effects produced by the black hole’s gravity from those caused by specific features of the particular source emitting the photons that we eventually detect. At first glance, one would expect that significant uncertainties in the detailed properties of the astrophysical source would always be present and dwarf any observable deviation from GR, rendering the very idea of a test of Einstein’s theory in this context moot.

↓ Next paragraph was generated here

The highlighted text is near the writing position. This is a layout of the evaluation input, not a PDF screenshot.

A matched comparison at the same writing positionJudgment: No-Evidence version preferred

Both arms use the same model, task and manuscript context. Only the retrieved passages differ. The first recorded output from each arm is shown below.

Evidence

Same model and position · retrieved passages supplied · first output

Not preferred in this comparison

The product produced no insertable continuation text.

The model generated prose, but the product admission rule blocked it as needing evidence. That raw prose has not been independently reviewed here.

Passages supplied to the Evidence version

Experimental Relativity with Accretion Disk Observations ↗

In this work we study our current ability to constrain and detect deviations from General Relativity using the accretion disk spectrum of stellar-mass black holes in binary systems. Our analysis combines relativistic ray-tracing and Markov-Chain Monte-Carlo sampling techniques to determine how well…

Read full passage excerpt

In this work we study our current ability to constrain and detect deviations from General Relativity using the accretion disk spectrum of stellar-mass black holes in binary systems. Our analysis combines relativistic ray-tracing and Markov-Chain Monte-Carlo sampling techniques to determine how well such tests of General Relativity can be carried out in practice. We show that even when a very simple astrophysical model for the accretion disk is assumed a priori , the uncertainties and covariances between the parameters of the model and the parameters that control the deformation from General Relativity make any test of General Relativity very challenging with accretion disk spectrum observations.

No Evidence

Same model and position · no retrieved passages · first output

Preferred by blind review

This concern is well founded: the observed image or spectrum is shaped jointly by the spacetime metric and by the poorly constrained microphysics of the emitting plasma, so that a deviation attributed to gravity may instead reflect an incorrect model of the source. The central question for any electromagnetic test of general relativity is therefore whether there exist observables that depend on the spacetime in a way that is sufficiently insensitive to these astrophysical uncertainties.

Judgment and task check

The blinded model review preferred the no-Evidence version. Its original rationale appears below.

Read the original model judgment

In this rationale, Output A = No Evidence; Output B = Evidence.

Output A provides a coherent, well-supported continuation that directly addresses the challenge described in the draft—disentangling gravitational effects from astrophysical source uncertainties. The paragraph flows naturally from the draft's concern about source uncertainties dwarfing GR deviations, using language ('observed image or spectrum,' 'spacetime metric,' 'emitting plasma') consistent with the VLBI/black hole imaging context. The central question posed at the end sets up the pedagogical exploration promised in the introduction. Output B returns an empty string with 'needs_evidence' action, which violates the task requirement to produce a completion paragraph; the draft clearly needs content here, and the source provided offers relevant factual context about the challenge of testing GR with accretion disk observations, making a deliberate empty response inappropriate.

The passages are input material; their presence does not mean an output used them or that they support every claim. Draft and source passages are excerpted; no simulated scores or PDF appear here.