FlowingFlowing
← Back to all development cases
Development case · blind model judgment pending independent human review; not a formal Benchmark conclusion.
AutocompleteNo-Evidence version preferred78 / 162 · 066052b7fc0775ac

Double-diffusive processes in stellar astrophysics

Astrophysics · 2103.08072v1

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

Double-diffusive processes in stellar astrophysics

4 Oscillatory double-diffusive convection and layered convection > 4.6 Conclusions for now

…, Nu T ≫ 1 {\rm Nu}_{C},{\rm Nu}_{T}\gg 1 ). This is not really consistent with Figure 19 b, where it appears that γ tot − 1 \gamma_{\rm tot}^{-1} may be decreasing instead of increasing with R 0 − 1 R_{0}^{-1} . This discrepancy most probably comes from the fact that neither ( 79 ) nor ( 80 ) correctly account for the weak dependence of Nu T {\rm Nu}_{T} and Nu C {\rm Nu}_{C} (and in turn D semi D_{\rm semi} ) on R 0 − 1 R_{0}^{-1} . This problem can only be resolved by running simulations for a wider range of R 0 − 1 R_{0}^{-1} , which is computationally very challenging as discussed above.

…F_{T,C}\sim H_{L}^{4/3} ). In all the simulations performed to date, the layers were ultimately seen to merge until a single one remains in the domain. This poses two problems. First, it has prevented us from proposing a model for what the layer height may be in a star, since it is not clear what mechanism (if any) eventually stops the merger process. Secondly, these results seem to challenge the original assumption of an equilibrium staircase, suggesting instead that layered convection may be a short temporary phase in the evolution of a star, whose duration depends on the layer merger rate.

↓ 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 layer-merger problem also complicates the interpretation of the earlier scaling relations, since a staircase that does not reach a steady state cannot be characterized by a single layer height.

Passages supplied to the Evidence version

A scaling relation between merger rate of galaxies and their close pair count ↗

The use of virial masses, instead of the stellar mass, is motivated by the fact that the dynamical friction time scale is mainly determined by the dark matter surrounding central and satellite galaxies. This fact can also minimize the error induced by uncertainties in modeling star formation in the…

Read full passage excerpt

The use of virial masses, instead of the stellar mass, is motivated by the fact that the dynamical friction time scale is mainly determined by the dark matter surrounding central and satellite galaxies. This fact can also minimize the error induced by uncertainties in modeling star formation in the simulations. Since the virial mass can be read from the well-established relation between the virial masses and the stellar masses in observation, our scaling relation can be easily applied to observations to obtain the merger rate and merger time scale.

No Evidence

Same model and position · no retrieved passages · first output

Preferred by blind review

The merger of layers in every simulation to date therefore leaves the equilibrium staircase an open question, and the layer height in a star cannot yet be predicted because the mechanism that would halt the merger process has not been identified.

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 = Evidence; Output B = No Evidence.

Both outputs are usable and grounded in the draft. Output B more directly and comprehensively addresses both problems explicitly stated in the draft: (1) inability to propose a model for layer height in a star due to unknown stopping mechanism for merger, and (2) the challenge to the equilibrium staircase assumption. Output B captures both the layer height prediction problem and the equilibrium staircase question with precise causal language ('because the mechanism that would halt the merger process has not been identified'). Output A introduces 'scaling relations' and 'single layer height' characterization, which slightly narrows and shifts the focus—the draft discusses 'model for what the layer height may be in a star' and 'equilibrium staircase' as broader conceptual problems, not specifically about characterizing by 'single layer height.' Output B preserves the full scope of both stated problems with less interpretive filtering.

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.