How we grade research

Sentiosonics summarises published research on sound, brainwave entrainment and sound therapies, and gives each study an evidence grade. This page explains how those grades are made, what they can and cannot tell you, and how to tell us when we have something wrong.

What a grade is

A grade is our judgement of how much weight a single paper can bear. It is about the paper’s methods and results, not about whether a technique works in general and not about the claims other people make on the topic. A strong paper can report a modest effect; a weak paper can report a dramatic one. The grade tells you how far to trust what the paper found.

The scale, from most to least support:

Strong evidence: replicated, or supported by a meta-analysis.
Moderate evidence: controlled trials, not yet replicated.
Mixed evidence: the paper’s findings are partial, or the studies it sits among disagree.
Weak evidence: small samples, no control group, or a design that cannot separate the effect from expectation.
Not supported: the claim was tested and the paper did not find the effect.
Untested: claimed, but not yet studied. A summary of a real study is never graded untested.

Studies we have not yet scored say Not yet graded, and should be read as a pointer to the paper, not a verdict on it.

The rubric

Every graded study is scored on eight criteria, each from 1 to 5, with a one-line note explaining the score. The scores and notes are printed on the study’s page under “How we graded this study”.

Criterion	What earns a 5	What earns a 1
Study design	Randomised controlled trial, or a systematic review with quality assessment	Case report, commentary or theoretical paper
Sample size	100 or more participants (reviews: 20 or more studies)	Fewer than 10, or not reported
Control condition	An active sham matched for sound and attention	No control
Blinding	Double-blind	Unblinded, with a subjective outcome
Outcome measure	Objective and validated (EEG, cortisol, normed task performance)	Anecdote, or unclear
Replication	Independently replicated with consistent results	Replication attempted and failed
Conflicts of interest	Independent funding, none declared	Funded by the maker of the product
Peer review and venue	Peer-reviewed, indexed journal	Not peer-reviewed, or the paper could not be located

Study design and replication count double, because they matter most for whether a result is real:

(2 × design + 2 × replication + sample + control + blinding + outcome + conflicts + peer review) ÷ 10

That gives a score between 1 and 5: 4.2 or above is strong, 3.4 to 4.19 is moderate, and anything lower is weak. We score the method, not the result. A trial can find a big effect and still score low, because the grade answers “how much should this change what you believe?” rather than “what did the authors conclude?”

Two rules override the score, because they come from what the paper found rather than how it was run. If the paper tested the claim and found no effect, the grade is not supported, however well designed the study. If its findings are partial, or the literature around it disagrees, the grade is mixed.

Where something is unknown, it is scored as unknown, never assumed in the paper’s favour. A paper that does not state its funding scores 3 on conflicts of interest, not 5.

Two worked examples

Near the top. Xiong and colleagues (2025) pooled 15 randomised trials of binaural beats around surgery, covering just over a thousand patients. Design 5 (a systematic review of RCTs), sample 5, control 4 (compared against blank audio and against non-beat audio), blinding 3 (varied across the trials), outcome 4 (standardised anxiety scales, pain, blood pressure), replication 4 (independent trials pooled, but with very high heterogeneity), conflicts 3 (not stated in the abstract), peer review 5. Weighted: 4.2, the floor of strong, and worth reading with that in mind.

Near the bottom. Yang and colleagues (2025) played theta, alpha and beta binaural beats to 65 students and measured blood pressure and heart rate before and after. Design 3 (everyone heard all three, before-and-after), sample 4, control 2 (no silence or plain-audio comparison), blinding 1, outcome 5 (objective cardiovascular measures), replication 3, conflicts 3 (not stated in the abstract), peer review 5. Weighted: 3.2, weak. A well-measured study can still score low when there is nothing to compare against.

Both come from the same eight questions. That is the point of a rubric.

What “graded from” means

Each grade says what it was based on:

Full text read: we read the methods and results in the paper itself, usually because it is open access.
Abstract and metadata: the paper is paywalled, so the grade rests on the abstract and the publication record. Details that only appear in the full text, such as how participants were blinded or who funded the work, may be scored as unknown.
Metadata only: we could find the publication record but not the abstract. These grades are the least certain.

Most grades on the site today are based on the abstract and metadata. Treat them as provisional; a grade can change when someone reads the full paper. This confidence is about our reading, not the paper: a strong grade read from the abstract is a well-built study we have only seen the summary of.

Who does the grading

Grades are assigned by an AI research agent following this rubric. The editor reviews them.

The agent finds the paper behind each summary by searching OpenAlex, Semantic Scholar, Crossref and PubMed. We only publish a summary when we have located the paper itself and the summary describes that paper. The agent checks that the authors, year and subject match, records the DOI and a link to the paper, scores the eight criteria with a note for each, and corrects the summary where it misstates what the paper did. It is instructed to grade the paper as published, not the summary and not the popular claim about the topic.

The editor reviews the agent’s work: the match to the paper, the scores and notes, and the summary text, and can change any grade. Each study page says whether it was graded by the research agent or by the editor, and when; once the editor has reviewed a grade, the page says so and gives the date, and later automated passes do not overwrite it.

New studies arrive graded by the agent. Each study page says whether the editor has reviewed its grade yet, and when. We say this rather than hide it, because a rubric is only worth anything if the gaps in how it has been applied are visible too.

We use an agent because checking every citation by hand is slow, and because it applies the same rubric the same way each time. It can also be wrong: it can misread an abstract, match the wrong paper, or score a criterion too generously or too harshly. That is why the scores and notes are published in full, and why we want to hear from you when they look wrong.

News verdicts

When we check a news story rather than a paper, the page carries a verdict instead of a grade:

Accurate: the story says what the paper found.
Fair: some looseness in the framing, nothing that misleads a reader.
Overstated: the finding is real; the headline is bigger than the finding.
Misleading: the story would leave a reader believing something the paper does not support.

The verdict is about the story. It is not a judgement on the researchers, who usually had nothing to do with the headline.

Corrections

If you think a grade, a summary or a link is wrong, please tell us. Corrections and counter-evidence are especially welcome. It helps if you include:

the page it concerns;
what you think is wrong;
a link or DOI for the source that shows it.

We read every correction. When we get something wrong, the page gets a dated correction note saying what was wrong and what it now says; we do not edit a mistake out silently. If you think a grade is wrong, the most useful message names the criterion and why, because the scorecard is right there on the page.

What this site is not

Sentiosonics is an information site, not medical advice. A grade describes the strength of the published evidence for a specific outcome in a specific study population. It is not a recommendation to use, or avoid, any product or therapy.

How we grade research, and how to correct us