Evidence Library

The data:
ERP CORE (Kappenman et al., 2021, Scientific Data), the canonical open EEG resource: 40 participants across six paradigms (~133,000 single trials) covering seven classic components — N170, MMN, N2pc, N400, P3, LRP, ERN. Raw BIDS files were downloaded from the NEMAR mirror (nm000132 v1.1.1), SHA-256 verified, and processed with MNE-Python (automatic ICA, global 200 µV artifact ceiling, pre-specified measurement windows per component). All numbers on the dashboard are the locked outputs of that pipeline.

What you'd see, step by step:

1. Question — what can a transparent open-source pipeline reproduce from a public EEG dataset, and how well?
2. Data — download and verify the ERP CORE raw files (40 participants, six paradigms, ~133,000 trials).
3. Cleaning — re-reference, filter, automatic ICA ocular correction, artifact rejection.
4. Design — pre-specify each component's measurement window, channel ROI, and statistical test before analysis.
5. Results — grand-average waveforms and per-component statistics revealed step by step.
6. Validation — the N400 cross-check against the published anchor, deviation documentation, and honesty labels.
7. Conclusion — five of seven effects match the published benchmark clearly; LRP and ERN show the right direction but stay sub-threshold, and that is reported honestly.

Try or inspect it:
Follow the seven-step walkthrough, compare each step with the previous one, and inspect or download the supporting tables, trial-level CSVs, and scripts.

Limitations:
Artifact rejection differs from the original paper (automatic ICA + global 200 µV ceiling vs per-subject tailored thresholds), so participant counts differ slightly from Kappenman et al. Table 1 (N=34–39). LRP and ERN did not clear the pre-registered effect-size gate (amber, reported honestly — not a pipeline failure). The N400 cross-check matches the published anchor in sign and magnitude (Δdz=0.022) but falls outside a very tight 0.01 tolerance, which is documented rather than suppressed. No item random effects.

The data:
The dataset comes from a 2024 PLOS ONE study by Chen, Zhang, Zhou, Chan and colleagues, and contains acoustic speech measurements from 46 Cantonese-speaking children (23 with autism, 23 typically developing) across nearly 10,000 recorded sentences. The data and analysis code were deposited publicly by the original authors, so everything shown is real, peer-reviewed research material. Because the pipeline re-runs on the deposited snapshot, results are stable — but small differences from the published numbers are expected and are explained honestly during the demo.

What you'd see, step by step:

1. The deposited data file loads: 55,000 rows of syllable-level acoustic measurements appear on screen.
2. The pipeline cleans the data automatically, flagging and removing a small number of incomplete rows.
3. Statistical models run for pitch, duration, and intensity — progress updates show in real time.
4. Results appear: children with ASD produce consistently smaller pitch range on focused syllables than typical peers.
5. A summary panel shows 20 out of 20 key statistical tests reproduced, matching the paper's conclusions.
6. Side-by-side numbers highlight where replicated values are close to published figures and where minor differences exist, with a plain-English explanation of why.

Try or inspect it:
Inspect the group-comparison table and prosody plots for the bundled study data. Contact the team for another run under an agreed data-handling arrangement.

Limitations:
The study is small by clinical standards (23 children per group), so effect sizes should be treated as indicative rather than definitive. The pipeline reproduces all directions of difference and statistical significance, but the raw test-statistic numbers differ from the paper's tables due to a difference in how model comparisons were sequenced — the conclusions are identical either way. Post-hoc pairwise comparisons are not re-run here, so fine-grained contrast tables from the original paper are taken on trust.

The data:
This demo analyzes 153,566 of 230,412 word observations from 84 native English readers in the public Provo Corpus. Five eye-movement measures match the authors' published values at Pearson r = 1.0000 (within 1 ms); predictability effects are reproduced for first-fixation duration (15.4 ms, t(83) = 11.37, p = 1.3e-18, dz = 1.24) and gaze duration (43.1 ms, t = 15.24, p = 8.9e-26, dz = 1.66), alongside word-length effects. Limitation: only 153,566 of 230,412 word observations are analyzed, so this is not a full-observation reanalysis.

What you'd see, step by step:

1. Start with the denominator decision: 153,566 complete observations are analysed from 230,412 public word observations.
2. The pipeline computes five eye-movement measures per word (first-fixation duration, gaze duration, go-past, total reading time, regression-in) using the published definitions.
3. Two contrasts run automatically: predictability (high vs zero cloze predictability) and word length (short ≤ 4 vs long ≥ 7 letters).
4. A validation panel compares the pipeline's five measures against the authors' published values — Pearson r = 1.0000, exact within 1 ms.
5. Results show robust predictability effects (first-fixation Δ15.4 ms, t(83) = 11.37, p = 1.3e-18; gaze Δ43.1 ms, t = 15.24, p = 8.9e-26) and word-length effects, with per-participant charts.
6. An Export tab lets you download the full analysis as PDF, DOCX, CSV or Markdown — validation table included.

Try or inspect it:
Open https://demo.patrickchu.net/eetracking-real/ and run the pipeline live — every figure, table and statistic can also be exported.

Limitations:
Only 153,566 of 230,412 word observations carry complete data across all five measures and are analyzed (the rest are excluded per the published definitions, e.g. first-pass skips); this is not a full-observation reanalysis. Word-frequency effects are not included because the official Provo files publish no word-frequency column — only predictability and word-length effects are claimed.

The data:
The dataset is 1,000 academic and general-interest texts — 500 pieces of translated English and 500 pieces of native English — drawn from two established public corpora and spanning four genres (academic, news, prose, fiction). The data and the authors' original analysis code are openly archived by Chou, Li, and Liu (PLOS ONE, 2023) and can be downloaded by anyone. Because the dataset is fixed and archived, results are stable; the numbers you see in the demo will be the same every time.

What you'd see, step by step:

1. Start with the comparison design: 1,000 texts across two provenance groups, four genres and six language features.
2. The pipeline runs the same statistical tests as the original paper — results appear within seconds.
3. A comparison table appears: 20 results match the paper exactly, 4 are very close with reasons shown.
4. One flagged anomaly is highlighted: a number in the published paper that is internally inconsistent.
5. A plain-English summary panel shows which language features differ most between translated and native English.

Try or inspect it:
Inspect the displayed results table and compare translated with native English by genre. For a run on other material, contact the team to agree data handling first.

Limitations:
The corpus is real but modest in size (1,000 texts, four genres), so findings describe these specific collections, not all English writing. Four of the 24 replicated results are close but not exact — two because the original paper used different statistical settings than its own archived code, and two because of what appears to be a misprint in the published table. Effect sizes throughout are small, meaning the translated/native differences are real but not large in practical terms. Any live re-run will produce the same numbers as the demo, since the data is static.

The data:
The demo uses 11 real, open-access, top-cited papers on ChatGPT in higher education (2023–2026), fetched from OpenAlex and pinned for reproducibility. The sample is small and selective (a top-cited sample, not an exhaustive search), and because live data can drift, results are reported as ranges rather than single numbers.

What you'd see, step by step:

1. First assess scope: 11 pinned, top-cited abstracts support an extraction tutorial, not an exhaustive literature review.
2. Watch as each abstract is read and structured into a matrix row.
3. See the design column cross-checked against a keyword baseline.
4. Themes and gaps appear automatically from the matrix.
5. A one-page summary shows key numbers and limitations.
6. You can click any cell to verify it against the abstract.

Try or inspect it:
Inspect the abstract-extraction matrix and compare two bundled papers. For a run on another literature set, contact the team first under an agreed data-handling arrangement.

Limitations:
Extraction is from abstracts only, not full texts, so some details are thinner. The LLM is stochastic but pinned to a cache for reproducibility; observed agreement with the baseline ranges from 0.76 to 0.88. The corpus is a small, top-cited sample, not an exhaustive review.

The data:
The data come from a real published experiment (Ma, Han & Li, 2022, PLOS ONE) in which participants read translated and paraphrased texts while their eye movements were recorded; the original spreadsheets and analysis scripts are freely available online (OSF). Because the dataset is fixed and not live, results are stable and directly comparable to what appears in the published paper. There is no synthetic or simulated data involved.

What you'd see, step by step:

1. First inspect what can be checked: participant-level eye-tracking measures are available, while one ordinal model remains outside the rerun.
2. The pipeline cleans the data exactly as the authors did: removing extreme readings, transforming skewed values.
3. Statistical models run automatically, testing whether translation and paraphrase produce different reading patterns.
4. A results table appears comparing the pipeline's numbers to the published paper's numbers side by side.
5. 20 out of 21 statistical verdicts match the paper exactly; one borderline result is flagged and explained.
6. A summary panel highlights the headline finding: readers' eye behaviour genuinely differs between translation and paraphrase.

Try or inspect it:
Inspect the displayed results table, choose one of the 21 eye-tracking results, and compare the published and checked values. For another dataset, contact the team to agree data handling first.

Limitations:
The dataset is small (one published experiment, a limited number of participants and texts), so findings should not be generalised beyond the original study's scope. One result out of 21 does not match the published value — it sits right on the borderline of significance and is sensitive to minor differences between the original software and the replication software; this is flagged honestly rather than glossed over. An ordinal model used in the original paper for translation-difficulty ratings could not be re-run and remains a known gap.

The data:
The dataset comes from two experiments by Ji, Liao & Hayward (BMC Psychology, 2025) and is publicly archived on OSF alongside the original analysis scripts. Experiment 1 tested 88 young Hong Kong participants estimating the ages of White faces showing neutral, happy, or angry expressions; Experiment 2 added older participants and Asian face stimuli. The data are real and stable (a fixed, cleaned deposit), so results are consistent across runs rather than shifting with live feeds.

What you'd see, step by step:

1. First compare the two experimental designs: face expression in Experiment 1 and participant-age modulation in Experiment 2.
2. Participant-level age estimates are calculated automatically for each expression condition.
3. Statistical tests run and reproduce the paper's key comparisons within seconds.
4. A results summary appears: 20 of the published statistics reproduced, directions and significance all matching.
5. Side-by-side numbers from the paper and the pipeline are displayed for inspection.
6. A plain-English verdict flags any differences (here: only a minor effect-size notation convention).

Try or inspect it:
Ask the demo to show only Experiment 2 results — type 'exp2' when prompted — to see whether the older-participant effect still appears in isolation.

Limitations:
Both experiments used White or Asian face stimuli rated by Hong Kong participants, so findings may not generalise broadly. Sample sizes are modest (88 young adults in Exp 1; roughly 72 across age groups in Exp 2). One effect-size number differs from the paper because the paper and this pipeline follow different but equally valid conventions for calculating it — the underlying finding is the same. The pipeline reproduces the core paired comparisons and the participant-age modulation, but not every sub-model from the original R script.

The data:
The demo uses 25 real student essays drawn from a large public dataset of grade 7–8 persuasive writing, each independently scored by two trained human raters on a 1–6 scale. The data is real, not synthetic, but it comes from a US classroom context — results on Hong Kong student writing or different rubrics would need separate validation before any conclusions are drawn.

What you'd see, step by step:

1. Twenty-five public, anonymised US middle-school essays load with two existing human scores for a small evaluation exercise—not operational grading.
2. The AI reads every essay against the official rubric and assigns its own score.
3. A results table appears: AI matched human scores exactly 68% of the time.
4. The key comparison appears: human raters agreed with each other at a similar rate, setting the honest benchmark.
5. Click any essay row to see the AI's full written feedback — three specific, rubric-linked comments per essay.
6. A summary panel highlights where AI and human scores diverged most, flagging essays for closer review.

Try or inspect it:
Select an included essay, compare both human ratings with the AI output, and critique the feedback against the displayed rubric. Do not upload student work.

Limitations:
The 25-essay sample is small, so treat the agreement figures as indicative ranges rather than fixed numbers — a larger local sample would tighten them considerably. All essays here are from US middle-school students writing in English; the AI's scoring reliability on Hong Kong student writing, or against a different institutional rubric, is unknown and would need its own validation study (typically 50–100 essays with two local raters). The AI also produces only a single overall score in this demo; separate scores for content, organisation, and language conventions are possible but not shown here.

The data:
14 Chinese-language government press-release sentences, hand-authored to include a mix of everyday official terms (e.g. 教育局, 立法會) and rarer policy-specific terms (e.g. 簡約公屋 → Light Public Housing, 明日大嶼 → Lantau Tomorrow Vision). The data is small and synthetic — it is a mechanism demo, not a full study. Because the underlying AI model can be updated at any time, exact numbers may shift slightly between runs; the demo quotes ranges from repeated runs, not single best scores.

What you'd see, step by step:

1. The test begins with a concrete decision: does a 22-term glossary increase correct terminology use across 14 synthetic sentences?
2. Each sentence is translated twice: once with no glossary, once with the glossary injected.
3. Results appear side-by-side; correct and incorrect term choices are highlighted per sentence.
4. A summary panel shows overall terminology accuracy: 80–90% without the glossary, 100% with it.
5. You can scroll to sentences 13–14 to see where the glossary makes the clearest difference — rare policy terms.

Try or inspect it:
Select another bundled government-style sentence and compare the glossary-assisted and unassisted outputs.

Limitations:
Only 14 sentences — enough to demonstrate the mechanism, not to publish. A real study would need 100+ sentences with professional translators setting the gold standard. The quality scores (fluency, adequacy) are generated by an AI judge, not a human, and both translation versions are already fluent, so those scores are close and should not be over-interpreted. Results are ranges across runs, not guaranteed single figures.

The data:
Sixteen short research abstracts were written by hand for this demo: eight that genuinely fit the review question (technology-enhanced feedback on school students' writing) and eight that do not, covering common exclusion reasons such as wrong age group, wrong subject, or no real study data. The dataset is synthetic and intentionally clean; real-world records are messier, and a proper validation would need 100+ abstracts checked by two independent human raters. Because this is a fixed pilot set, the headline numbers are illustrative ranges rather than a certified benchmark.

What you'd see, step by step:

1. The review question, PICOS criteria, and 16 intentionally clean synthetic abstracts load for a fixed pilot—not a review benchmark.
2. Sixteen abstracts appear one by one; the AI reads each and gives an Include or Exclude verdict.
3. Every verdict is accompanied by a plain-English reason you can read and challenge.
4. A PRISMA-style flow diagram builds live, showing how many records survived each stage.
5. A summary panel shows the AI agreed with the gold-standard labels on 15 of 16 abstracts.
6. Click any row to compare the AI's reason with the human gold label side by side.

Try or inspect it:
Choose one of the supplied synthetic abstracts, inspect the provisional include/exclude decision, and test which PICOS criterion drove it.

Limitations:
This is a 16-abstract pilot built on hand-written, unusually tidy records — not a real-world benchmark. Gold labels come from a single rater, not the two independent reviewers a real review requires. The one missed study (a meta-analysis mis-labelled by the demo) shows the AI is not error-free; recall figures in production will likely be lower and should be reported as a range. A live deployment would also need deduplication and full-text screening stages not shown here.

The data:
The demo uses 11 public-domain items from the GAD-7 and PHQ-9 scales. It compares three translation drafts (plain, glossary-guided, and colloquial Cantonese) against the official Traditional Chinese versions. The data is synthetic in the sense that the drafts are generated by an AI, and results may vary slightly with each run.

What you'd see, step by step:

1. You see the original English items and three Chinese drafts.
2. The pipeline runs forward translation and blind back-translation.
3. An AI judge scores each draft for meaning preservation.
4. A flagged item is automatically refined and re-checked.
5. A deliberately wrong translation is caught and corrected.
6. Final results show which draft best matches the published version.

Try or inspect it:
Compare the plain, glossary-guided, and colloquial drafts for any of the 11 bundled items. Contact the team for another run under an agreed data-handling arrangement.

Limitations:
The AI acts as both translator and judge, so this validates the workflow, not a finished instrument. Results are ranges, not exact numbers, because the AI is stochastic. The comparison with published Chinese is indicative, as regional variants exist.

The data:
A publicly archived survey of 3,032 adults in Hong Kong, Singapore, and the United States, collected around the WHO's March 2020 pandemic declaration (published data deposit by Prof Catherine Wing-Man Yeung, CUHK Business School, and co-authors). The dataset is real, publicly available, and fixed — it will not change between demos, so results are stable. Because the deposit does not include the date each person responded, the before/after comparison from the paper cannot be fully re-run; results for that part are therefore reported as ranges, not single figures.

What you'd see, step by step:

1. The pipeline loads the fixed public deposit and first checks what variables are—and are not—available for reproduction.
2. Summary statistics appear on screen: 3,032 participants confirmed, split across three regions.
3. Worry scores and precautionary-action rates are calculated and displayed alongside the paper's own numbers.
4. Correlation results appear for all three regions — each matches the published figures exactly.
5. A clear flag highlights the one gap: the before/after split cannot be verified because response dates are absent from the deposit.
6. A summary panel shows what reproduced (20 of 21 checkable items) and what the missing piece would take to fix.

Try or inspect it:
Inspect the deposit-coverage table to see which analyses can and cannot be reproduced. Contact the team for another run under an agreed data-handling arrangement.

Limitations:
The dataset is real but the deposit is incomplete: response dates were not archived, so the paper's headline before/after ANOVAs cannot be independently verified. Worry scores and action rates are therefore shown as overall figures that fall between the published before and after values, not as a true replication of the time-split analysis. All other results — correlations, sample counts, action rates — reproduce cleanly. The fix is straightforward: if the authors share even a simple before/after flag, the remaining checks run in minutes.

The data:
The dataset comes from a published study by Dr Yen Na Yum (EdUHK) and co-author Wang (2025), deposited publicly on OSF. It covers 88 Chinese-English university students learning 16 technical English words under different reading conditions — the data and original analysis scripts are both openly available. Because the dataset is fixed and fully public, results are stable; the numbers you see today will be the same tomorrow.

What you'd see, step by step:

1. First verify the study scope: 88 students learning 16 technical words under the deposited experimental conditions.
2. Three statistical models run automatically, one for each outcome: spelling accuracy, spelling errors, and meaning accuracy.
3. A comparison table appears showing the published findings alongside the pipeline's independent results.
4. Green 'reproduced' labels highlight where the two sets of results agree on direction and significance.
5. A plain-English summary box names the three headline findings that held up across both analyses.

Try or inspect it:
Inspect the word-length comparison table and trace an effect to the published OSF data. Contact the team for a run on other material under an agreed data-handling arrangement.

Limitations:
The study is relatively small (88 students, 16 words), so effect sizes should be treated as indicative rather than definitive. The replication uses a different but comparable statistical method to the original (the two approaches measure subtly different things), which is why some coefficient numbers differ even when the conclusions agree — sign and significance, not exact magnitude, are the honest replication target. Four of the 19 comparisons are flagged 'close' rather than 'reproduced' because one significance threshold shifted; the headline findings are unaffected.

The data:
Two small experimental datasets (roughly 150–160 participants each) from Prof Sara Kim's published PLOS ONE study, downloaded directly from the open-science repository OSF. The data are real experimental responses, not synthetic, and are frozen at the published version — so results are stable and will look the same every time the demo runs.

What you'd see, step by step:

1. Begin by comparing the two small experimental samples and the distinct role each plays in the reported mediation claim.
2. Participant exclusions run automatically, recovering the exact sample sizes from the paper.
3. Group-comparison tests run; the F-values and group means appear on screen.
4. The mediation analysis runs, showing how emojis reduce a sense of being objectified, lifting creativity.
5. A moderated-mediation test checks whether this pathway is stronger for relationship-oriented team members.
6. A summary panel flags each result green (reproduced) or red (diverged) against the published numbers.

Try or inspect it:
Re-run the mediation step with a different random seed, then compare the confidence interval and conclusion with the bundled result.

Limitations:
Both studies are small (under 160 participants each) and used student samples in a controlled online setting, so effect sizes may not transfer directly to real workplaces. The confidence intervals reported are ranges, not single definitive numbers — a slightly different resampling method shifts the interval endpoints a little, though the direction and significance of every finding remain the same. This demo shows statistical reproducibility only; it cannot speak to whether the original experiment would replicate in a fresh sample.

The data:
The dataset comes from the paper's own openly shared archive (OSF): error-rate recordings from 53 student interpreters, scored across four error types (conceptual, lexical, syntactic, phonological) alongside each student's language proficiency, working-memory span, and anxiety score. The data are real but small — 53 participants — so findings should be read as exploratory rather than definitive. Because the dataset is a fixed, archived file (not a live feed), results are stable and fully reproducible each time the demo runs.

What you'd see, step by step:

1. First assess the evidence base: 53 student interpreters measured across four error types and three candidate predictors.
2. A correlation and collinearity check runs automatically; a summary confirms no problematic overlap between predictors.
3. Five regression models run (one per error type plus total); a results panel appears showing which predictors matter for each error type.
4. A relative-importance chart appears, showing which of the three factors — proficiency, working memory, anxiety — drives each error type most.
5. A final comparison table lights up: all 15 published findings from the paper's Table 5 are marked 'reproduced', confirming the pipeline matches the original.

Try or inspect it:
Inspect the regression results table and point to an estimate whose uncertainty you want explained. For another dataset, contact the team first under an agreed data-handling arrangement.

Limitations:
The sample is small (53 students), so individual regression estimates carry meaningful uncertainty — treat the direction of effects as the headline finding, not the precise numbers. The relative-importance scores (e.g. proficiency contributes roughly 18% of explained variance in total errors) are summaries of a pattern, not exact quantities to over-interpret. The replication re-expresses the authors' original R analysis in Python and matches it closely, but cannot go beyond what the original study's design allows.

The data:
The study asked 540 Hong Kong participants to rate a single recorded speaker — in either Cantonese, English, or Mandarin — on personality and credibility traits using 7-point scales. The dataset is publicly available on OSF (open-access, with a PLOS data-quality badge) and is fixed, so results are stable and reproducible to the decimal point. One caveat: the deposited file appears to be missing some of the original survey items, which affects one secondary analysis (reliability scores) but not the main findings.

What you'd see, step by step:

1. Begin with the central comparison: how 540 Hong Kong participants rated one speaker across three language conditions.
2. The pipeline runs statistical tests comparing all three language groups across twelve personality traits.
3. A results table appears: seven traits show clear language differences, five show none.
4. Headline findings highlight: English leads on perceived intelligence, competence, and prestige; Mandarin edges ahead on employability; Cantonese trails on both.
5. A side-by-side panel shows published numbers versus replicated numbers — they match exactly for all main results.
6. A flagged warning draws attention to the one mismatch: reliability scores cannot be fully reproduced due to missing items in the public data deposit.

Try or inspect it:
Inspect the displayed results table, choose a language condition, and ask what a group difference means. For a run on other material, contact the team to agree data handling first.

Limitations:
The original study used only 540 participants in one city at one point in time, so findings should not be generalised beyond Hong Kong or this period. The deposited data is missing some survey items, meaning the authors' internal reliability scores (Cronbach's α) cannot be reproduced — a genuine transparency gap in the public archive, though it does not affect the main statistical conclusions. Participant demographics were not fully balanced across language groups, a limitation the original authors themselves acknowledge.

The data:
The demo uses the publicly archived datasets from Prof Amy N. Dalton's PLOS ONE paper (Liu, Dalton & Lee, 2021), deposited openly on OSF. The data are real survey responses collected during COVID-19 — two US samples, one US experiment, and one Hong Kong sample — totalling around 1,200 participants across four studies. Because these are fixed, archived files (not live feeds), results are stable and fully reproducible each time the demo runs.

What you'd see, step by step:

1. First compare the four samples and ask whether the self-concept association is consistent across settings.
2. It applies the original paper's own filters: attention checks and eligibility screens run instantly.
3. Regressions and correlation tests run, replicating the paper's SPSS analysis in Python.
4. A results summary appears showing 9 key statistics alongside the published values.
5. Green 'reproduced' labels confirm matches; any discrepancy is flagged for discussion.
6. A plain-English summary highlights the headline finding: role disruption predicts feeling inauthentic, and that link is stronger for people who valued those roles highly.

Try or inspect it:
Change the bundled filter—for example, retain participants who failed one attention check—then compare the updated sample sizes and results.

Limitations:
All data come from one published study, so this demo shows reproducibility of existing findings, not new discovery. The HK sample effect (r ≈ .21) is noticeably smaller than the US sample (r ≈ .59), which the paper itself acknowledges — the demo does not paper over this. Minor numerical differences between the original SPSS output and this pipeline's output are expected and normal (different software, slightly different standard-error conventions); the substantive conclusions are unchanged across all nine tests. Results should be read as a range of plausible values, not single definitive numbers.

The data:
The demo uses a clearly labelled synthetic summary of a Hong Kong adolescent sleep and well-being study, written in the style of a typical RGC GRF application. It is not a real proposal. Because the underlying model and live data can vary, results are shown as ranges, not single fixed numbers.

What you'd see, step by step:

1. A clearly labelled synthetic proposal summary loads for a fixed criteria-mapping demonstration; no real application is assessed.
2. The pipeline runs automatically, scoring the summary against official RGC criteria.
3. A colour-coded matrix appears, showing scores and evidence quotes per criterion.
4. Gaps and revision suggestions are listed below the matrix.
5. A coverage table shows which form sections the summary addresses.
6. You can hover over any score to see the exact wording from the RGC guidance.

Try or inspect it:
Inspect the criteria-alignment table and point to a score or evidence gap you want explained. For another application, contact the team first under an agreed data-handling arrangement.

Limitations:
The demo uses a synthetic summary, not a real application, so scores reflect only what that summary evidences. The scoring model is a proxy reviewer and can vary slightly between runs; the keyword baseline is simple. Always check the current RGC guidance, as wording may change.

The data:
The study uses two public datasets: roughly 11,000 common English words rated by nearly 800,000 crowd workers on Amazon Mechanical Turk (asking 'at what age did you first learn this word in print?'), plus a matching set of zero-shot estimates produced by GPT-4o for the same words. Both datasets are openly deposited on OSF by the authors and are stable snapshots, so the numbers you see in the demo will be consistent across runs.

What you'd see, step by step:

1. Begin with coverage: 11,074 words have deposited crowd and AI estimates, while some paper analyses require a separate dataset.
2. The pipeline calculates average 'learned-in-print' ages: ~8.2 (reading) and ~8.9 (writing).
3. Those averages are checked against the published paper — both match exactly.
4. AI estimates appear alongside: GPT-4o rates words roughly one year later than humans.
5. A simple agreement chart shows where AI and human ratings agree or diverge.
6. A summary panel flags each published result as reproduced or within expected range.

Try or inspect it:
Pick any word from the list on screen and ask 'what age did GPT-4o assign this word versus the crowd?' — the demo will pull up both values instantly.

Limitations:
The data covers English words only and the crowd sample skews toward US adults, so norms may not transfer directly to other languages or populations. The AI-versus-human correlations are moderate (roughly 0.67–0.70), meaning AI estimates are a useful but imperfect proxy for human judgements. Some regression analyses in the original paper require an additional external dataset not included in the deposit, so those tables are noted as a follow-up rather than shown live. All figures reported are ranges across slightly different word-overlap subsets, not a single definitive number.

The data:
The dataset is the publicly released SPSS file attached to the original paper (open licence, CC-BY 4.0): 381 children living in Hong Kong residential care homes, assessed repeatedly over two years by 76 care workers. Because this is a fixed, archived dataset — not a live feed — the numbers are stable and will look the same every time the demo runs.

What you'd see, step by step:

1. Begin by checking the longitudinal evidence structure: 381 children, 76 care workers and repeated observations over two years.
2. The pipeline reproduces the study's sample descriptives — 76 workers, 381 residents, stay-length split.
3. Growth-curve models run; the screen shows behavioural-problem trajectories matching the paper's two-phase decline.
4. Burnout regressions appear; standardised effect sizes land close to the published values.
5. A summary panel highlights which findings reproduced, which are slightly attenuated, and the one flagged discrepancy.

Try or inspect it:
Inspect the replication results table and point to an estimate you want compared with the published result. For another dataset, contact the team first under an agreed data-handling arrangement.

Limitations:
The dataset is small (76 workers) and covers only one sector in Hong Kong, so results should not be generalised broadly. The replication used a standard multilevel approach rather than the original software (Mplus), which means slope values look numerically different until a simple scaling adjustment is applied — the direction and significance still match. One phase-2 effect appeared marginally significant in the replication but not in the paper; this is flagged honestly rather than explained away. Effect-size ranges, not single 'best' numbers, are the appropriate way to read the burnout results (roughly 25–33% of variance explained).

The data:
The data come from the published supplementary file of Fong, Lo & Ho (2023, Scientific Reports), released under an open licence. It contains questionnaire responses from 184 Hong Kong stroke survivors at baseline and 148 of the same participants two months later, covering quality of life, physical health, mood, hope, self-esteem, and disability. Because the dataset is fixed and public, results are stable — but small real-world samples like this always carry some uncertainty, so the pipeline honestly reports ranges rather than single headline numbers.

What you'd see, step by step:

1. Start with retention and measurement coverage: 184 participants at baseline and 148 at two months.
2. Descriptive summaries appear: quality-of-life scores and SF-12 health scores match the paper exactly.
3. The pipeline tests the factor structure and displays fit — poor fit for the simple model, acceptable fit once the paper's modifications are applied.
4. Reliability numbers appear: internal consistency (omega 0.85) and two-month stability (r = 0.67) both match the paper.
5. A validity table appears showing quality-of-life links to mood, hope, disability, and self-esteem — all in the expected directions.
6. A summary panel flags which findings reproduced, which are close, and where the software difference causes a known gap.

Try or inspect it:
Inspect the open-tool results table and point to a result you want compared with the published analysis. For another dataset, contact the team first under an agreed data-handling arrangement.

Limitations:
The pipeline uses a freely available Python tool rather than the specialised software (Mplus) used in the original study. This means one technical step — testing whether the scale measures time points on an identical numerical scale — can only be approximated here, not exactly replicated. Fit index numbers therefore differ slightly from the paper's (the pattern is the same), and that gap is flagged openly throughout the demo. The sample is also small (184 people), so individual numbers should be read as indicative ranges.

The data:
The demo uses a fully synthetic dataset: 240 simulated respondents, 12 questions, answered on a 1–5 scale, with a known three-group structure built in from the start. Because the correct answer is known in advance, we can prove the pipeline finds it before anyone trusts it with real data. A real deployment would use your own survey responses instead.

What you'd see, step by step:

1. The validation question comes first: can the workflow recover the known three-factor structure planted in 240 synthetic responses?
2. Reliability scores appear for each of the three subscales — all comfortably above the accepted 0.70 threshold.
3. Item-level statistics show how well each question pulls its weight in its subscale.
4. A factor analysis runs and assigns every item to a group — 100% match to the known structure.
5. The screen displays a ready-to-edit Methods paragraph and Results paragraph, written in plain academic English.

Try or inspect it:
Select one of the bundled scale-label sets and watch the draft update with those fixture names.

Limitations:
The data here is synthetic and tidy — real surveys bring missing responses, scale bunching at one end, and cross-cultural wording effects that can lower these numbers. The factor analysis shown is exploratory only; a journal submission would also need a confirmatory stage. The AI-generated paragraphs are a first draft: the interpretation, emphasis, and final wording must come from you as the author.

The data:
The dataset is 1,000 short Chinese texts — 500 originally written in Chinese and 500 translated into Chinese from English — drawn from two established public corpora and deposited openly on OSF under a Creative Commons licence. Each text is described by seven numerical scores measuring the statistical 'randomness' of its word and character patterns (Shannon entropy). The dataset is fixed and offline, so results are stable; reported figures are ranges across 20 runs rather than a single lucky number.

What you'd see, step by step:

1. Begin with the reproducible stage: seven deposited entropy scores for 1,000 labelled texts; raw segmentation cannot be rechecked.
2. Four classifiers train automatically on 70% of the texts; the remaining 30% are held back for testing.
3. Each classifier's accuracy and ranking appear on screen — SVM and LDA both reach roughly 88–90% accuracy.
4. A feature-importance chart shows which entropy scores drove the predictions — POS bigram entropy ranks first.
5. A stability panel reveals that accuracy ranges across 20 random splits, showing the published figure was one draw from that range.
6. A summary panel flags what reproduced, what was a statistical tie, and what could not be verified from the deposited data.

Try or inspect it:
Inspect the classification table and compare rows for original and translated Chinese texts. For a run on other texts, contact the team first under an agreed data-handling arrangement.

Limitations:
The study's original text-processing step (segmenting raw texts and computing entropy scores) cannot be re-checked because the public deposit contains only the final scores, not the raw texts or processing code; this demo reproduces the classification stage only. The original paper did not record its random split, so an exact number-for-number match is impossible — the right comparison is whether the pattern of results holds, and it does. All accuracy figures are ranges across 20 runs, not single best-case numbers. The dataset is modest in size (1,000 texts) and covers two specific genres, so conclusions may not generalise to all Chinese text types.

The data:
The dataset is 35 Hong Kong Chinese children with developmental dyslexia, drawn from the pretest phase of a 2021 PLOS ONE study by Prof Yetta Kwailing Wong (CUHK Educational Psychology) and colleagues. It is real, publicly deposited on OSF, and fixed — it does not change between runs, so results are stable and fully reproducible from the deposit. Post-training outcomes were not included in the public file, so only the pretest correlations and regressions can be checked here.

What you'd see, step by step:

1. Start with the evidential boundary: 35 children and pretest measures are available; post-training outcomes are not.
2. Four correlations run; reproduced values print alongside the published values.
3. Two hierarchical regression models run, one for speeded reading, one for non-speeded.
4. A side-by-side comparison table appears: every number either matches or is flagged.
5. A plain-English summary confirms all 10 published statistics were reproduced exactly.

Try or inspect it:
Remove one bundled control, such as IQ, then compare how the estimated contribution of character fluency changes.

Limitations:
Only 35 children, one school context, pretest data only — effect sizes this large are unlikely to replicate in a broader or more diverse sample. The pipeline reproduces the paper's own numbers perfectly, but that tells us the analysis was done correctly, not that the findings will generalise. Post-training outcomes and the full longitudinal data are not in the public deposit and cannot be checked here.

The data:
The data come from three real behavioural experiments (visual search reaction-time tasks) published in 2025 by Prof Shelley Xiuli Tong's lab at HKU, with all files openly deposited by the authors on OSF. The pipeline works from those original files directly, so the numbers shown are grounded in a real published study, not synthetic examples. Because the pipeline re-estimates the models each time it runs, reported results are given as ranges rather than single fixed figures — minor variation is expected and normal.

What you'd see, step by step:

1. The archived OSF files load for a GEE reanalysis; because this targets a different estimand from the published mixed model, exact coefficients are not replication targets.
2. Data cleaning runs: the screen shows how many trials are removed and why.
3. Statistical models run automatically on all three experiments at once.
4. A results summary appears: which published findings were reproduced, which were close.
5. Two independent methods are compared side-by-side, showing they agree on the headline pattern.
6. A final verdict panel highlights the key takeaway: the paper's core claim holds up.

Try or inspect it:
Inspect the results table, choose one experiment, and compare the reanalysis pattern with the published result. For another dataset, contact the team to agree data handling first.

Limitations:
The pipeline uses a different statistical method (population-averaged GEE) from the one in the paper (subject-specific mixed model), so exact numbers will not match — direction and significance pattern are the right things to compare. Data-trimming percentages come close but not always exactly, reflecting small differences in how ambiguous cases are handled. With only a few dozen participants per experiment, effect sizes are modest and some secondary results sit near the significance boundary, so treat any single p-value with caution.

The data:
The dataset comes directly from the authors' own public archive (OSF): 3,694 Korean hip-hop artists, with each artist's number of featuring collaborations and a quality score built from likes and prior fame. The data is cleaned and anonymised by the original authors and is freely reusable. Because it is a fixed, archived snapshot, results are stable across re-runs — no live-data drift.

What you'd see, step by step:

1. The authors' archived dataset of 3,694 Korean hip-hop artists loads for the reproduced subset; the paper's simulations and second-domain analysis are not run.
2. The pipeline fits two statistical curves to the collaboration network and runs bootstrap tests.
3. Results appear: the 'rich-get-richer' (power-law) curve is rejected; the alternative curve is not.
4. A head-to-head comparison confirms the alternative fits better — matching the published numbers exactly.
5. Five regression models run in sequence; R² values and key controls appear, matching the paper's table.
6. A summary panel shows which published findings reproduced and flags what was not re-run.

Try or inspect it:
Inspect the model-comparison table and compare the power-law and log-normal results; the coauthorship analysis is not included. For another dataset, contact the team to agree data handling first.

Limitations:
Bootstrap p-values shift slightly each run by design; the reliable target is the direction (power-law rejected, log-normal plausible), not an exact decimal. The 'S-shaped' growth pattern described in the paper belongs to the academic coauthorship data, not the hip-hop data — that component was not re-run here. The simulation experiments from the paper (synthetic network models) were also not reproduced in this demo. The dataset is relatively small for network science and covers one music scene, so findings may not generalise beyond Korean hip-hop.

The data:
The dataset is 66 Cantonese-speaking children (children with Developmental Language Disorder, age-matched typical peers, and younger typical peers) tested on sentence-comprehension tasks; it is the authors' own real data, openly shared on the OSF repository under a Creative Commons licence. Because the pipeline pulls directly from that public archive, results are stable — the source files do not change. Exact numbers may land in a small range across runs due to differences in statistical software, so the demo reports ranges rather than a single pinpoint figure.

What you'd see, step by step:

1. The paper, open dataset for 66 children, and original R script appear; the live reproduction uses a Python approximation rather than the authors' exact model.
2. Raw data loads automatically; the same 62–64 children the authors kept are selected.
3. The statistical models run; progress ticks by in roughly 30 seconds.
4. A results panel appears showing each of the paper's key findings side-by-side with the reproduced ones.
5. Green 'reproduced' labels confirm: subject relatives easier than object relatives, CL-type sentences easier than ge3-type, and children with DLD scoring lower than typical peers — all matching the published paper.
6. A plain-language summary box highlights the three headline results and flags what a full replication would still require.

Try or inspect it:
Select a headline group-level result and request a plain-language methodological gloss. The response must not classify, diagnose, or recommend action for an individual child.

Limitations:
The dataset is small (66 children, one study) and cannot support broad generalisations on its own. The replication uses a different statistical method from the original (a Python approximation rather than the exact R model), so coefficient sizes shift slightly — direction and significance are what reproduce, not every decimal place. A word-for-word software replication would require running the authors' own R script, which is included in the demo package but not shown live.

The data:
ZuCo (Zürich Cognitive Language Processing Corpus) 1.0 — the natural-reading task: 12 healthy adults read English sentences while EEG and eye-tracking were recorded simultaneously. The public research data was mirrored and verified from the original distribution; this demo processes 36,562 word-level fixation-locked FRP epochs through a full MNE-Python pipeline (band-pass filter, ICA, fixation-to-EEG synchronisation at 2 ms/sample precision).

What you'd see, step by step:

1. The research question: does word frequency reliably modulate the N400 during natural reading, when reading is unconstrained by a task?
2. Data loading: simultaneous EEG + eye-tracking from 12 participants reading real sentences.
3. Cleaning: automatic ICA, fixation–EEG synchronisation, word-level epoch extraction — with the exact thresholds shown.
4. Design: pre-specified N400 window (300–500 ms), median frequency split, subject-level paired contrast.
5. Results: the headline frequency effect is null (p = 0.797, n = 12) — reported honestly, with power caveats.
6. Validation: alignment QC (2.0 ms/sample slope, median residual ≤ 1 ms), subject-exclusion review, provenance hashes.
7. Conclusion: what the null means, what a stronger test would require (≥ 40 subjects, continuous predictors), and the tentative cross-modal gaze–N400 correlation.

Try or inspect it:
Open the Explore assumptions tab, then change the N400 window, subject-exclusion or sync thresholds and review the recomputed statistics. Non-default runs are marked EXPLORATORY.

Evidence:
ZuCo 1.0 natural-reading, 12 participants, 36,562 word-level fixation-locked FRP epochs (EEG + eye-tracking synchronized at 2 ms/sample). Headline frequency effect is null — p=0.797, d=−0.076, n=12 — reported honestly with power caveats (≈9% at the pre-specified bound). The Explore assumptions tab lets visitors vary the N400 window, subject-exclusion and sync thresholds with live client-side recomputation; non-default runs carry an EXPLORATORY banner and underpowered analyses are blocked. validated 2026-09-01.

Limitations:
With 12 participants the study is underpowered for confirmatory claims — all inferential results are descriptive (power ≈ 9% at the pre-specified bound). The frequency split discards graded frequency information, sentence context was not controlled, and the next-fixation response overlaps the analysis window (fixation-overlap contamination risk). One subject's run was recovered via alternate runs; one run is absent from the mirror and disclosed. The cross-modal correlation is model-dependent and cannot be confirmed at this sample size.

The data:
The study uses 566 street-view photographs of Shenzhen collected via a mapping service, each rated by survey respondents for how psychologically restorative the space feels. The images and ratings are publicly available on GitHub alongside the original analysis code. Because the publicly shared files appear to differ slightly from the exact version used to produce the published results, some numbers come out as ranges rather than single figures — that gap is itself part of what the demo illustrates.

What you'd see, step by step:

1. The shared 566-image dataset, human ratings, and previously generated model predictions load; the commercial vision model itself is not rerun.
2. A conventional machine-learning model runs on those photos; its prediction score appears.
3. The AI vision method (with built-in urban knowledge) runs; a score of R²=0.76 appears — matching the paper exactly.
4. The same method runs without that built-in knowledge; its lower score of R²=0.37 appears, again matching.
5. A sensitivity check varies how many reference images the AI uses (2, 3, or 5); scores drop in the same order the paper reports.
6. A summary panel flags the one result that cannot yet be fully verified — the baseline comparison — and explains why.

Try or inspect it:
Inspect the audit table and compare survey ratings with the image-model predictions for a bundled street image. For another dataset, contact the team to agree data handling first.

Limitations:
The dataset is small (566 images, one Chinese city, one season) so results should not be generalised to other cities or climates. The AI vision predictions themselves were made by a commercial model and are taken from the published files — the demo verifies the scoring of those predictions, not the full end-to-end pipeline. Most importantly, the paper's headline claim of a 0.535 improvement over the baseline cannot be independently confirmed yet: the baseline score differs between the published notebook and the shared data files, likely due to a data-version change. The improvement is probably real, but the exact magnitude is unverified.

The data:
The corpus is 77 official English-language news articles on inclusive education in China, drawn from a public archive deposited by the paper's authors on OSF. The texts are real published news items, not synthetic, and the deposit is stable — the same file is retrieved every run. Because the authors' original line-by-line coding sheets were not deposited, the pipeline builds its own independent coding for comparison rather than replaying the authors' exact work.

What you'd see, step by step:

1. The first decision is comparability: the paper counts concordance instances, while this case independently codes sentences.
2. The pipeline scans every sentence and flags those discussing inclusive-education discourse — about 1,000 sentences highlighted.
3. Top words appearing near 'inclusive education' display: children, schools, quality, promote, concept.
4. Each flagged sentence is automatically labelled with one of four published categories: efforts, consensus, challenges, or other.
5. A comparison panel appears showing the published split alongside the pipeline's own split — 'efforts' leads in both, at roughly 65% (published) and 61% (pipeline).
6. A plain-language summary card appears noting where the two codings agree and where they diverge.

Try or inspect it:
Select a bundled news-sentence fixture and inspect its proposed category and reasoning.

Limitations:
The authors' original sentence-by-sentence coding was never made public, so this demo cannot reproduce their exact numbers — it produces an independent re-coding for comparison only. The two codings use different counting units (the paper counts 520 concordance instances; the pipeline counts roughly 1,000 sentences), which partly explains differences in the minority categories. The automated coder agrees only moderately with a simple keyword check, meaning borderline sentences — especially praise-heavy policy lines — could reasonably be labelled 'efforts' or 'consensus' by different coders, human or machine. Finally, the corpus covers only official-channel English reporting, so findings describe how authorities project inclusive education abroad, not the full range of media coverage.

The data:
Thirty real consumer-complaint narratives from a US financial regulator's public database, standing in for open-ended survey or interview responses. The regulator's own category labels serve as the human-coded ground truth. Because this is a fixed public dataset the results are stable, but if you substituted live or updated text the model's agreement scores would shift — treat any figures as a range, not a fixed number.

What you'd see, step by step:

1. Thirty public regulator complaint narratives and a fixed five-category codebook load for a small classification-agreement exercise; the source labels are not validated interpretations of qualitative meaning.
2. The AI reads each text and assigns one category; a progress indicator shows coding in real time.
3. A summary appears: 27 of 30 texts coded identically to the human — 90% agreement overall.
4. A reliability score (kappa ≈ 0.73) is displayed; a plain-English label reads 'substantial agreement'.
5. Three disagreements are highlighted; clicking one shows the text, both codes, and a note on why definitions may overlap.
6. A clean CSV of all coded items is ready to download for further analysis.

Try or inspect it:
Select an included public complaint, inspect the assigned category, and compare it with the source label. Do not paste confidential interview, survey, or participant text.

Limitations:
The dataset is small (30 items) and comes from one specific domain, so the kappa figure (roughly 0.70–0.80 across runs) is illustrative rather than definitive. Categories with only one or two examples — such as 'credit card' or 'debt collection' here — can show wide swings in agreement by chance. A real study would need a larger, domain-matched sample and a codebook refined through the disagreement-review step shown in the demo.

The data:
The demo uses a standard open-access teaching dataset: a 60-channel EEG recording of a person responding to auditory and visual cues, roughly two and a half minutes long. It is a real (not synthetic) recording widely used in neuroscience training worldwide, so results are stable and repeatable across runs. Because this is a single-subject tutorial file rather than a full study dataset, the numbers shown are illustrative benchmarks, not publishable group statistics.

What you'd see, step by step:

1. The raw brainwave recording loads and its 60 channels appear as scrolling coloured traces.
2. The pipeline filters out electrical noise and eye-blink artefacts automatically; a clean signal remains.
3. The recording is sliced into ~200 brief trial windows, one per stimulus event; 90% pass quality checks.
4. Averaged waveforms (ERPs) appear for auditory vs. visual conditions, showing distinct timing and scalp location differences.
5. A scalp topography map lights up showing where brain activity peaks for each condition at each moment in time.
6. A time-frequency colour map reveals how oscillatory brain power rises and falls around each visual stimulus.

Try or inspect it:
Point to a brain-map or signal panel and ask what it shows for a reading or language study. Contact the team for another run under an agreed data-handling arrangement.

Limitations:
This is a single participant from a tutorial dataset, so no group statistics or individual differences are shown — a real study would add per-participant tuning and mixed-effects models. The pipeline settings (filters, epoch length, artefact thresholds) are reasonable defaults, not optimised for a specific research question; a live project would involve a pre-registered analysis plan. Results should be read as a proof-of-concept demonstration of the workflow, not as findings.

The data:
Fifteen scripted Cantonese student messages, split evenly across everyday chat, moderate distress, and crisis statements (including self-harm and suicidal ideation). The messages are purpose-written for this demo, not drawn from real student records — this is a proof-of-concept mechanism test, not a clinical study. Because the scripts are fixed, results are stable across runs; a live deployment with real users would behave differently and would need proper ethics-approved evaluation before any conclusions could be drawn.

What you'd see, step by step:

1. Fifteen purpose-written Cantonese test messages load with predefined scenario labels; none comes from a real student or live conversation.
2. The chatbot replies to each message in Cantonese with an empathetic response.
3. A separate automated judge reads each reply and scores distress level, empathy, hotline presence, and whether handoff was recommended.
4. A summary panel appears: 15/15 distress levels correctly identified, all 5 crisis messages triggered a real hotline number and a human-handoff recommendation.
5. You can scroll the full conversation log and see how the tone shifts — warm reassurance for everyday chat, gentle concern for distress, immediate safety referral for crisis.
6. A limitations card is shown last, reminding you that scripted demos are not clinical trials and a real service needs further validation.

Try or inspect it:
Use the supplied selector to compare everyday, moderate-distress, and crisis-handoff scripts. Do not enter personal experiences, identifying details, or requests for crisis support.

Limitations:
All 15 messages were written by the research team, so the chatbot has never been tested on real students; strong results here confirm the design works in principle, not in practice. The automated judge is itself an AI, so empathy scores are indicative, not clinically validated — the one hard check is a simple text-search confirming the real hotline number appears in every crisis reply. This is a triage-and-handoff tool only: it does not diagnose, treat, or counsel, and any live deployment would need ethics approval, clinician review of the scripts, and verification of local crisis resources.

The data:
The demo uses 2,000 real IMDB consumer reviews drawn from a publicly available dataset, with a 120-review sub-sample that carries human-verified sentiment labels used to check the AI's accuracy. Because this is a fixed public dataset the results are stable for today's walkthrough; if the same pipeline were pointed at live social-media or regulatory feeds, scores would shift over time and should be treated as ranges rather than single fixed numbers.

What you'd see, step by step:

1. Start with the validation boundary: 120 labelled English reviews calibrate sentiment within a fixed 2,000-review corpus.
2. The AI reads each review and labels it positive or negative; accuracy checks at 96% against human labels.
3. The topic model groups all 2,000 reviews into six themes automatically — no manual coding.
4. A sentiment score is attached to each theme, revealing which topics drive praise or complaint.
5. An AI-drafted executive summary appears, translating the theme table into plain business findings.

Try or inspect it:
Select a bundled review or policy-text fixture and inspect its sentiment label and nearest theme.

Limitations:
The accuracy figure (96%) comes from English-language reviews; Chinese or Cantonese text needs additional language-specific setup before the same reliability applies. Topic names are generated from the most frequent words in each cluster — a human researcher should review and rename them for any publication. The pipeline is validated on a 120-review sample; a full deployment on your own corpus would require a fresh calibration run on that material.

The data:
The demo uses a simulated word-response experiment: 32 participants each completing 80 trials, with a known 45 ms speed difference built in so the pipeline can be checked against a correct answer. The data are synthetic (not from a real study), designed to mimic standard cognitive-psychology experiments. Because the data are fixed for this demo, results are stable — a live pipeline on real collected data would produce ranges rather than single numbers.

What you'd see, step by step:

1. Raw trial-by-trial response times load and display on screen.
2. The pipeline automatically removes invalid trials (errors, extreme outliers).
3. Group averages appear: related-prime responses average 45 ms faster than unrelated.
4. A significance test and effect-size score are calculated and shown instantly.
5. An AI-drafted Results paragraph appears, ready for the researcher to edit.

Try or inspect it:
Inspect the bundled trial results and exclusion summary, then point to a rule you want explained. To run your own study, contact the team first under an agreed data-handling arrangement.

Limitations:
The data are simulated, so the pipeline has not yet faced the messier trial structures of a real study — though the steps carry over directly. Only a simple two-condition comparison is shown here; real multi-experiment papers typically need more complex models, which are an extension not yet in this demo. The AI-drafted paragraph is a writing aid and starting point, not a submission-ready text.

The data:
Ten Chinese–English sentence pairs were hand-crafted for this demo: each pair contains one polished translation and one with a single planted error (mistranslation, omission, unnecessary addition, awkward phrasing, or wrong register). The data is small and purpose-built to test the mechanism, not to benchmark real-world performance. Because this is a fixed test set rather than live text, scores are stable across runs.

What you'd see, step by step:

1. Ten Chinese sentences appear on screen, each paired with two English translations.
2. The AI scores both translations for meaning (adequacy) and readability (fluency), blind to which is correct.
3. A results table appears: the good translation scores 5/5 every time; flawed translations average around 3.5/5.
4. For each flawed translation, the AI flags the error type — mistranslation, omission, awkward phrasing, etc.
5. A plain-English summary appears, confirming the AI preferred the good translation in all 10 cases and spotted the flaw in 9 of 10.
6. Click any row to see the AI's written justification for its scores and error label.

Try or inspect it:
Choose one of the bundled held-out translation pairs and inspect the score, preference and flagged concern.

Limitations:
This is a 10-sentence proof-of-concept with errors that were deliberately planted, so the strong numbers (100% preference accuracy, 90% error detection) reflect a controlled setting, not a live translation workflow. Error-type labelling is less reliable — the AI correctly named the error category about two-thirds of the time, and occasionally confused related error types. AI scoring correlates with human judgment but should be treated as a first-pass triage tool, not a replacement for human assessment in high-stakes contexts. A real study would use 100+ sentences evaluated by at least two human raters alongside the AI.

The data:
The demo uses simulated Chinese sentence-reading data (30 readers, 40 sentences each) with known word-frequency differences built in, so the pipeline can be checked against a correct answer. This is synthetic validation data, not a real corpus; real deployment would use your lab's own recordings or public datasets such as ZuCo or GECO. Because the demo data is fixed, the numbers you see today will be stable — but results on live lab data will naturally vary across runs and participant samples.

What you'd see, step by step:

1. Raw fixation records load: one row per word per participant, nothing else.
2. Pipeline calculates five standard reading measures for every target word automatically.
3. Results appear: low-frequency words take longer to read on every single measure.
4. A summary paragraph — ready to paste into a manuscript — is drafted instantly.
5. You compare the pipeline's recovered differences to the known built-in effects as a quality check.

Try or inspect it:
Choose a bundled word-level fixation fixture and re-run the calculations using the supplied column mapping.

Limitations:
The data today is synthetic and small (30 simulated participants), so this validates that the measures are computed correctly, not that the norms match any real population. The pipeline does not yet include mixed-effects models, which are the current statistical standard for trial-level eye-tracking data — the paired t-tests shown are a useful first pass only. The AI-drafted paragraph is a starting point; authors must review it before submission.

The data:
The demo uses a short synthetic Cantonese passage (45.6 seconds, 135 aligned units) generated by a text-to-speech voice, so results are perfectly reproducible for a live showing. Real interview or corpus recordings can be substituted directly — the pipeline handles them the same way. Because synthetic speech is cleaner than natural conversation, pitch and timing figures on real recordings will vary; treat any numbers as illustrative ranges, not fixed benchmarks.

What you'd see, step by step:

1. A Cantonese audio clip loads and plays — you hear the sentence spoken aloud.
2. The pipeline transcribes the audio; a full Chinese-character transcript appears on screen.
3. Each word lights up with its precise start and end time in the recording.
4. A results table appears showing pitch (F0) and loudness for every word.
5. A summary line confirms 135 aligned units processed, all with pitch measurements, at ~2.96 units per second.

Try or inspect it:
Choose another bundled Chinese sentence fixture and re-run the pipeline; compare its transcript and pitch table with the supplied reference.

Limitations:
The audio here is synthetic (computer-generated speech), so it is unusually clean; natural recordings — with background noise, overlapping speakers, or heavy dialectal variation — will produce messier pitch tracks and occasional transcription errors. Timing is at the word level, not the individual sound level, so fine-grained phoneme work would need an extra alignment step. Pitch figures are raw and would require speaker-specific tuning before appearing in a publication.

The data:
Two audio recordings of the same 83-character Chinese passage were used: one read fluently, one read slowly with four deliberate word substitutions (e.g., 森林 replaced by 公園). Both voices are synthetic — real children's recordings would be the next step. Because the conditions are fixed and controlled, results are stable for this demo, but scores on real classroom audio would vary with recording quality and individual children.

What you'd see, step by step:

1. The passage text and two audio files load on screen — you can see the Chinese characters.
2. The system listens to each recording and produces a word-by-word transcript automatically.
3. Rate, accuracy, and pause time are calculated and displayed side by side for both readings.
4. An AI rubric rates expression, phrasing, and pace on a standard 1–4 scale for each reading.
5. A summary panel shows fluent beating disfluent on every expected measure — the self-check passes.

Try or inspect it:
Ask the demo to play back one of the two recordings while you watch the live transcript appear word by word.

Limitations:
Both recordings use a synthetic (text-to-speech) child voice, so pitch-variety scores do not meaningfully separate the two conditions — that finding needs real expressive versus flat child readings to be trusted. Accuracy figures are a conservative lower bound because the speech recogniser occasionally 'corrects' substituted words back to the original text; a production system would use forced alignment against the passage to catch every error. This is a two-condition demo of the measurement machinery, not a norming study — it shows the pipeline works, not what typical scores look like across a class or grade level.

The data:
The demo uses ten hand-written realistic academic emails, one fictional student CV (for Chan Hoi Yan, an MPhil Linguistics candidate), and a transcript of one fictional departmental meeting — all synthetic and modelled on a Hong Kong university context. Because the data is fixed and self-contained, results are stable across runs; a real deployment using your own emails would need two to four weeks of tuning and results would vary.

What you'd see, step by step:

1. Ten fictional academic emails load on screen; no real inbox, student record, or university system is connected.
2. The agent labels every email — urgent, reply, delegate, or ignore — and suggests a one-line action.
3. Results appear beside gold labels; all 10 categories matched, but priority levels were only half right.
4. A reference letter for the MPhil student appears; a coverage check confirms every CV fact was included, and three soft character claims are flagged as unverified.
5. Paste-in meeting notes are processed; three of four decisions and four of five action items surface, each with owner and deadline.
6. A summary panel shows where the agent succeeded and where you should review before acting.

Try or inspect it:
Choose a bundled fictional email, edit its wording without adding personal or confidential information, then compare the suggested category and action with your judgment.

Limitations:
All three tasks run on ten emails, one CV, and one meeting — a proof of concept, not a tested system. Gold labels reflect one person's judgment, so accuracy figures should be read as illustrative ranges rather than firm benchmarks. The agent drafts and flags; it does not send emails or submit forms — integration with Outlook or university portals is a separate engineering project not covered here. Priority scoring was only 50% accurate on this small set, so treat priority suggestions as a prompt to check, not a decision.

The data:
A short Cantonese parent–child play session (roughly 48 seconds, 12 scripted turns) was synthesised for the demo by pitch-shifting an adult voice to approximate a child's voice — a standard acoustic technique. The audio is fully synthetic and clean, not drawn from real fieldwork. Because the demo uses a fixed recording, results are stable; a live deployment with real noisy recordings would produce ranges rather than exact counts.

What you'd see, step by step:

1. A 48-second Cantonese audio clip loads and plays in the browser.
2. The pipeline transcribes speech turn-by-turn and labels each turn Adult or Child.
3. A summary panel appears: 6 child vocalisations detected, 11 conversational turns, adult talk share 73%.
4. A pitch graph shows child turns consistently higher than adult turns, confirming speaker labels are acoustically grounded.
5. A scrollable transcript lets you read every turn with its timestamp and speaker role.
6. A comparison panel shows how these numbers map directly onto the metrics a LENA device would report.

Try or inspect it:
Select one of the bundled short audio fixtures and compare how turn counts and talk-share estimates change.

Limitations:
The audio is synthetic and clean — no overlapping speech, no background noise. Real nursery or home recordings are much messier and would require additional tuning before the counts are reliable. Speaker labels here were verified against a script written by the demo author, not independent human coders. Treat the headline numbers as a proof of concept, not a validated research instrument.

The data:
Ten target words were recorded using synthetic Cantonese child and adult voices, each word embedded in a standard picture-naming phrase used in real clinical assessments. The voices are computer-generated, not real children — this is a mechanism test, not a clinical dataset. Because the errors and targets are built in by design, we know the exact right answer for every item, which makes the accuracy figures fully verifiable.

What you'd see, step by step:

1. The demo loads two voices — one adult, one child — reading the same Cantonese passage.
2. Both recordings are fed to a speech-recognition engine; transcripts appear side by side on screen.
3. Ten child-voice clips play, each containing a deliberate mispronunciation typical of speech sound disorder.
4. The transcripts appear; you see whether the mispronunciation survived into the text or was silently corrected.
5. An AI language model reads each transcript and names the disorder pattern — fronting, stopping, etc.
6. A results summary appears: 7 out of 10 disorder patterns correctly identified, 8 out of 10 errors confirmed present.

Try or inspect it:
Point to a word in the results table and ask why it was marked unclear; use the displayed per-item evidence to check the explanation.

Limitations:
All voices are synthetic, so the demo proves the pipeline works in principle — it does not yet measure how well it handles real children's voices, which are less precise and more variable. Ten items is a pilot; a clinically meaningful evaluation would need 100 or more items checked against human speech therapist judgements. The speech-recognition engine occasionally 'corrects' a mispronunciation back to the standard word before the disorder detector even sees it, which is a known technical hurdle this demo is designed to expose, not hide.

The data:
The demo uses a short synthetic Cantonese maths lesson (under a minute, 10 exchanges between one teacher and one student) recorded in a clean, controlled way so the pipeline always runs the same way live. Because it is synthetic, results are stable for this demo; a real classroom recording would introduce background noise, overlapping voices, and children's accents, so accuracy figures should be treated as a best-case range rather than a fixed number.

What you'd see, step by step:

1. A short Cantonese maths lesson audio file loads — you see the waveform and duration.
2. The pipeline transcribes the audio; a turn-by-turn transcript appears on screen in Cantonese.
3. Each turn is automatically labelled: Teacher or Student, and a discourse code (Instruction, Question, Response, or Feedback).
4. A summary panel shows the teacher-talk ratio (roughly two-thirds) and a breakdown of question, response, and feedback turns.
5. A reliability table appears comparing the automatic codes against a hand-written gold standard — about two-thirds of discourse codes matched, which is moderate agreement for a fully automated first pass.

Try or inspect it:
Inspect the bundled synthetic lesson and compare its coded exchanges. Do not upload real classroom data; any real-data run requires prior contact, consent, and agreed data handling.

Limitations:
The audio is synthetic — two clean, adult voices with no overlap — so accuracy will be lower on real classroom recordings with children's voices, noise, and cross-talk. Discourse codes are illustrative; a real study would develop the codebook together with your research team. Treat the two-thirds code-agreement figure as an upper bound, not a guaranteed result.

The data:
The demo uses 20 short passages: one genuine human excerpt (from Wikipedia) and one AI-generated passage on the same topic, repeated across 10 Hong Kong-relevant topics (e.g. Cantonese, MTR, Hong Kong cuisine). Because the labels are known by construction — we made the AI passages ourselves — no human annotation is needed to check whether the detectors are right. The texts are a small proof-of-concept set; a real study would require hundreds of passages across many writing styles and AI models.

What you'd see, step by step:

1. The 20 passages load on screen, each tagged with its true human-or-AI label.
2. The pipeline extracts writing-style statistics (word length, punctuation patterns, vocabulary variety) for every passage.
3. A simple classifier trains on 19 passages and predicts the 20th — repeated for all 20; final score appears: 16 out of 20 correct.
4. An AI judge reads each passage blindly and gives a verdict with a short reason; per-text results populate a results table.
5. A side-by-side summary appears: the style classifier got 16/20 correct; the AI judge got all 10 topic-pairs right when comparing the two passages directly, but only 11/20 when judging each passage in isolation.
6. The researcher is invited to inspect any row — clicking a topic shows the original passage, the judge's stated reason, and whether the verdict was right or wrong.

Try or inspect it:
Point to any topic in the results table and inspect how the judge reached its decision for that passage. Use only the 20 bundled passages.

Limitations:
This is a 20-text mechanism demo, not a benchmark — results would shift with a larger or more varied dataset. The style classifier (16/20) is a weak-but-real signal; the AI judge is stronger at paired comparison (10/10) but unreliable on single passages (11/20), and it tends to flag encyclopedic human writing as AI. Detection accuracy is also a moving target: newer AI models write less predictably, so these numbers reflect today's AI output style and should be treated as indicative ranges, not fixed figures.

The data:
The source material is a real, completed PhD thesis (289 pages) on accented-speech intelligibility, submitted to UNSW in 2013, plus four raw experiment data files from the same Dropbox folder — together covering several thousand listening trials across multiple experiments. The data are real and finished (not synthetic), so results are stable; however, the pipeline uses a simplified statistical approach rather than the full mixed-model analyses in the original thesis, which is why some numbers come out slightly different from the published figures. Any figures quoted should be treated as ranges, not single authoritative values.

What you'd see, step by step:

1. A preconfigured, authorised Dropbox source supplies the public demonstration files: a real 289-page thesis and associated experiment data.
2. The PDF is parsed; a structured chapter-and-experiment summary populates on screen.
3. Raw data files are loaded; subject-level accuracy scores are recomputed automatically.
4. A side-by-side comparison table appears: recomputed statistics versus thesis-reported values.
5. Three charts render showing listener-group accuracy differences and tone-error effects.
6. A drafted Results paragraph appears, ready to read and compare with the thesis prose.

Try or inspect it:
Use only the supplied files: point to a comparison-table row and ask the panel to explain it. For another project, contact the team first to agree scope, permissions, and data handling.

Limitations:
Only one experiment's statistics were fully verifiable from the raw files provided; other chapters had no machine-readable data attached, so the pipeline cannot check those. The statistical tests used here are simplified (basic t-tests on group means) rather than the mixed-model ANOVAs in the original thesis — this explains small differences in t-values and effect sizes, and means results should be read as approximate reproductions, not exact replications. The eye-tracking session log was parsed for trial structure only; actual gaze data requires specialist export files not included. All findings come from a single, decade-old thesis with modest sample sizes (50–60 participants per study), so nothing here should be generalised.