studentbench

This is the single authoritative release-facing description of task membership for the final StudentBench GRE dataset. It replaces the former instrument-specific inclusion documents. Quant and Verbal share one policy; the few instrument-specific values are stated explicitly below.

The machine-readable source of truth is studentbench_overall_parameters.json. If this document and that registry ever disagree, the release must fail verification rather than silently choose one. 1_main_leaderboard_data/ is already the physically pruned primary analysis cohort, so paper builders read it directly and apply no additional eligibility filter.

Instrument-specific values

Property	GRE Quant	GRE Verbal
Instrument bank	`gre_quant:assembly_main:27q-section47m:2c31aa3e`	`gre_verbal:27q-section41m:037ffb6b`
Questions per assessment	27	27
Nominal time per assessment	47 minutes	41 minutes
Eligible pre-test score	5–24, inclusive	4–24, inclusive
Primary timing window	first submissions through minute 42	first submissions through minute 36
Ignored end-of-test period	final 5 minutes	final 5 minutes

Quant and Verbal are separate instruments. Section-specific results are reported separately. Combined analyses are allowed only when they explicitly include GRE section in the model and label the pooled estimand; raw scores must not be pooled as though the instruments were interchangeable.

Rules shared by the final main cohort

A task enters 1_main_leaderboard_data/ only when every applicable rule below passes:

Research provenance. Exclude admin, test, demo and staff-QA records; require research consent; enroll within the applicable half-open study window; and apply the known identity backstops. Non-consenting and administrative records are not public research data.
Instrument and score range. Require the assigned instrument, a valid pre-test score within the instrument-specific range in the table above, and enough improvement headroom. These are pre-treatment eligibility rules, not outcome filters.
Assessment effort. For every completed timed assessment, allow at most 10 tab-outs and require at least 600 seconds of summed item-level time on task.
First attempt. Retain the first same-instrument attempt per account. Within-run duplicate accounts retain the completed row and then the earliest row.
Completion. Require both pre-test and post-test completion and at least 14 answered questions on each 27-question assessment. “Answered” means a nonblank submitted answer.
Assigned-treatment integrity. Exclude provenance-pilot arms and any task with a documented cross-model lesson-plan fallback or substantive additional treatment outside the assigned arm. Small incidental exposure retained by PI adjudication does not by itself trigger exclusion.
AI tutoring participation. For AI-tutored tasks only, require all three:
- at least 3 student-authored chat messages;
- at least 5 total practice-answer attempts; and
- at least 3 distinct practice problems attempted.
A practice problem is attempted when answer_submission_count >= 1; the released data require that indicator to agree with was_answered. These gates do not apply to human or control arms.
Rapid-submission screen. On each pre-test and post-test, retain the earliest server-recorded submission for each question, order those submissions chronologically within the instrument-specific primary timing window, and compute adjacent inter-submission intervals. Exclude the task if the linear 90th percentile of those intervals is under 10 seconds on either assessment. The rule is evaluated only with at least 13 intervals; an under-supported phase is retained unless another gate fails. The final five minutes are excluded from this calculation.

Observed lift, post-test score, correctness and every other learning outcome are forbidden as selection inputs. There is no lift-based exclusion rule.

The identically constructed p80 timing rule is a registered sensitivity analysis only. It does not define membership in the physically pruned main dataset.

Instrument-specific source and provenance rules

GRE Quant

Main AI and human data use the initial comparison cohort, with the corrected reruns and Gemini 3.7 collection documented below.
Final control: experiment run 8, with graded pre-test review for at least five minutes before study videos. The process gate uses the first pretest_submitted/pretest_auto_submitted event to the first control_video_session_start.
The earlier Quant control without graded pre-test review remains outside the main cohort. Its assessment-response tables are retained under 4_supplementary_and_excluded_research_data/quant_control_no_pretest_review/; this is a partial historical assessment release, not a reconstruction of the full cohort.
The original Opus 5 and Gemini 3.6 Flash max-prompt conditions are not main-leaderboard arms. They were replaced by correctly configured run-8 lean-v3 collections and are preserved under 4_supplementary_and_excluded_research_data/quant_ai_max_prompt_runs/.
The final corrected Quant counts are 75 Opus 5 students and 89 Gemini 3.6 Flash students.
Gemini 3.7 Flash (medium) contributes 87 retained run-11 students from the final study snapshot and uses the final lean-v3 Quant prompt pair.
Human source data come from the final human comparison cohort and the PI-pinned transcript archive. Human lesson plans attach only through exact participant-linked teacher_plan rows; missing plans or transcripts are marked unavailable and never imputed.
Quant task_id_1282 received substantive AI tutoring before its human-tutoring session and is excluded as mixed treatment. Human release profiles neutralize latent enrollment-time AI model and prompt assignments.

GRE Verbal

AI and human data use the final Verbal comparison cohort.
Final control: experiment run 9, with at least five minutes of graded pre-test review before study videos.
The 70 earlier completed controls without that review remain outside the main cohort. Their assessment-response tables are retained under 4_supplementary_and_excluded_research_data/verbal_control_no_pretest_review/ as partial historical records.
All retained Verbal AI students use gre.verbal.plan.lean.v2 and gre.verbal.tutor.lean.v2. Sessions receiving a lesson-plan fallback from another model are excluded.
Gemini 3.7 Flash (medium) contributes 92 retained run-12 students from the final study snapshot.
Earlier same-instrument starts participate in first-attempt adjudication even when the earlier record itself is outside the public release window.

Final physically pruned main cohort

Instrument	AI	Human	Control	Total	Pre/post response rows
GRE Quant	1,139	61	91	1,291	69,714
GRE Verbal	1,000	79	99	1,178	63,612
Total	2,139	140	190	2,469	133,326

Every main student has exactly 27 pre-test and 27 post-test response rows. The paper build must use all 2,469 physical main tasks and must report runtime_filter_applied = false.

Excluded and supplementary data

The task-level excluded cohorts listed below retain their profiles, assessment responses and selected research material under 4_supplementary_and_excluded_research_data/. Physical placement establishes their release role, and every profile has a plain-language primary_main_leaderboard_exclusion_reason. Ancillary files for specified exclusion groups are omitted. The three historical comparison cohorts described above retain assessment tables and cohort descriptions only; their original cohort/session-folder placement identifies their scope, and they have no released student.json in this partial restoration.

Earlier mutually exclusive failure routing contains 874 Quant and 348 Verbal task trees.
The final outcome-blind protocol-adherence migration moved 107 Quant and 112 Verbal tasks from the former main tree into 4_supplementary_and_excluded_research_data/excluded_from_final_leaderboard_by_protocol_adherence/.
All task-level exclusions are physically separated under 4_supplementary_and_excluded_research_data/; each excluded student.json records its primary main-leaderboard exclusion reason.
Alternative, superseded and pilot conditions are documented separately because they are not task-level failures of the final protocol. They must never be pooled into the main leaderboard.

Physical placement and each excluded profile's primary exclusion reason define release membership. Independent validation requires every excluded record to state its applicable exclusion and every retained main task to pass all applicable current raw protocol gates; internal migration evidence is maintained outside the distributed dataset.

Release and integrity gates

For every main task, the external maintainer-side release verifier requires:

exactly 27 pre-test and 27 post-test response rows using the canonical response schema;
response-derived scores agreeing with student.json;
each response joining exactly once to the appropriate question bank;
stable, unique task and anonymous student IDs;
complete arm-required files and coherent question, prompt, provider, model, lesson-plan, practice, trajectory, feedback and telemetry joins;
chronological phase timestamps and valid practice-problem references;
cost-table membership matching the retained main AI cohort;
geography only as anonymous aggregates in 6_population_demographics/, with no per-session country/state, city, raw IP address or geography membership links;
no names, email addresses, raw participant/account/tutor/recording IDs, IP addresses, user agents, identity mappings, credentials, source caches, progress logs, internal manifests, .complete, .DS_Store, or other unintended release artifacts.

Restricted identity mappings remain outside the public release. They are used only by authorized maintainers for exact identity-linked privacy and administrative checks.

Sole pre-cutoff public exception

No ordinary pre-cutoff Quant or Verbal session is released. The sole exception is the frozen, separately labeled Quant prompt-optimization pilot under 4_supplementary_and_excluded_research_data/quant_exploratory_prompt_optimization_pilot/, because it was used to select the production prompt configuration. Its 326 analyzed rows and 30 excluded rows are exploratory and never enter the main leaderboard.

Reproducibility boundary

This document defines data membership and provenance, not a paper result. Effects, confidence intervals, p-values, figures, costs and rankings must be recomputed from the finalized physical folders and studentbench_overall_parameters.json; they must not be copied from an older report.