Analysis & methods

The peak day is 2026-08-29: 233 submissions from 14 students. The busiest publicly visible four-hour cell is 2026-08-29, 12:00–16:00 Beijing time: 153 submissions from 12 contributors (12.75 per active contributor).

HW1: median 12.5, range 8–98; HW2: median 13, range 3–62; HW3: median 14, range 7–108; HW4: median 12, range 8–141; HW5: median 14, range 1–209. Only window submitters enter each distribution.

13 of 43 problem slots retain a mean gap above 10 percentage points after three attempts. HW3 26s3HW3-EvenOrOdd has the largest gap: 66.15 through three versus 100.00 best observed, a 33.85-point difference among the same 39 attempters. This suggests a question for reviewing feedback; it does not establish why the problem was difficult.

HW1's actual mean changes from 76.93 on attempt 1 to 64.53 on attempt 2. Event counts change from 304 to 106; distinct students from 38 to 34. Only pairs reaching a later attempt contribute to that point. A lower mean does not mean the same group got worse. HW1–HW5 have separate panels.

Among eligible retries: HW1 32.3% (95/294); HW2 45.0% (125/278); HW3 34.6% (119/344); HW4 43.2% (136/315); HW5 63.2% (72/114). These are event rates; students can contribute multiple retries.

Each window is split into five equal-duration phases. Color shows the share of eligible retries setting a new best; each cell reports new-best counts, eligible retries and distinct contributors. Different phases can contain different students, problems and prior scores.

The available midterm scores have a mean of 90.02 and median of 93.25. Canvas lists a nominal maximum of 100; values above 100 are preserved as raw points. No grade is capped or inferred.

26 matched students from a 41-member eligible roster; 15 grades were not in the supplied excerpt. This is a partial, non-random cohort, not the class grade distribution.

In this excerpt, final-day submission share has the clearest negative rank association with exam scores (ρ = -0.48). Its leave-one-out correlations range from -0.60 to -0.42, while the all-23-problem and discrepancy-excluded subsets give -0.43 and -0.47. All three leave-one-out ranges retain their original signs, yet the submission-count and mean-first-score bootstrap intervals cross zero. Stability to one-record deletion does not imply a precise estimate. These checks do not remove selection bias or confounding.

All selected homework windows end before the user-confirmed exam start.

What the three behaviors mean

• Submission count: all submissions inside each configured HW1–HW3 window, including attempts after acceptance.
• Mean first-submission score: first normalized score per attempted assignment–problem slot, then the mean within each student. Unattempted slots are not imputed as zero. Coverage varies, so a separate all-23-slots sensitivity check is shown.
• Final-24h share: the percentage of a student's in-window submissions made in the final 24 hours before that homework's configured deadline. It does not measure time spent studying or account for individual extensions.

A first submission can follow off-platform work, and students may submit drafts at different stages. More submissions do not necessarily mean more effort; lower initial scores do not establish lower prior ability. The signs here are exploratory and can reflect cohort selection, problem coverage and different submission practices.

Group means are 81.00, 95.27 and 93.92 for under 35, 35–59 and 60 or more submissions. The middle group has the highest mean, while the highest-count group has the highest median. A single summary cannot establish that additional attempts improve an exam score.

All 26 available grade records have a unique normalized-name match to a Course Pulse student account. Of those, 11 have recorded pre-exam platform activity, including nine with resource-open events. The other 15 have no observed activity in this tracking window; this does not establish that they did not study.

Personal tracking began on 8 September at 13:49 Beijing time. Main analysis ends on 16 September at 08:00, four hours before the exam, because metrics are aggregated by UTC day. OJ metrics span all three configured homework windows, so the sources cover different periods.

The 11 students with recorded activity average 95.27 points, compared with 86.17 among the 15 without a record. Their medians are 94.5 and 87.5. A 9.11-point difference between self-selected groups is not an estimate of what the platform caused.

Any positive recorded metric qualifies, including a brief visit. Students can download materials and study offline, use other tools, or differ in prior knowledge and motivation. These factors are not controlled here.

Final-day OJ submission share remains negatively associated with exam scores. Platform active time, recorded days and resource opens each show positive rank associations, but all three platform intervals cross zero. These measures do not yet establish a reliable usage–achievement relationship.

Active time estimates foreground interaction, with an idle limit; it does not measure attention or total study time. Missing platform events are represented as zero recorded events only within the instrumented window. No claim is made about activity before tracking began.

What the combined evidence supports

Similar homework endpoints can conceal different submission schedules and resource-use patterns. In this partial cohort, concentrated final-day submissions accompany lower exam scores, while recorded platform users have higher average scores. Platform intensity remains uncertain, and prior ability, motivation and off-platform study could explain these patterns.

Why AI and practice effectiveness are not plotted

The export includes assistant replies, staff activity and messages that cannot be linked to current accounts. Linked student messages are concentrated on one date, and overlap with the available grade cohort is too small for a useful public comparison. Practice records are current draft snapshots, with no submitted outcomes; they cannot reconstruct edits, completed exercises or learning gains. These sources need further validation before effectiveness claims.

The next useful evidence is the remaining grades and a longer, consistently instrumented observation period. More variables alone will not resolve cohort selection or confounding.

Only real recorded usage is included; editable reported totals, demo accounts and staff activity are excluded. Raw CSVs, names, account IDs, messages and code remain outside the public repository. Public groups have at least five students; this reduces disclosure but is not a formal anonymity guarantee.

Download platform aggregate JSON · Sources, definitions and reproduction ↗