Blind 75 Sweep — Gap Audit

Review boundary and corrected contracts

Learning objective: Give each of 75 problems a 60-second approach-retrieval attempt, preserve the raw answers, then validate them and rank the patterns needing follow-up. This is a classification audit, not 75 implementations and not a readiness score.

A fast, confident answer may be wrong. Labels below therefore separate provisional retrieval status from validated correctness. The time boxes, ranking weights and review intervals are author-chosen planning rules, not calibrated hiring predictions.

Series: LeetCode — From Basics to Production-Ready · Session 59 / 65 · Phase 4 · Design review simulation

Watch first

If you want a taxonomy refresher, use it on a separate study day and record the exposure. During the sweep, hide topic labels and avoid solution material. No particular overnight delay guarantees an unprimed measurement.

The original optional resources were freeCodeCamp's problem walkthrough, CS Dojo's problem-solving tips and freeCodeCamp's DP course. Their historical labels and links are retained as optional resources, not freshly verified content, duration or effectiveness claims. Watching a solution beforehand makes that problem a retrieval check on known material, not an unseen classifier test. The sweep does not require watching anything.

Why this session exists

Choosing the next exercise from a vague feeling can overlook gaps. A table makes the actual observations inspectable: which approach you produced, its invariant, its cost, whether it was prompted and whether it was correct.

The key distinction is stating an approach versus implementing it. A 60-second answer can expose a missing state definition; it cannot show that you will implement without bugs, satisfy every boundary or communicate well in an interview. Keep the L57/L58 evidence beside this audit rather than treating one score as a replacement for the others.

A mixed list samples several families, but 75 is not a statistical guarantee and the chosen collection may overrepresent some patterns. A large failure count can produce a useful repair plan; a small one still needs validation.

Blank-file warm-up

No template warm-up. Spend about three minutes preparing a sortable table; allow more if importing the list takes longer. Record the exact list/version and a saved mixed order. Many published lists are grouped by topic, so published order is not necessarily mixed order. Shuffle before opening statements, hide topic headings/tags and save that order for reproducibility. Shuffling reduces consecutive topic cues; it does not make familiar problems unseen or observations independent.

Use these columns:

id · problem · named pattern · approach + invariant + complexity · provisional status · seconds · prompt/exposure · validation + acceptable approach · gap note

Prefill all 75 IDs/titles. Leave validation blank until after the cold sweep. Do not replace your raw answer with the editorial answer.

Pattern anatomy

Maintain two process invariants: every attempted problem has a row, and every row records the timing/assistance condition under which it was produced.

Each 60-second slot:
0–15s   Read the statement/constraints; examples are allowed if needed for meaning.
15–45s  State a pattern, concrete approach/invariant, and time/space cost.
45–60s  Save the actual answer and provisional status; move on.
 
Provisional retrieval labels:
cold    complete unprompted answer within 45s, without substantial hesitation
warm    complete unprompted answer by 60s, but slower or hesitant
hint    a tag, solution hint or other outside cue was used; record which
failed  no complete unprompted approach by 60s
 
Do not code or open editorials during the sweep.
Statement examples are part of the problem, not automatically a hint.
An unfinished answer at the beep is failed, not warm merely for being close.
After the sweep validate correctness; fast confidence is not correctness.

Use a repeating timer or log actual elapsed time. If you pause for a break, stop the session clock and record the break; do not silently give one problem extra time.

The cue

Run this audit when you have completed broad study but lack a repair plan, when practice choices feel arbitrary, or when a mock failure has no clear diagnosis. Revisit it when old data no longer seems representative, not because a month is a scientifically fixed expiry date.

If you already know that a specific implementation is broken, a focused regression exercise may be the better next step. A sweep is useful for breadth; it is not mandatory before every repair and cannot substitute for learning an unfamiliar topic.

Guided solve

Cold pass — 75 minutes of active slots. Work in the saved mixed order. Preserve partial answers and uncertainty. Do not fill the “acceptable approach” column from memory as if it were ground truth.

Validation pass — after the sweep. Compare the stated invariant and cost with a reliable solution or your own proof/counterexample. Record valid, invalid or unresolved, the evidence used and acceptable alternatives. A different pattern name is not a misclassification if the algorithm meets the contract. If 20 minutes is insufficient to validate 75 rows, finish later and leave remaining rows unresolved; do not fabricate completed validation to fit the session budget.

Read the table in three passes:

1. Aggregate by pattern with denominators. Report, for example, four incomplete attempts out of five, not just “four failures.” A single failure out of eight might reveal a real variant gap; it is not automatically noise.
2. Find invalid approaches. Mark a misclassification when the chosen rule cannot solve the contract, not whenever it differs from an editorial tag. Count self-corrected wrong turns separately so the final valid approach is not labelled wrong.
3. Find false confidence. A provisional cold row that fails validation stays in the raw record but enters the repair queue as invalid. Confidence is a measurement about the attempt, not certification.

For a concrete classification check, consider sorted two-sum with target 8 on [1, 2, 4, 6]. Endpoints sum to 7, so move the left pointer: every pair with the leftmost 1 is at most 7. The next pair, 2 and 6, succeeds. This discard argument—not merely “sorted, so two pointers”—justifies the move. An unsorted array does not support that same elimination proof.

Rank a short weak list. One transparent rule is F + 2M. Among unassisted, resolved rows, F counts attempts with no complete approach and no assessable wrong algorithm; M counts assessable invalid chosen approaches. If an incomplete response also contains a checkably wrong algorithm, count it in M only. Treat hinted and unresolved rows as separate review needs. Keep the numerator, denominator and failure notes next to the score. The factor 2 is a prioritisation choice, not a measured risk ratio.

Example: family A has F=2, M=1 out of five attempts, score 4; B has F=3, M=0 out of ten, score 3. A ranks first under this rule, but B still has three concrete failures. Break ties using the needs of your stated practice plan or prerequisite dependencies, not unsupported claims about employer frequency.

Choose at most five families, with three for the next L60 session. Attach the exact failed examples and missing mechanisms to each. Keep the raw table for comparison in L65.

Solo timed

Seventy-five slots are the main timed block. Useful checks, regardless of order:

• Arrays/hashing: does the proposed method meet the constraints, or did a familiar title hide a quadratic scan?
• Binary search/two pointers: state the monotone predicate or discard argument. A label alone is incomplete.
• Trees/graphs: distinguish representation, traversal, state and termination. Do not assume this will be your weakest family.
• Dynamic programming: state what a DP cell means and the transition. There is no expected 50% failure rate for this exercise.
• Fatigue: if attention deteriorates, stop at a slot boundary and finish the remaining rows later with the same 60-second rule. Record the split.

Timing rubric

The initial plan is 98 minutes, not a 75-minute session including analysis. At active minutes 25 and 50, expect 25 and 50 recorded slots. Falling behind may mean overrun, navigation or reading difficulty—not necessarily solving. Keep the same slot length when resuming; shortening to 40 seconds midstream would change the measurement. If you intentionally use a shorter audit, label it a separate condition and do not combine its rates with the 60-second run.

Self-scoring checklist

Give 0–2 for each process item: all 75 rows accounted for; consistent slot rules; no unrecorded assistance; raw and validated approaches kept separately; and a ranked, evidence-backed weak list. Use 2 for complete, 1 for partial and 0 for absent. Unresolved validation makes that item partial, even when all cold-pass rows exist.

Author-chosen bands of 8–10, 5–7 and 0–4 can help decide how much cleanup is needed. They do not establish that the data is unbiased or that the learner is ready. Never award points for having few gaps; the purpose is an honest record.

Performance criteria

Report observed rates instead of invented cohort benchmarks. Separate complete unprompted answers, validated correct approaches, hinted attempts, unresolved checks and implementation samples. State each denominator; validation of only ten rows does not establish correctness for all 75.

Look for clustering, but use it as a hypothesis: repeated missing DP states suggests a state-definition exercise; scattered errors might reflect several gaps, reading load, fatigue or timing. They do not uniquely diagnose general retrieval speed.

Each gap needs a specific note, such as “could name BFS but could not justify marking visited on enqueue.” Test three provisionally strong rows by actually implementing them. A slow solve identifies a possible implementation gap; it does not retroactively prove that every classification answer was wrong. Keep the two measurements distinct.

Recovery scripts

Want to solve one now: “Interesting gap; saved for the repair pass. Next slot.”

Ten rows behind: stop at a boundary, record the cause and remaining IDs, and resume with the same slot length. Do not accelerate the remainder to make the spreadsheet look complete.

Discouraged by the count: “These are observations to validate, not a verdict.”

Cannot distinguish recognition from retrieval: write the actual approach without looking at a solution and mark uncertainty. Familiarity alone is insufficient, but uncertainty does not mean the approach is necessarily wrong.

Too many weak families: rank them and choose the next three, keeping the rest visible. Deferring a repair is not declaring it unnecessary.

Common failure modes

Solving during the sweep changes its purpose. Recording only a status loses the diagnostic mechanism. Topic order and visible tags provide cues; mixed order and recorded exposure make those limitations clearer. A fast wrong answer needs correction even when its confidence felt good.

Misconception: “I must re-solve every forgotten problem before choosing what to practise.” A sweep can cheaply identify candidate gaps, while full solves test execution and edge cases. Use both according to the question you need answered. Worked problems can transfer learning when their underlying rules are understood; neither “individual problems transfer nothing” nor “one pattern transfers to every variant” is justified.

Think of the table as a map of observed attempts, not a diagnostic blood test with calibrated thresholds. It helps direct attention but cannot establish that your weakest family determines hiring outcomes. If strong familiar-list recall conflicts with a poor unseen mock, check exposure, correctness, implementation and task difficulty before deciding which practice to add.

Complexity

The sweep itself costs 75 × 1 minute = 75 minutes. At an assumed 15 minutes per full solve, 75 implementations cost 1125 minutes = 18.75 hours. The sweep-only ratio is 75 / 1125 = 1/15, or about 6.67%, not 1.5%. Including the initial 3-minute setup and 20-minute analysis gives 98 / 1125, about 8.71%, before any extra validation. These are comparisons of assumed budgets, not measured equivalent learning outcomes.

A four-to-six-week re-audit can be a planning default, adjusted to your schedule and delayed checks. Keep list identity, assistance and timing conditions so that later comparisons are interpretable.

Self-check (answer follows each question).

Q: What does the sixty-second slot measure?

A: Whether you can produce a concrete approach, invariant and complexity under that condition. It does not establish implementation ability; correctness is validated afterwards.

Q: Why give misclassification extra priority?

A: An invalid confident approach can hide a conceptual gap. Double weighting is this audit's heuristic, not proof that it costs twice as much or that every wrong turn wastes fifteen minutes.

Q: Why use a saved mixed order rather than topic groups or blindly trusting list order?

A: Consecutive topic cues can help later answers. A shuffled order with tags hidden reduces that cueing, but familiar problems remain familiar and samples are not guaranteed independent.

Q: Why organise repair by pattern rather than only by problem?

A: Repeated failures of the same mechanism can be addressed together. Keep the concrete problems as counterexamples and test transfer to distinct variants; do not dismiss an isolated failure or assume universal transfer.

Q: Why defer editorials until after the sweep?

A: They provide the answer or cues you are trying to retrieve. If consulted, record the exposure and preserve the earlier response; the row becomes an assisted attempt, not useless data.

Q: What if 70 rows are cold and five warm but L57 went badly?

A: First validate the supposedly strong answers and check prior exposure. Then sample implementations and compare the mock's phase log. Familiar-list retrieval, unseen classification, implementation and narration can differ. Choose the repair from that evidence rather than ruling out retrieval automatically.

You can now… run a consistent sweep, preserve raw answers, distinguish retrieval from correctness, validate alternative approaches, maintain pacing without changing the condition, identify candidate clusters and produce an actionable weak list with concrete notes.

Spaced queue

Schedule from validated observations, not confidence alone. Suggested defaults are cold +60 days, warm +21, hint +7 and failed +2 then +7 after success. Invalid and unresolved approaches need prompt checking even if the provisional label was cold. These intervals are conventions; adapt them to subsequent performance.

Keep the original table, corrections and dates. L65 can compare status changes under stated conditions, while new problems and full implementations test transfer and execution. Movement up this table is one measure of progress, not the only honest one.

Corrective source notes

The saved 2026-10-07 APS learning-techniques review summary was re-read. It supports practice testing and distributed practice in the reviewed educational settings, not this audit's thresholds, double weighting, calendar or hiring predictions. Timing ratios and the sorted two-sum discard argument are derived above; no measured learning advantage is inferred from cheaper timing alone.