Methods · Metacognition & self-regulation
E2
Error pattern mining
Run your accumulated error log for the patterns that only appear at scale.
When to use it
- You have weeks or a term of error-log entries and weekly reviews keep finding only local problems.
- You fixed something, logged it, and keep making it again, and you need to see that persistence.
- You suspect the same wrong move is wearing different topic labels.
What to ask AI
Paste your accumulated log and say: ignore the topic labels; group my errors by the underlying cognitive move that went wrong.
Ask the persistence question: which misconceptions did I log, work on, and then make again?
Ask whether errors cluster by condition (time of day, position in the session, pressure) rather than by content.
How it helps
Weekly error review sees each mistake as a topic problem. A term of data can show one habit wearing three coats, sign handling in algebra, direction in vectors, sense in equilibrium.
The highest-value finding is persistence after correction: you logged it, addressed it, and are still making it. That usually means the real gap is a level below the symptom.
Conditions matter too. Errors that cluster late at night, under time pressure, or on the second half of long questions are not fixed by more content revision.
Error-type drift is a progress signal: early “didn't know” should shift toward “misapplied,” then “slipped.” If yours never shift, learning is not progressing the way it should.
How to run it
- Wait until you have enough entries for scale, roughly a term, not three weeks dressed up as mining.
- Paste the full log and ask for clusters by cognitive move, not by chapter title.
- Run the persistence question separately: logged, worked on, made again, for each, was the fix shallow or was the real gap lower?
- Tag a sample by time of day, session position, and task type; ask whether condition beats content.
- Compare early-term versus late-term error types. Demand which entries any proposed pattern does not cover.
- Convert one cluster into a concrete fix and a re-test, mining without action is description.
Example exchange
Here's my full error log for the term, about forty entries across calculus and physics. I want clusters by what went wrong in my head, not by topic name.
Before I group anything: does every entry have a 'what I believed' field filled in?
Most do. A few early ones just say 'wrong answer.' I can skip those.
Skip the thin ones. From the rest, I see two clusters: skipping a second factor or inner function, and rushing setup under time pressure. A few leftovers don't fit either. Which cluster do you want a re-test set for first?
The inner-function one. That keeps biting me.
Pick five leftovers that look like that cluster. Solve them closed-book this week. We'll only call the cluster fixed if those clear.
Copyable prompt
Here's my full error log for the term. Ignore the topic labels. Group my errors
by the underlying cognitive move that went wrong, not by subject. What clusters?
Then: which misconceptions did I log, work on, and make again?
LOG:
[paste]The Tell
If your mining finds the same pattern every term and the remedy is always “be more careful,” you are describing a permanent condition, not fixing a cause.
Carefulness is not a remedy. Name the wrong rule or the lower gap, then re-test it.
Principle evidence
Strength of the underlying learning idea, not a claim about AI products.
Moderate as a principle: mining logs for patterns is formative diagnosis plus knowledge tracing. It only works if the log named causes, not just wrong answers. Overfitted “patterns that explain everything” are a known failure mode.
AI delivery evidence
Whether an AI tutor delivers this method well is a separate question.
AI pattern claims can overfit a noisy log. Treat clusters as hypotheses. Demand the entries that do not fit, and verify with new attempts rather than accepting a flattering story.