Lesson 06 — B2 Failure Literacy: Turning A Cutoff Miss Into A Rule

Learning objective (Bloom's: Evaluate): Evaluate a real AI-assisted mistake from your own workflow — a stale fact or a misattributed detail — to diagnose its cause and crystallize a standing rule that prevents its recurrence, rather than deleting and retrying.

Standard: AIHC.1.B2 (Failure literacy), Band 1. Touches X11 (deficit-free engineering — an unwanted output gets analyzed as an information gap the collaboration didn't fill yet, not as "the model is bad").

Incoming mastery (from Lesson 01): AIHC.1.B2 = 0.30 — mistakes get fixed in the moment and forgotten, no standing rule yet.

Structure: PBL. Band 1; timing-tolerance honored: 3 cycles, each a full pause point. This is the last lesson in the set — scaffolding here is the lightest of the six, on purpose.

Trust boundary named up front (R13): this lesson's worked case is the exact cutoff-risk pattern Lesson 05 flagged [S15] — a real, doc-set-documented failure mode, not an invented one.

Cycle 1 — Delete-and-retry vs. diagnose-and-rule

• SAY: The doc-set's own iteration advice says feedback and refinement beat starting over — "give Claude feedback: follow-up instructions, clarifications, or ask it to rewrite an answer, rather than starting over from scratch" [S17]. Failure literacy takes that one step further: fixing this answer isn't the same as writing a rule that stops the same mistake next time. Delete-and-retry fixes the sentence; it leaves the underlying risk exactly where it was.
• SEE: A described three-column log: mistake → cause → rule, one row worked as example (labeled representative, not her real case).
• DO: Recall a real mistake Claude made somewhere in "The Last Mill" project (from Lesson 01's diagnostic, or a fresh one you've noticed since). Fill in its cause using the three columns. (Evaluation: checks a genuine cause is named — "Claude was wrong" isn't a cause, it's a restatement.)

⏸ Pause point. Banked: one real mistake, diagnosed to a cause.

Cycle 2 — The cutoff-miss case, worked live

• SAY: The single most common consumer-side failure mode this doc-set actually documents: asking about something recent without checking the active model's cutoff or toggling web search. Per S15, models "may not be aware of events or information that occurred after their respective cutoff dates"; per S09, web search is the documented fix. This is a real, named-in-the-doc-set risk — not a hypothetical.
• SEE: A described side-by-side: the same present-tense question about the mill site, asked once without web search (confidently worded, no citation) and once with it toggled on (citation-backed, possibly different answer).
• DO: Find one present-tense claim about "The Last Mill" you haven't checked yet. Run it once without web search, once with the slider toggled on [S09]. Log the difference. (Evaluation: checks both runs actually happened and a real difference — or confirmed absence of one — was logged.)

⏸ Pause point. Banked: one real cutoff-miss test, logged.

Cycle 3 — Crystallizing the rule

• SAY: A crystallized rule is one sentence you can check yourself against next time — for example: "Any claim about current or ongoing status gets web search on before it ships." Writing it down is what turns a lesson learned once into a standing habit, the same way Lessons 02 and 03 turned a spec and a delegation split into standing project instructions.
• SEE: A described view of her real project instructions panel, with one new line added: the crystallized rule from Cycle 2.
• DO: Add your real crystallized rule as a new line in your project's standing instructions. (Evaluation: checks the rule is specific and checkable — "be more careful" fails; "web search on for any current-status claim" passes.)

⏸ Pause point. Banked: one real standing rule, added. Three cycles complete.

Independent at-bat

Audit your whole current draft for present-tense/ongoing-status claims using your new rule, with no scaffolding. Note what you catch now that you wouldn't have caught before Lesson 01.

Exit ticket (climbing to Evaluate — the objective's level)

1. (Remember) What are the three columns of a failure-literacy log?
2. (Understand) Why does "delete and retry" leave the underlying risk in place even after the sentence itself looks fixed? One sentence.
3. (Apply) You catch Claude stating a business's ownership as settled fact with no web search toggled on. What's the immediate fix, and what's the standing rule?
4. (Evaluate — objective level) Looking back across Lessons 02 through 05 of your own record, pick the ONE mistake pattern most likely to recur in your workflow. Defend why it's the highest-risk one — not just the most recent — and write the rule that would catch it.

Ledger write

ledger_append:
  learner_id: L3-CONS-PRIYA
  lesson_id: L3-cons-priya-06-B2-failure-literacy
  standards:
    - {standard: AIHC.1.B2, mastery_before: 0.30, mastery_after_simulated: 0.65}
  bloom_reached: evaluate
  auto_score: ""
  self_score: ""
  structure_used: PBL
  referents_used: []
  journal_prompt: >
    What's the rule you're proudest of writing this week — and which of the five practices from
    Lesson 01's diagnostic do you think needs a lesson like this one next?
  next_lesson_seed: >
    all five diagnosed anchors now have at least one dedicated lesson; recommend the next
    generation-time lesson target either B3 (trust calibration) or C1 (provenance, her
    original lowest score at 0.25, still untouched by this set)

RUBRIC SELF-AUDIT

Escalation: all load-bearing indicators PASS → auto-ship.