Lesson 05 — B1 Verification: Certifying Claims Before They Ship

Learning objective (Bloom's: Analyze): Analyze every factual claim in an AI-assisted passage of "The Last Mill" against your interview notes and outside sources — including whether any claim falls past a model's knowledge cutoff — to certify exactly what can ship to your editor.

Standard: AIHC.1.B1 (Verification), Band 1. Touches X1 (trust as evidence, not confidence) and X8 (convergence is not validity — Claude agreeing with itself twice is not a check).

Incoming mastery (from Lesson 01): AIHC.1.B1 = 0.45 — "verifies sporadically (journalist instincts help)" — today makes it systematic instead of instinctive.

Structure: PBL. Band 1; timing-tolerance honored: 3 cycles, each a full pause point.

Referent (flavor only, not load-bearing): in the IPL, the third umpire doesn't re-ask the on-field umpire whether he's confident — he goes back to the replay, the actual evidence. Verifying a claim means going back to your actual interview tape or notes, never re-asking Claude how sure it sounds.

Trust boundary named up front (R13): this lesson leans on knowledge-cutoff facts that are the single most volatile entries in the doc-set — re-check them on the freshness beat before trusting them past today [doc-set frontmatter; S15].

Cycle 1 — Claim by claim, not paragraph by paragraph

• SAY: Verification happens one claim at a time, checked against your actual interview notes — receipted, no source, or source-disagrees. Not a vibe-check of the whole paragraph.
• SEE: A described two-column view: five real claims from her draft, left; matching (or missing) fields in her interview notes, right; one flagged red for no match.
• DO: Pull five real claims from your current draft passage. Check each against your actual interview notes and tag: receipted / no source / source disagrees. (Evaluation: tags checked against whether a matching note genuinely exists, not against confidence.)

⏸ Pause point. Banked: five real claims, tagged.

Cycle 2 — The cutoff check

• SAY: A claim can be fully receipted in your notes and still be a verification risk if it's a present-tense claim — "the mill's parent company still owns the site" — because Claude's own knowledge has a hard stop. Per the doc-set: Sonnet 5, Fable 5, Opus 4.8, and Opus 4.7 have a January 2026 cutoff; Sonnet 4.6 and Opus 4.6, August 2025; Haiku 4.5, July 2025 [S15] — "these models may not be aware of events or information that occurred after their respective cutoff dates" [S15]. The fix for a present-tense claim is web search, toggled via the slider icon in the chat input [S09] — not asking the model to try harder from memory.
• SEE: A described flag on one real present-tense claim in her draft ("current owner," "still in operation") tagged cutoff-risk, next to the actual model-cutoff table from S15.
• DO: Find any present-tense/"as of now" claims in your real draft passage. Name which model you're using and its cutoff date [S15]. For any cutoff-risky claim, toggle web search on and check it fresh [S09]. (Evaluation: checks a real cutoff date was named correctly for the model in use, and that web search was actually toggled, not assumed.)

⏸ Pause point. Banked: cutoff-risk claims flagged and checked.

Cycle 3 — Convergence is not validity

• SAY: If you ask Claude the same question twice and get the same answer, that is not verification — that's the model agreeing with itself. The source is your interview notes or an outside document, never Claude's own repeated confidence (X8).
• SEE: A described contrast on one real claim: (a) asking Claude again — same answer, false reassurance; (b) checking it against her actual primary source — a different or confirming result either way.
• DO: Pick one claim you're tempted to just re-ask Claude about. Instead, check it against your actual primary source, and note whether the two methods would have agreed or not. (Evaluation: checks the primary source was actually consulted, not just re-asked.)

⏸ Pause point. Banked: one real claim, checked the right way. Three cycles complete.

Independent at-bat

Verify the remaining claims in that same real passage independently — no checklist columns given this time; build your own tally as you go.

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

1. (Remember) Name the three verdicts a claim can receive when checked against your interview notes, and give one real example of each from today's session.
2. (Understand) Why can a claim be fully receipted in your notes and still be a cutoff-risk? One sentence.
3. (Apply) A claim in your draft says "the mill has sat empty since 2019." Your notes confirm
4. Is this claim fully cleared, or does it still need a second check? Explain.
5. (Apply) Claude gives you the same answer twice when you ask about a fact with no source in your notes. Is that verification? What should you do instead?
6. (Analyze — objective level) Take the full claim set from today's passage: sort every claim into one of the three verdicts, flag any cutoff-risk claims separately, and state which claims are cleared to ship to your editor today.

Ledger write

ledger_append:
  learner_id: L3-CONS-PRIYA
  lesson_id: L3-cons-priya-05-B1-verification
  standards:
    - {standard: AIHC.1.B1, mastery_before: 0.45, mastery_after_simulated: 0.68}
  bloom_reached: analyze
  auto_score: ""
  self_score: ""
  structure_used: PBL
  referents_used: ["sports — cricket (IPL)"]
  journal_prompt: >
    Which claim almost shipped before you checked it today — and was the catch a source
    mismatch, or a cutoff-risk you hadn't thought to name before this lesson?
  next_lesson_seed: "cutoff-miss pattern flagged here becomes Lesson 06's worked failure case"

RUBRIC SELF-AUDIT

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