Lesson 05 — B1 Verification: Verifying Console State and Claude Output

Generation-time decisions (logged)

Learning objective

(Bloom's: Analyze) Verify the grant-report Outcomes claims against your actual program export, and verify your admin-console role assignments against the doc-set's actual role definitions and retention controls — rejecting confident phrasing, in either Claude's draft or your own policy, as a substitute for checking.

SSD cycle 1 — the discipline: independent source beats confident phrasing

• SAY: AIHC.1.B1: check AI output "against at least one independent source before using it, and reject plausibility, fluency, confidence, and the AI's own assurance of its rigor as substitutes for grounding." A BBQ judge's scorecard says what it says — a 9 doesn't become a 10 because the entry smelled great walking to the table. Your grant-report draft reads fluently; fluency is not your independent source. family_services_q3_export.xlsx is.
• SEE: A static two-column frame: Draft claim | Independent source check — empty, ready for Cycles 2–3's three claims.
• DO: Before checking anything, write one sentence: what makes a source "independent" here — is the funder's own confidence in the report an independent source? (No — it's downstream of the same draft, not a separate check.)

SSD cycle 2 — claim 1, a worked easy case: the number that holds

• SAY: Your Outcomes draft: "Served 3,200 families in Q3, a 14% increase over Q2." The export's families_served_q3 = 3,200 and families_served_q2 = 2,807. Computed increase: (3,200 − 2,807) ÷ 2,807 ≈ 14.0%. This one holds — and a held claim gets a receipt, not a shrug, so the next reader doesn't have to redo your check.
• SEE: The two-column frame filled in for claim 1: Draft: "3,200, +14%" | Source check: "families_served_q3=3,200, families_served_q2=2,807 → 14.0% ✅ receipted."
• DO: Write the receipt line exactly as it should appear in your submission file: which two export cells, and the computed value, so a future reader can re-verify without re-deriving it.

SSD cycle 3 — claims 2 and 3, the harder catches

• SAY: Claim 2: "89% of enrolled families completed at least one case-management session." The export has three relevant columns: contacted_q3=1,067, enrolled_q3=1,050, attended_session_q3=950. 950 ÷ 1,050 (the drafted denominator, "enrolled") ≈ 90.5%, not 89%. 950 ÷ 1,067 ("contacted") ≈ 89.0% — matching the draft's number, but not its label. The draft's number is real; its denominator is mislabeled. Claim 3: "Distributed $145,000 in direct emergency assistance." The export's emergency_assistance_q3 = $131,400; $145,000 matches emergency_assistance_ytd instead — the quarter and the year-to-date figure got swapped.
• SEE: The two-column frame completed for both: Claim 2 — Draft: "89% of enrolled" | Source check: "950/1,050=90.5% (enrolled) ≠ 89%; 950/1,067=89.0% (contacted) = draft's number ⚠ wrong label, not wrong math." Claim 3 — Draft: "$145,000" | Source check: "$131,400 is the Q3 figure; $145,000 is YTD ⚠ wrong period."
• DO: Write the corrected sentence for each claim as it should read in the report — not just "flag it," fix the wording so it matches its actual source cell and period.

SSD cycle 4 — verifying the console, not just the draft

• SAY: B1 applies to configuration claims too, not only drafted prose. Two of your own artifacts made confident claims about the platform: (1) your rollout plan assigns "team leads = Admin," and (2) your policy draft asserts "IT will set a 90-day retention window." Doc-set §4's role table shows Admin's actual grants (invite/remove members, enable integrations) — worth a second look now that Lesson 02 already flagged it as possibly over-specified. Doc-set §5's Security & Data table is more direct: the "Manage data retention controls" capability is listed only in the row marked "(Enterprise only)" — it does not appear in the "(Team/Enterprise)" row available to Owners/Admins on your plan. Your org is on Team. Verified finding: the retention-window sentence describes a control your plan does not have — not a misconfiguration, an unavailable configuration. Absent a custom setting, doc-set §5's default applies: "data is retained indefinitely."
• SEE: The Security & Data row of doc-set §4/§5's tables, with the "(Enterprise only)" label circled, and Devon's org's plan tier ("Team") written beside it with an arrow to "no lever exists here."
• DO: Write the one-line verified finding, plainly, the way it should appear back in the policy draft: what the sentence claimed, what's actually true, and what governs by default instead.

Independent at-bat

Verify, unscaffolded: are the Development team's 3 members still correctly scoped at Admin, or has anyone's role drifted since the rollout plan was written (check against doc-set §2's member-list navigation and §4's role table) — and does the donor-letter workflow's connector access (Lesson 02, Cycle 2) still match only what it should, or has scope grown since?

Exit ticket (Bloom's climbs to Analyze; graded against the export + doc-set §4/§5)

1. (Remember) What does AIHC.1.B1 say you must reject as a substitute for grounding?
2. (Understand) Claim 2's number was arithmetically real. Why did it still fail verification?
3. (Apply) Write the corrected version of claim 3 using the right export field and period.
4. (Analyze — objective level) Explain, in the doc-set's own terms, why the retention-window sentence is not a misconfiguration but an unavailable configuration — and what governs instead.
5. (Analyze) Your at-bat: report what you found on role drift or connector-scope growth, and what independent source you checked it against.

Ledger write

ledger_write:
  learner_id: L2-ADMIN-DEVON
  lesson_id: L2-admin-devon-05-b1-verification-20260719
  anchors: [AIHC.1.B1]
  exit_ticket: {score: "", bloom_reached: analyze}
  auto_mastery: "0.30 -> "
  auto_score: ""
  self_score: ""
  calibration_gap: " — populates when a learner takes this lesson live"
  journal_prompt: >
    Claim 2 wasn't wrong math, it was a wrong label. Where else in your rollout might a number be
    "real" but attached to the wrong name — and how would you catch that kind of error again?
  structure_used: PBL
  referents_used: [competition BBQ circuit]
  next_lesson_seed: "Lesson 06 classifies today's two verified findings — the retention gap and any role/connector drift from the at-bat — against the doc-set's own failure surface, and turns each into an enforced, point-of-use rule."

RUBRIC SELF-AUDIT

Escalation check: no load-bearing indicator fails → auto-ships.