L6 — UAT & Closure
Did It Actually Do What L0–L5 Claimed?
Every earlier layer labeled its numbers Actual or Assumption. L6 is the check on that discipline itself — re-verifying the claims against what's actually embedded in the L2–L4 code and output, logging what didn't work the first time, and closing the project against the criteria set back in L0–L2.
Source: L2/L4 code comments + KPI labels, re-checked directly
Scope: 6 test cases + 1 defect — limited to what's verifiable, not padded
Built: Aug 2026
This is not a full regression suite. It's an honest re-verification of the specific claims L0–L5 made, using only evidence that exists in the built artifacts — six test cases and one defect, not an inflated list.
This is a self-directed case study — there's no separate business stakeholder to hand this off to for acceptance testing. Every test below was designed and run by the same person who built L0–L5, wearing both the "builder" and "acceptance reviewer" hat. That's a real limitation on independence, and it's stated here directly rather than implied away by formal-sounding UAT language.
To keep that limitation from becoming an excuse for weak testing, the test cases below are restricted to claims that can be checked two ways: against a code comment or label already embedded in L2–L4 at build time, and against a fresh recalculation done while writing this page. A test only counts as run if both checks agree.
What this UAT is not
Not a full regression suite, not independent QA, not a claim that every corner of L0–L5 was tested. It's a re-verification of the specific numeric and behavioral claims those layers made, using evidence that already exists in the code and output — nothing was tested that isn't traceable to something built.
Each row below traces to a specific line in L2/L4: a SQL comment, a KPI card label, or an AI-package field. "Expected" is what that layer claimed; "actual" is what re-checking the artifact directly shows.
| id | test | expected | actual | result |
| TC-01 | Row-count reconciliation, L4 vs. L2/L3 | 1,618 total · 622/614/255/83/30/14 by type | Matches — per verification comment in l4_risk_assessment query | PASS |
| TC-02 | Business-category rollup sums to total | 4 categories sum to 1,618 | 735 + 614 + 255 + 14 = 1,618 | PASS |
| TC-03 | Financial exposure reconciliation | R$271,891.01 total | Matches L4's "Actual — verified vs. L2/L3" KPI card | PASS |
| TC-04 | Metric-name disambiguation | L4's 98.37% must not be confused with L2's 98.85% | KPI sub-label explicitly states "distinct from L2's 98.85% amount-match rate" | PASS |
| TC-05 | NULL order_total not silently treated as confirmed value | AI package must flag the gap, not substitute payment_total | 2 affected orders — recommended_action explicitly says "do not rely on payment_total alone as confirmation" | PASS |
| TC-06 | Zero-exposure tie-break determinism (14 APPROVAL_GAP orders, all R$0) | Tier assigned by purchase-date order, not fabricated dollar variance | Package text for all 14 explicitly attributes tier to the purchase-date tie-break | PASS |
6 of 6 pass. Every failure mode a portfolio reviewer would probe first — do the totals actually add up, does a similar-looking percentage get conflated with a different one, does a missing field get glossed over — was checked directly against the artifact, not asserted.
| id | found in | severity | description | fix |
| DEF-01 | L2 — exception classification | MED | CASE WHEN branch order placed amount-based flags before business-status flags; since CASE WHEN resolves at the first true condition, canceled-with-payment orders were being misclassified before the cancellation check ever ran | Reordered branches so business-status flags evaluate first; financial exposure for canceled orders recomputed with SUM(payment_total) directly instead of the reconciliation-difference column |
Why this is the only entry
This defect was caught during L2's own build and fixed before L2 was marked done — it's included here rather than hidden because a defect log with one honest, real entry is more credible than one with several invented to look thorough.
| criterion | target | actual | result |
| Reconciliation match rate | ≥95% (set in L2) | 98.85% amount-match rate | MET |
| Row-count integrity L2→L3→L4 | Exact match, no drift | 1,618 confirmed at every layer (TC-01, TC-02) | MET |
| No AI write-back to system of record | Zero write-back code paths (L4 design constraint) | Validation buttons log a decision string only; no ledger-write path exists | MET |
| Every financial claim traceable to L2/L3 source | No unlabeled or invented figures in L0–L5 | All Actual-labeled figures re-checked in Step 02 above | MET |
Signed off byArthunya Kanoklertwongse (self-certification — sole contributor, see Step 01)
Scope certifiedL0–L6, O2C Reconciliation Modernization
Basis6/6 test cases pass, 1 defect logged and resolved, 4/4 acceptance criteria met
DateAug 2026
The original Gate 1 target was L0–L2 complete plus a first wave of job applications sent by 31 August 2026. All seven layers — L0 through L6 — are complete ahead of that date, which is worth stating plainly without overselling it: the extra layers exist because the case study kept surfacing real BA/BSA-relevant questions (governance, ROI honesty, acceptance testing) worth answering, not because the original scope required it.
| layer | what it actually proved |
| L0 | Understanding the business problem before touching a tool |
| L1 | Designing a process on paper before writing a query |
| L2 | Turning business rules into verified SQL — and catching a real branch-order bug along the way |
| L3 | Communicating findings through a dashboard, and treating a tooling pivot (static HTML over Power BI publish) as a deliberate choice, not a failure |
| L4 | Deciding where AI belongs in a workflow, and designing a system where it structurally can't overstep that boundary |
| L5 | Separating what's measured from what's assumed in a business case, instead of blending them into one impressive number |
| L6 | Applying that same honesty to testing itself — six real checks beat thirty invented ones |
Closed. This case study set out to show how a Business Analyst redesigns a financial review process end-to-end — discovery through UAT — on public data, with every claim either verified or explicitly labeled as an assumption. L6 is the layer that checks whether that discipline actually held up, and it did.
★
Lesson: The same instinct that made L5's R$199.52 more credible than R$271,891.01 applies to testing — a short, honest test matrix that's actually traceable to the code beats a long one that isn't. In an interview, "here are the six things I checked and how" holds up to follow-up questions in a way a padded list doesn't.