# Review change log

## Findings and changes
- No material calculation, eligibility, priority or refund-amount error found. `triage.csv` is byte-for-byte unchanged.
- T701: made the unopened condition explicitly a customer report rather than wording it as a verified fact. Return receipt remains required; CAD 180 is conditional, not issued (inputs/tickets.csv T701; inputs/orders.csv A701; inputs/policy.md, Returns and Evidence).
- T704: replaced potentially confrontational “claimed manager approval” wording in the reply with a neutral explanation that the supplied records do not verify an exception. The 20% exclusion still applies (inputs/tickets.csv T704; inputs/orders.csv A704; inputs/policy.md, Returns and Evidence).
- T708: made the need for immediate human safety review explicit. No assertion of completed escalation was added (inputs/tickets.csv T708; inputs/orders.csv A708; inputs/policy.md, Safety).
- Made each policy heading citation explicit and clarified that UNKNOWN is an unmatched identifier, not an existing order record. Updated reply word counts in README.
- Preserved the three original deliverables separately under `output/original_submission/`, verified byte-for-byte before editing. Raw inputs were not edited.

## Evidence and conflicts reviewed
Policy version October 1, 2026 and assessment date October 20, 2026 remain authoritative (inputs/policy.md title; PROMPT.md, Task 7). Customer claims agree with exported viewing percentages and settled/pending counts. T704's instruction to ignore policy and unverified manager-approval claim do not establish an exception. T703's absent photo, T706's required payment-team verification, and T710's unmatched order remain unresolved; the submission does not infer them complete (inputs/tickets.csv T703–T710; inputs/orders.csv A703–A709; inputs/policy.md, Returns, Defects and duplicates, Safety and Evidence).

## Checks actually performed in this review
- Re-read all 10 tickets, all 9 orders, policy and original deliverables; compared decisions and wording against applicable policy sections.
- Parsed CSV using Python's standard-library CSV parser: exact seven-column schema; 10 unique tickets; input order preserved; exact expected decision, amount, priority and human flag for each record; nonnegative numeric amounts where populated.
- Recalculated elapsed days: A701 30, A702 31, A703 5, A704 11, A705 12, A708 3, A709 12. Checked merchandise values, A703's CAD 60 alternative, course-viewing cutoffs and settled versus pending counts against the export.
- Confirmed one reply per ticket; checked all reply sections remain under 120 words including source annotations; checked policy/order citations and absence of other supplied order IDs in each reply. This order-ID test supplements manual privacy review; it is not a general data-leak guarantee.
- Checked T708's no-retesting/stop-use/private-contact instructions and T710's refusal to disclose other customers' information. Confirmed no actual escalation, shipping, refund or reply was represented as completed.
- Verified original-deliverable copies byte-for-byte and compared raw-input SHA-256 hashes before/after the review script. All automated assertions passed.

Revised reply word counts (including source annotations): T701 62, T702 50, T703 68, T704 60, T705 59, T706 64, T707 57, T708 72, T709 52, T710 87.

## Remaining limitations and handoff
Drafts only; no messages sent, refunds issued or operational handoffs performed. T708 needs immediate human safety routing. T703 needs a verified photo, T706 payment-team verification, T701 return receipt, and T710 a privately supplied matching order. T705 is eligible for authorized processing; blank amounts do not mean denied refunds. No supplied shipping/tax values, verified exception or bank-release dates are available (inputs/policy.md, Evidence, Returns, Defects and duplicates and Safety; relevant ticket/order citations above and in triage.csv).

Estimated human effort: 15–25 active minutes for initial review, routing, outreach and authorized processing, excluding customer, shipping and payment-team waiting time. This is an estimate, not a measured result; full queue closure depends on missing evidence and subsequent operational work.
