A writing skill for your agent.

Trip.com retention evidence for hybrid-policy review

Evidence for the HR director

Bloom, Han and Liang's 2024 Nature study covered 1,612 university-graduate engineering, marketing and finance employees in Trip.com's Airfare and IT divisions in Shanghai. Interns and new employees on probation were excluded. At this single employer, random assignment by odd or even birthday dates put 792 odd-birthday employees in a group allowed to work from home on Wednesday and Friday, with three office days, and 820 even-birthday employees in a five-office-day control group. The two home days were optional. The 518 initial volunteers started in the week of 9 August 2021, and the remaining 1,094 non-volunteers in the week of 13 September.

During the experiment, quit rates by assigned group were 7.20% in control and 4.80% in treatment, with P = 0.043 for the difference. These are experiment-period rates, not annual rates. The 2.40-percentage-point difference is about a 33% reduction relative to control. The nominal six-month experiment effectively ended on 21 January 2022 when a COVID case led the company to allow everyone to work from home, although the attrition figure caption reports follow-up through 23 January. This was not two years of randomized policy exposure.

Among employees assigned eligibility, actual take-up was about 55% for volunteers and 40% for non-volunteers across Wednesday and Friday opportunities, excluding public holidays, personal holidays and excused absences. These percentages use neither all employees nor all weekdays as their denominator. Typical use was about one day weekly. The quit-rate comparison therefore estimates assignment to the optional policy, not the effect of actually working from home twice every week.

Implication for our review

My judgment is that this evidence warrants considering a comparable optional-policy trial for graduate office roles, subject to checking how similar our work and commuting conditions are. That is a policy judgment, not a research finding that another employer will achieve the same quit-rate reduction or that fully remote work is equivalent.

Quarterly profit and operating cash flow review

This fictional manufacturer reports on an unchanged scope for the standalone April 1 through June 30 quarters of 2025 and 2026, not year-to-date periods. All amounts are exact KRW million. Net income increased by 200, or 20% of the Q2 2025 base of 1,000, while operating cash flow fell by 300 but remained positive in both quarters. Net income is a supplied bottom-line input rather than a reconstructed income statement.

Reconciliation

For the balance rows, the table gives each quarter's opening and closing balances followed by the cash flow adjustment.

Other adjustments total zero by design, and no acquisitions, currency translation, write-offs, reclassifications, or other noncash movements affect these balance changes. Depreciation is a noncash add-back, not a cash receipt. Accordingly, the reconciliations are 1,000 + 200 - 100 - 250 + 50 = 900 and 1,200 + 240 - 400 - 600 + 160 = 600.

The combined working-capital adjustment moved from -300 to -840. Its additional -540 outweighed the +200 net income change and +40 depreciation adjustment change. The operating cash flow change bridge is +200 + 40 - 300 - 350 + 110 = -300, with the last three components reflecting the larger receivable increase, larger inventory increase, and larger payable increase, respectively.

Interpretation limits

Cash balances, investing and financing flows, collection terms, receivable aging, inventory plans, and profit-growth causes are not supplied. Unknown is not zero. The reconciliation does not establish overdue customers, unsold inventory, or financial distress. Operating cash flow alone cannot establish a cash-balance change or support a complete liquidity assessment.

Warehouse automation investment review

Fictional scenario. Prepared for the operations director and investment committee. All figures are pre-investment assumptions, not measured equipment performance. Money is in constant KRW million unless stated otherwise.

Four comparable zones each process 1,200 eligible cases daily for 300 operating days annually, with fixed demand and no sales uplift. Of eight scheduled hours daily, 75%, or six hours, are occupied processing time. A two-worker team processes 200 cases per occupied team-hour manually. Automation targets an unvalidated 300 with the same staffing and case mix. Processing would take four rather than six team-hours, releasing four worker-hours daily per zone, or 1,200 annually and 4,800 sitewide. Scheduled time, occupied team-hours and worker-hours are different measures.

Cash realization assumes cancellation of 600 agency hours per zone annually at KRW 30,000 per invoiced hour. This yields 18 per zone and cancels only 2,400 sitewide hours, not all released hours. Permanent payroll stays unchanged. Other released hours support redeployment without cash savings. Spend cancellation remains unvalidated. Annual gross cash savings are 18 for the pilot and 72 sitewide. Incremental maintenance, software and energy cost 4 and 12 respectively, leaving net savings of 14 and 60.

Immediate four-zone deployment costs 240 upfront. A one-zone pilot lasts one operating year and costs 75 upfront. Conditional expansion reuses pilot equipment but costs another 190, totaling 265. Spending and go-live coincide. The model assumes no ramp, constant savings and uniform cash accrual, with expansion immediately after year one. It excludes tax, depreciation, financing, discounting, working capital and salvage. Simple payback is 240 / 60 = 4.00 years for full deployment. Continuing the pilot without expansion gives 75 / 14, about 5.36 years. Conditional staged payback from initial outlay is 1 + (75 - 14 + 190) / 60, about 5.18 years. Year one earns only 14.

Expansion would require target throughput on comparable case mix, a case-error rate at most 0.5%, no unresolved critical safety findings and documented cancellation of 600 agency hours during the pilot year. Finance and operations must substantiate the remaining zones' economics before separate expansion approval. Management recommends authorizing only the 75 pilot tranche because throughput and spend cancellation remain unvalidated. Neither tranche is approved or completed. The pilot commits 165 less initially, but staging adds 25 in capital and delays modeled payback. Immediate deployment has better assumed economics but commits more capital before cash realization is proven.

Evidence request on New York congestion pricing

To the transport-policy reviewer

Reported results and missing definitions

Page 13 of MTA's December 17, 2025 board presentation, 2025 in Data, reports 11% fewer vehicle entries into the central business district, or CBD, and 71,000 fewer per day. These are entry counts, not unique cars or vehicle-miles, and the rounded figures do not establish an exact baseline count. Observation and comparison periods and baseline construction are unspecified on the slide.

MTA also reports car speeds within the CBD up 4.3%, which does not mean travel time fell 4.3%. Absolute speeds, weighting, hours covered, measurement provider, baseline, and observation and comparison periods are unspecified.

MTA reports PM2.5 down 22% in the first six months, crediting Cornell University, 2025. The slide places this claim under a Congestion Relief Zone heading without specifying exact study dates, the concentration statistic, the percentage denominator, or detailed geographic coverage. The concentration claim does not establish a 22% reduction in emissions, all pollutants, or individual exposure, and cannot be generalized citywide. The presentation title does not establish a common full-year comparison for these metrics.

Evidence requested

I treat the implementing agency's summary as a record of attributed claims, not independent causal proof of the policy's effects. For independent review, I request the traffic series and calculation methods, the missing periods and baselines, and the underlying air-quality study's geography, statistic, counterfactual method, and uncertainty. This is my request for evidence, not MTA's recommendation or a recommendation to repeal or expand congestion pricing.

ADR 01. Durable invoice-event receipt and local entitlement grants

Context

This proposed ADR is a fictional application design, not a Stripe, API, or production change. It assumes one account, one environment, snapshot invoice.created and invoice.paid events at version 2026-10-28.preview, existing invoice-to-order mappings, and one non-revocable local entitlement per invoice.

Stripe documents unordered delivery and duplicate events. Snapshot created values have second-level precision and establish neither order nor uniqueness. Repeated delivery of one Event and separate duplicate Event objects differ. Stripe recommends data.object.id with event.type for the latter, not blanket suppression of future object updates.

The fictional arrival trace is invoice.paid E2, repeated E2, late invoice.created E1, and distinct duplicate invoice.paid E3. These four deliveries contain three Event IDs and two types for one invoice and order. E1 was generated before E2, possibly within the same timestamp second.

Decision

Use a durable inbox as the work queue, without external queue publication. Store payload, receipt time, and pending or completed state under a unique (account_id, event_id) key. account_id is configured account context, not an assumed payload field. event_id, event_type, and invoice_id come from Event.id, Event.type, and data.object.id. The business-operation key is (account_id, invoice_id, operation_kind=grant_entitlement), not the Event ID.

Following Stripe's handler guidance, verify the signature against the unmodified raw body. Commit the inbox row or confirm its durable predecessor before promptly returning 2xx. Persistence failure receives no success response. Receipt acknowledges delivery, not business completion. Repeated E2 adds no work, including when its existing row is still pending. It neither discards that row nor marks it completed. A crash after receipt commit but before acknowledgment leaves the committed work available on redelivery.

One ACID database holds inbox, operation ledger, and entitlement, with locks and unique constraints for concurrent workers. Paid processing retrieves the invoice and requires confirmed status=paid and its matching existing order_id. Retrieval failure or an unmet condition leaves work pending. For a new operation, ledger insertion, entitlement mutation, and inbox completion commit atomically. An existing operation permits completion without another grant, including for E3. A transaction crash cannot commit only completion. Created processing completes its receipt without resetting entitlement, so late E1 cannot undo a grant.

Alternatives

Event-ID-only deduplication misses E3. Acknowledgment before persistence risks losing work. Timestamp sorting cannot establish order. None replaces business-operation uniqueness.

Consequences and limits

Database contention and pending-work recovery are costs. Pending rows remain retryable, and inbox and operation records are not pruned within this example. Under these database and retention assumptions, eventual successful receipt, worker retries, retrieval, paid confirmation, and order matching produce one local grant for this trace, not global exactly-once execution. Refunds, disputes, subscription changes, external fulfillment, and outbound payment API idempotency are excluded.