# Tariff Instrument — Phase 1 pre-registration (rate-as-collected)
Registered 2026-08-08, BEFORE any chart or finding is drawn from tariffs.db.

## What this instrument is
US Census national imports, HS6 × country × month, 2017-01 → present:
consumption value, dutiable value, calculated duty (regular tariffs as
estimated at entry). Effective rate := calc_duty / dutiable_value, always
derived at query time, never stored, never mixed across tables.

## Known limitations (disclosed on every response)
1. CAL_DUT excludes anti-dumping/countervailing duties and MPF/HMF fees —
   understates true burden on AD/CVD-heavy codes (steel, aluminum, solar).
2. Duties attach to imports-for-CONSUMPTION; the port ledger measures GENERAL
   imports. The two are never divided into each other. FTZ/bonded-warehouse
   timing differs by code.
3. De minimis (<$800) shipments filed no entries → invisible pre-2025 for
   China e-commerce; the 2025 removal is a REGIME BREAK, not a trade surge.
4. HS revisions (notably 2022) split/merge codes; within-code series respect
   code lifespans; long series use concordance or stop at the boundary.
5. Region rollup codes (0003, 0014, 1XXX…) are stored but excluded from
   country analysis; used only for reconciliation.

## Kill-switch (must pass before ANYTHING publishes)
K1. Reconciliation: summed CAL_DUT_MO across all real countries must match
    (a) the all-countries rollup rows within 0.5%, and (b) published
    CBP/Treasury monthly duty collections within a stated tolerance (target
    ±10%; CBP totals include AD/CVD and fees, so a stable, explainable gap
    is expected — an UNSTABLE gap kills publication).
K2. Spot-rate sanity: 20 hand-checked cells (code × origin × month) vs
    statutory expectations — e.g. pre-301 China rates near MFN; EU MFN-only
    codes near schedule rates. >2 unexplainable misses kills publication.
K3. Continuity: no month with >5% missing pulls (pull_progress 'fail').

## Registered questions (written before looking)
Q1. China's effective collected rate, all goods, monthly 2017→now: does the
    curve match the known 301/IEEPA staircase in TIMING (±1 month of
    effective dates)? Falsifier: jumps appear where no action exists, or
    actions produce no jump on covered codes.
Q2. Duty-per-kg on the CW-1 furniture codes: did duty/kg fall alongside the
    declared-value collapse within identical codes (undervaluation cross)?
    Falsifier: duty/kg flat while declared $/kg fell (would point to rate
    changes, not undervaluation).
Q3. Exclusions: on 301-excluded code lists (2019-2020), does the collected
    rate visibly drop during exclusion windows and recover at expiry?
    Falsifier: no rate response inside published exclusion windows.

## Isolation guarantees (how this cannot disturb the live system)
- Separate DB file (tariffs.db); ledger.db and trimtab.db untouched.
- Backfill runs on the Mac only; the box receives a finished .db by scp.
- Server integration is additive: new module + new tool; existing routes and
  tools unmodified; per-file node --check before every scp; rollback =
  remove one require + restart.
- The server opens tariffs.db read-only, same as every other DB.
- Nothing publishes (no page, no tool) until K1–K3 pass and are recorded here.

## KILL-SWITCH OUTCOMES — recorded 2026-08-08
K1b (external, Treasury MTS "Customs Duties" net receipts, 114 months): PASS.
  Yearly mean ratio ours(assessed)/Treasury(net collected): 0.89–1.03, every year
  2017–2026. Expected <1 in normal times (we exclude AD/CVD + fees); documented
  divergences: 2025 spikes (IEEPA cash-timing) and 2026-05/06 where Treasury net
  goes ~0/negative on court-ordered IEEPA REFUNDS — our assessed-at-entry figures
  correctly do not follow refunds. Divergence understood = check passed.
K2 (spot cells vs statutory expectations): PASS 12/12 after correctly applying
  dutiable-value semantics — duty-free lines carry dut_val=0 (e.g. China wooden
  seats 2017-06: $348M consumption, $0 dutiable, $0 duty = MFN Free confirmed);
  Mexico avocados 2024-06: 99.92% of value entered USMCA-free, the quoted 2.8%
  is the rate on the 0.08% non-claiming sliver. DESIGN RULE: the tool serves BOTH
  rate_on_dutiable (duty/dut_val) AND burden (duty/con_val). Never conflated.
  Confirmed cells include: Germany & Japan cars 2.5% (MFN exact), Vietnam knit
  sweaters 32.0%, China seats 25.0% (2019-12, 301), China toys 30.3% (2025-06,
  IEEPA), Italy wine 1.1%.
K3 (completeness): PASS. 114/114 months, 0 failed pulls, 225–233 real countries
  per month (variation = genuine trade sparsity, not pipeline gaps).
K1a (internal, Census world-total rows, all 114 months): RUNNING — result to be
  appended verbatim when complete. Publication remains gated until it lands.
K1a (internal, Census official world-total rows I_COMMODITY=-&CTY_CODE=-,
  all 114 months): PASS — 114/114 reconciled, worst deviation 0.000%, no month
  beyond the 0.5% tolerance. GATE CLOSED 2026-08-08: all kill-switches green;
  Phase 1 (tariff_burden tool + /tariffs page + /api/tariff) cleared to publish.

## ATLAS STUDY REGISTRATIONS — written 2026-08-09 BEFORE final runs
R1 india-gsp-2019: US withdrawal of India's GSP benefits (effective 2019-06-05)
  produces a discrete, sustained step UP in India's dutiable share of import value
  in June 2019 (≥8pp vs May), sustained through December (≥+8pp vs Jan-May mean),
  with no donor origin (China, Vietnam, Taiwan, Japan — non-GSP majors) stepping
  ≥3pp in the same month. FALSIFIED IF: the step is absent/<8pp, gradual, pre-dated,
  or matched by any donor.
R2 tariff-wall: In the pre-war window (2017-01→2018-05), the mean effective burden
  of the apparel-origin set {Cambodia, Bangladesh, Sri Lanka, Pakistan} is ≥3x the
  mean burden of the rich-origin set {Germany, Japan, UK, France}. The 2025 wave
  NARROWED this ratio (2026 H1 ratio < pre-war ratio) by raising the bottom of the
  wall's beneficiaries (rich-set burden rises ≥4x its pre-war level). FALSIFIED IF:
  pre-war ratio <3x, or the ratio widened.
R3 laptop-exodus: On HS 847130 (portable computers), (a) China's share of world
  imports fell ≥45pp within 5 months of 2025-02; (b) world totals stayed within
  ±20% of the 2024-H2 monthly mean during the collapse (ruling out demand/reclass);
  (c) after the tariff's removal (China burden <2% by 2026-03), China's share
  recovered LESS THAN HALF of its loss by 2026-06. FALSIFIED IF any of a/b/c fails.
R4 232-vs-ieepa-retention: Chapters under Section 232 sectoral measures
  {72,73,74,76,87} retained ≥65% of their tariff increase (2024 baseline → peak
  2025-09..2026-01 → now 2026-04..06), while non-232 major chapters {84,85,61,62,94,95}
  retained ≤55%. Chapters whose peak < baseline are excluded (metric undefined) and
  disclosed. FALSIFIED IF the two groups overlap on retention.
R5 dodge-premium-methods (negative result, registered for publication AS a methods
  lesson): the naive claim "trade migrates to the lower-tariff origin in proportion
  to the rate spread" fails — Vietnam grew fastest where spreads were SMALLEST,
  because rate-on-dutiable is computed on a rounding error when duty-free share is
  high. Registered claim: bucketed China→Vietnam spread vs growth shows NO positive
  monotone relation. FALSIFIED IF the relation is positive and monotone.

R6 china-coverage-collapse (registered 2026-08-09, before decomposition run): the
  2025 rise in China's effective burden came MORE from ending duty-free entry
  (coverage) than from raising rates on already-dutiable goods. Registered claim:
  decomposing the 2017-06 → 2025-05 burden rise as Δburden = rate-effect
  (2017 coverage held fixed) + coverage-effect (new coverage at end rates), the
  coverage term accounts for the larger share. Known before registration: dutiable
  share 2017-06 ≈ 42%, 2025-06 92.6% (published on /tariffs); the decomposition
  split itself has NOT been computed. FALSIFIED IF the rate effect alone, holding
  2017 coverage fixed, explains half or more of the burden rise.

R7 rulebook-versus-receipts (methods study, registered 2026-08-09 before run): an
  effective tariff rate computed by summing the published schedule's applicable
  measures cannot be reconciled with the collected receipts on the same codes by
  any constant correction. Registered claim: across ≥4 China test codes spanning
  301-only, 301+2025, and heavily-excluded goods, the ratio collected ÷
  schedule-sum is non-constant (spread of ratios > 2x between codes). FALSIFIED IF
  a single scalar maps schedule sums onto collected rates within ±5 points on all
  test codes — arithmetic reconciliation would then be possible.

R6 VERDICT (run 2026-08-09, frozen data to 2026-06): CONFIRMED. Burden rise
  2017-06→2025-05: 2.71% → 47.92% (+45.2pp). Decomposition: rate effect at 2017
  coverage 19.24pp (42.5%), coverage effect at end rates 25.98pp (57.5%).
  Dutiable share 41.8% (2017-06) → 91.3% (2025-05) → 80.9% (2026-06). Falsifier
  (rate effect ≥ half) not met.

R7 VERDICT (run 2026-08-09, month 2026-06, China): CONFIRMED. Collected ÷
  naive-schedule-sum ratios: cars-1.5-3L .636, sweaters .613, upholstered seats
  .548, laptops .205, toys .133, li-ion batteries .054 (sum reads 642.5%,
  collects 34.4%). Spread 11.8x between codes — no scalar reconciles; falsifier
  not met. Naive sum = accepted China-scoped measures naming the code in
  extracted scope, plus measures with no machine-readable scope (PDF annexes) —
  exactly what a schedule reader cannot exclude.