# Gauge 1 verification

Verified on 2026-10-01 (America/New_York), using official observations through
August 2026. All seven stages are implemented and have run on real data.

## Registration and stage commits

Locked config/prereg were committed before any real confirmatory statistics:
`647aa46062826879af1cb9afdc0f013ce49c1784`.
Verified BEA acquisition compatibility was logged before confirmation in
`da9a8bd726d1079092350c23c5d54496ac4071ea`; this is the registration hash recorded
in the results. No primary parameter or statistical procedure was changed.

| Stage | Commit |
|---|---|
| Scaffolding and implementation | 647aa46, 1c251ba, 0219444 |
| Acquisition compatibility log | da9a8bd |
| Official monthly data and provenance | ebe6d90 |
| Exploration | 001dbc9 |
| Nested-project registration metadata fix | a1a3fc6 |
| Confirmatory analysis | a838786 |
| Registered robustness | 142e20f |
| Five final charts | 7fb72ef |
| Results brief and site reading | 33415d2 |
| Factual missing-data and robustness disclosures | 44c0d3c |

Commits used explicit fashion/ paths. Unrelated parent-project changes were not
staged. The CLI does not commit automatically.

## Commands and actual verification

- `make setup`: Python 3.12.13, dependencies synchronized from committed uv.lock.
- `make lint`: all checks passed.
- `make test`: 98 synthetic tests passed in 23.15 seconds on final source.
- `make g01-data`, `make g01-explore`, `make g01-confirm`, `make g01-robust`,
  `make g01-charts`, `make g01-publish`: completed on real official observations.
- `.venv/bin/python -m sillage.gauges.g01_lipstick --stage all --offline --force`:
  all seven stages completed from the same immutable raw vintage.
- Results and site JSON reproduce exactly when only updated_at is excluded.
  The eight processed files, sixteen SVG/PNG charts, and robustness CSV are
  byte-identical across cached replay (25 artifacts).
- A fresh local sparse checkout of the committed project, seeded with the
  verified BEA cache, ran literal `make g01` through every stage with both API
  credentials absent. Its results/site JSON matched the project after excluding
  updated_at. FRED raw files were already committed; no second live acquisition
  was needed for this checkout check.
- `make g01` after the final replay: every unchanged stage skips as current.
- Independent audits reconstructed registered event paths, seeded placebo and
  bootstrap draws, HAC regressions/BH, price checks and all robustness rows.
  They matched saved numerical results. Maximum placebo-draw rounding difference
  was 2.7e-15.
- Actual exploratory/final PNGs were visually reviewed; all are readable, all
  SVG/PNG source notes present, PNGs at 200 dpi, and SVG creation dates omitted.
  Site formula/percentile/direction/history/schema were independently checked.

All fabricated fixtures are confined to tests or temporary test roots. Real
outputs contain no synthetic observations or imputed values.

## Data, source vintages and QA

The combined dataset is 812 × 20 with a continuous monthly PeriodIndex from 1959-01 to
2026-08. All 15 BEA columns are complete and finite; quantity and price indexes
are positive. The latest month passes the six-month freshness requirement.
The nominal cosmetics share exactly equals n_cosmetics/n_total.

All 12 raw responses and their sidecars have HTTP 200, recorded UTC pull times,
and matching SHA256. UTC pull-date filenames are 2026-10-02; user-facing analysis
and registration dates use America/New_York. BEA ALL requests needed no chunks.
Full provenance, FRED source-update dates, BEA production times and notes are in
`data/processed/gauge01_vintages.json` and both results briefs.

| Category | Line | Quantity / price / nominal codes |
|---|---:|---|
| Cosmetics |139| DCOSRA / DCOSRG / DCOSRC |
| Clothing |104| DCLORA / DCLORG / DCLORC |
| Jewelry and watches |63| DJRYRA / DJRYRG / DJRYRC |
| Total PCE |1| DPCERA / DPCERG / DPCERC |
| Goods |2| DGDSRA / DGDSRG / DGDSRC |

Original source descriptions are retained. Quantity/price UNIT_MULT=0; nominal
UNIT_MULT=6 is millions of dollars, seasonally adjusted at annual rates, with
raw numeric values retained. Dictionary lists each source, unit and adjustment.

FRED raw coverage: USREC 1854-12–2026-09; cosmetics CPI 1977-12–2026-08;
apparel CPI 1947-01–2026-08; unemployment 1948-01–2026-08. Genuine October 2025
missing values remain in all three BLS series. Cosmetics CPI additionally has
227 months before its first observation. No interpolation or forward fill.
Unemployment changes for October and November 2025 are omitted in intensity;
its HAC 12 lags refer to retained observations. This spacing limitation is
explicitly reported, with the original complete-case calculation preserved.

## Headline and limitations

H1 mean gap +1.787958 log points, placebo/BH p = 0.06518696, 4 of 9 positive;
rough bootstrap 95% interval [-1.859010, 6.717381]. The primary test does not
establish the small-luxury hypothesis at the registered 5% threshold.
BH p: H1b 0.00254091, H1c 0.02819436.
Latest tilt 2026-08: +0.446367 pp, percentile 59.25, steady; history 120 months.

Only nine episodes; perfume is not separately identified. Dropping 1960 changes
H1 to -0.342043 (p = 0.62167566); the post-2000 mean is -0.237776 (p = 0.56508698).
All 13 robustness rows remain reported-only. Growth regressions require complete
quarters; descriptive price checks do not identify a causal substitution effect.
Details and factual evidence appear under What could be fooling me in the brief.

## Secrets and external setup

Ryder explicitly approved BEA terms/CAPTCHA and FRED terms/request submission.
Both issued keys are configured in gitignored .env with permissions 0600. They
are absent from committed raw FRED files, sidecars, processed data and reports.
BEA raw bodies may echo USERID; those exact files stay gitignored, untracked,
and permission-restricted. Keep that local cache for offline replay.
The project virtual environment verifies TLS successfully; no certificate
verification was disabled. No setup checkpoint or Ryder action remains.
