# Gauge 4 verification — 2026-10-14

**All nine stages are complete for the required Chanel and Dior basket.** Ryder
explicitly delegated the archive price review to Codex. Every decision identifies
Codex and its evidence; this is not an attestation that Ryder performed human
verification. The methodological substitution was logged and committed before
confirmatory analysis. No further primary review action remains.

## Review and provenance

- Independently inspected all 169 cached archived HTML snapshots and completed
  every one of the 308 sampling-row decisions: 110 accepted and 198 excluded.
  Exclusions comprise 139 missing captures and 59 unsupported/wrong-product
  captured pages. No alternate capture or inferred price replaced an exclusion.
- All 58 primary rows have y/n decisions: 30 accepted and 28 excluded. The fixed
  retailer selection remains Chanel’s own US site and Macy’s. The primary
  series contain 23 Chanel and 7 Dior observations. Secondary accepted prices
  are used only in the registered multi-retailer robustness series.
- Private local audit JSON in `data/manual/bottle_prices_review_audit/` records
  DOM/JSON locators, SKU/size-price mapping, currency context, raw SHA256, reviewer
  and reasons. Corrected prices and source fields are in the ignored review CSV.
  A backup of the earlier blank review CSV is retained in the audit directory.
- Manual corrections use explicit product-specific forms or mapped main variants
  inside the already sampled authorized URL. Unmapped ranges, wrong concentrations,
  recommendations on wrong-product pages, special gift packages, refills, bot
  pages and missing size/regular-price evidence are excluded.
- No unverified, fabricated or interpolated price enters metrics, charts or site
  data. Every primary observation has its exact snapshot-month official wage and
  both CPI values. All 110 accepted observations retain viewable snapshot links.

## Primary results

The registered alternative is an **increasing** bottle-hours trend. Neither
primary series supports that direction. Both fitted trends are negative. These
results concern the retained historical observations; Dior’s latest primary
capture is August 5, 2021, and Chanel’s is February 16, 2026. They are not estimates
of current Dior prices. The site interpretation explicitly names the capture month.

| Bottle | Primary observations | First → latest hours | Sampling window | Annual trend % | 95% HAC interval % | Raw p | BH p |
|---|---:|---:|---|---:|---|---:|---:|
| Chanel N°5 Eau de Parfum spray 100 ml | 23 | 6.0782 → 5.9301 | 2010-03-01 → 2026-03-01 | -0.860461 | [-1.38086, -0.337316] | 0.999355 | 1 |
| Dior Sauvage Eau de Toilette spray 100 ml | 7 | 4.1166 → 3.8285 | 2016-09-01 → 2021-09-01 | -2.01574 | [-2.72283, -1.30351] | 1 | 1 |

## All registered tests

OLS fits logs on decimal snapshot year, with Newey–West/HAC lags = 2. The six
H4a/H4b tests form one BH family at FDR .05. Conditional H4c was not run because
optional Creed collection remains deferred. Parameters were not tuned after review.

| Test | n | Annual % | 95% interval % | Raw p | BH p | Supports registered increase |
|---|---:|---:|---|---:|---:|---|
| H4a:chanel_no5_edp:bottle_hours | 23 | -0.860461 | [-1.38086, -0.337316] | 0.999355 | 1 | no |
| H4b:chanel_no5_edp:overall_inflation | 23 | -0.0240251 | [-0.472713, 0.426685] | 0.541692 | 1 | no |
| H4b:chanel_no5_edp:category_inflation | 23 | 2.37369 | [1.90602, 2.84351] | 4.98348e-24 | 1.49505e-23 | yes |
| H4a:dior_sauvage_edt:bottle_hours | 7 | -2.01574 | [-2.72283, -1.30351] | 1 | 1 | no |
| H4b:dior_sauvage_edt:overall_inflation | 7 | -0.417594 | [-0.843573, 0.0102151] | 0.972143 | 1 | no |
| H4b:dior_sauvage_edt:category_inflation | 7 | 3.08075 | [2.64457, 3.51879] | 5.52e-45 | 3.312e-44 | yes |

Both bottle prices outpaced the cosmetics-category index in the registered
windows. Neither outpaced overall CPI under the registered one-sided tests.
All null results and adjusted p-values are retained. Small samples, archive
missingness and irregular spacing limit generalization; the directional tests
do not establish that any individual’s fragrance budget became cheaper.

## Robustness and coverage limits

- All 14 reported sensitivity rows are saved in `reports/gauge04-sensitivity.csv`:
  alternative wages, verified sale-price substitution and multi-retailer medians.
  No separate supported sale prices were recorded, so that substitution uses the
  reviewed regular prices. Variants remain distinct from the primary BH family.
- The multi-retailer Dior window reaches 2026, unlike the fixed primary retailer
  window ending in 2021. Observation windows are shown in the brief and CSV;
  different coverage windows must not be treated as an equivalent same-window test.
- Source discovery retained 982 URL histories / 4,516 daily capture records,
  including 75 optional Creed histories retained locally. Required-bottle history
  coverage is limited to the registered prefixes and bounded domain searches.
  No captures in a queried scope is not proof of archive-wide absence.
- The machine extractor still reports 21 eligible candidates, 91 literal
  `status=needs_review` rows and 57 identity-mismatch rows. Manual evidence review
  recovered additional exact-size prices and rejected unsupported candidates.
  Extraction confidence flags persist after review; they are not pending decisions.
- The full local brief includes both discovery QA and collection QA, separately
  labeled, the common-window comparison, every accepted price’s snapshot link,
  excluded rows/reasons, caveats, source vintages, versions and deviations.

## Fresh execution and reproducibility

- Actual wages refresh and individual review, metrics, analyze, charts and publish
  stages completed. The review gate passed from the completed evidence decisions;
  it was not bypassed or relaxed.
- `.venv/bin/python -m sillage.gauges.g04_bottle_hours --stage all --offline --force`
  executed all nine real-data stages and returned 0. All nine manifests were current.
- `reports/gauge04-results.json` was identical after removing only `updated_at`.
  Fourteen checked artifacts were byte-identical, including candidate/review CSVs,
  macro and price parquet, macro vintages, sensitivity CSV and all final charts.
- All 366 recorded raw downloads and sidecars remained SHA256-identical across
  the replay. All 356 receipts that predated this October 14 work remain unchanged.
- `make lint test`: Ruff passed; **574 synthetic tests passed in 39.85 seconds**.
  Two new dated-reading/collection-QA regressions failed before their fixes and
  passed afterward. Existing synthetic end-to-end tests exercise the review gate,
  conditional three-bottle BH family and cached deterministic exports.
- Both site JSON schemas passed validation. Rendered final plots/cards were
  visually inspected; a title/subtitle overlap and clipped square title were
  corrected. PNGs are 200 dpi: 1200×840 plots and 1200×675 / 1080×1080 cards.
  Charts use white backgrounds, Okabe–Ito colors, recession shading where
  applicable and source notes. Sampling-date labels are explicit on the cards.

## Source vintages and registration

Archive acquisition retains October 2 and October 6 UTC pulls; all 169 sampled
HTML receipts are October 6 UTC. No new archive capture was substituted during
this review. The official macro refresh uses immutable October 14 UTC files;
all five series now reach September 2026. Prior source files remain intact.

| Series | Last observation | Source update |
|---|---|---|
| AHETPI | 2026-09-01 | 2026-10-02 08:26:56-05 |
| CES0500000003 | 2026-09-01 | 2026-10-02 08:26:56-05 |
| CPIAUCSL | 2026-09-01 | 2026-10-14 08:08:29-05 |
| CUUR0000SEGB02 | 2026-09-01 | 2026-10-14 08:08:26-05 |
| USREC | 2026-09-01 | 2026-10-01 18:02:31-05 |

The original locked config/prereg commit is
`81adabc009b63435e522302c478a7425b2594a52`. The Codex reviewer substitution and
product-form correction method were committed as `a517058` before confirmatory
analysis. No sampling, price definition, wage series, metric, hypothesis, HAC lag
or BH parameter changed. Full dated deviations remain in `prereg/gauge-04.md`.

Runtime versions: pandas 3.0.6, numpy 2.5.3, scipy 1.18.1, statsmodels 0.15.0, matplotlib 3.11.2, pyarrow 25.0.1, requests 2.34.2, beautifulsoup4 4.15.0, lxml 6.1.3, pytest 9.1.1, ruff 0.16.10.

## Outputs and remaining actions

- Full local results: `reports/gauge04-results.md` and `.json` (ignored because
  they contain the required row-level archive table). Local review/audit files
  remain ignored. Aggregate statistical QA, charts and site data are committed.
- Site artifacts: `site/data/gauge04.json` and `site/data/gauge04_calculator.json`.
  These are workspace data outputs; no website deployment was performed.
- `reports/gauge04-todo.md` records zero pending primary decisions. Required
  designer-basket work is complete. Creed remains an optional deferred extension.
- No essay prose, front-end work or another gauge’s files were changed by this task.
