# Gauge 4: Bottle-Hours preregistration

Locked on **2026-10-01**, before archived prices are extracted. This is the fixed
analysis plan supplied by Ryder. Config: `config/gauge04.toml`. The registration
commit is verified against Git before collection and confirmatory stages.

## Basket, identity and authorized retailers

Required: Chanel N°5 Eau de Parfum spray (100 ml / 3.4 fl oz, designer/heritage)
and Dior Sauvage Eau de Toilette spray (100 ml / 3.4 fl oz, designer). Stretch:
Creed Aventus Eau de Parfum spray (100 ml / 3.3 fl oz, luxury/niche), collected
only after the required two have usable extractions.

Accept parsed size within 100 ± 3 ml; convert fluid ounces using 29.5735 ml/fl oz.
Require exact product and concentration in title, h1 or JSON-LD product name.
Exclude N°5 L'Eau, Eau Première, Parfum/extrait, EDT; Sauvage EDP, Parfum,
Elixir; Aventus for Her, Cologne and Absolu Aventus. USD only. Keep list and sale
prices separately; sale observations never enter the primary list-price series.

Use the bottle's own US brand site first (chanel.com, dior.com,
creedboutique.com), then nordstrom.com, saksfifthavenue.com, neimanmarcus.com,
bloomingdales.com, macys.com, sephora.com and ulta.com in that order to break
ties. Exclude marketplaces, discounters and grey-market sellers, including
Amazon, eBay, Walmart Marketplace, FragranceNet, FragranceX and Jomashop.

## Archive sampling and verification

Group changing URLs into bottle × retailer families. Query CDX with repeated
AND filters, daily successful captures, lowercase bottle keywords, pagination
where needed and product-page filtering. Each year from the first capture
through the current year, choose the capture closest to March 1 and September 1
within ±45 days; ties select the earlier capture. Dates outside the window are
missing. Never interpolate or fabricate prices.

For each bottle choose the **single retailer with the most sampling dates with
an extracted 100 ml list price**, breaking ties by the order above. Freeze that
choice before human review. The robustness series is the median verified list
price across authorized retailers at each sampling date.

Ryder must mark every primary-series review row `y` or `n` before metrics,
analysis, charts or publication. `y` requires a verified positive USD list price;
`n` rows are excluded and counted. Preserve raw HTML, extraction candidates,
snapshot links and reviewer corrections separately. Robustness uses verified
rows only. Unverified prices cannot appear in analysis, charts or site data.

## Monthly macro data and metrics

FRED AHETPI (production/nonsupervisory total private hourly earnings) is the
primary wage; CES0500000003 (all employees total private) is the alternative.
CPIAUCSL is overall inflation and CUUR0000SEGB02 is the unadjusted cosmetics,
perfume, bath and nail category benchmark. Join each price to its **snapshot's
calendar month**. Download USREC only for shared recession chart shading.

Bottle-hours = verified list price / AHETPI. Alternative hours = price /
CES0500000003. Real price = price × latest CPIAUCSL / snapshot-month CPIAUCSL,
in latest available month dollars. Category-relative index =
100 × (price / first price) / (category CPI / first-price-month category CPI).
Price per ml = price / 100. Missing required macro values block primary analysis.

## Registered hypotheses

Fit OLS with intercept and decimal year **year + (snapshot month − 1)/12**;
Newey–West/HAC standard errors with **2 lags**, one-sided positive alternative.
Report 100 × (exp(beta) − 1) annual percent change and its transformed 95% interval.

| Test | Per-bottle outcome | Direction |
|---|---|---|
| H4a, primary | ln(bottle-hours) | > 0 |
| H4b overall | ln(price / CPIAUCSL) | > 0 |
| H4b category | ln(price / CUUR0000SEGB02) | > 0 |
| H4c, only with Aventus | ln(Aventus price / each designer price), matched sampling dates | > 0 |

Apply Benjamini–Hochberg at FDR 0.05 across **all H4a–H4c tests**, retaining raw
and adjusted p-values and reporting null results. The descriptive common
window uses the first and last sampling dates shared by all collected bottles;
Sauvage's 2015 launch limits the shared window. Regress logs rather than
correlating trending levels. No parameter tuning to improve outcomes.

## Sensitivities, outputs and reporting

Report alternative wages, including sale-price observations, and the verified
multi-retailer median as sensitivities only. Never promote them to primary.
LOWESS display uses frac=0.5. Bottle-hours lines use observed values; missing
dates remain gaps. Then-versus-now cards compare the first sampled year and
latest observation. Indexed prices, wages and both CPIs use the first common
sampling date as 100. Save white-background Okabe–Ito SVG and 200 dpi PNG charts
with USREC shading on time-series plots and bottom-left source notes.

Site readings: base layer, hours, null percentile; latest verified observation
and chronological sampling-date history. Direction uses nearest one-year
change divided by SD of historical one-year changes, steady for |z|<0.5;
five-year change uses the closest observation to five years earlier. Calculator
data retain latest verified list price, snapshot date and retailer.

Results include primary findings first, all tests/intervals/raw and adjusted p,
common-window comparisons, sensitivities, excluded rows and reasons, snapshot
links, archive coverage, pull/update dates, dependency versions, QA, deviations,
charts and evidence for archive/JavaScript gaps, promotions, size/formula changes,
retailer differences and wage choice. Seed any randomness with default_rng(42).
Row-level archive data stay ignored. All tests use synthetic fixtures only.

## Deviations

- 2026-10-01: None. USREC is an auxiliary download required by chart conventions;
  it introduces no change to the four macro measures or registered hypotheses.
- 2026-10-01: Corrected the operational CDX page size from 1000 to 50 index
  blocks, the maximum described by the archive's API documentation. Domain URL
  discovery uses a small 200-URL limit as expressly allowed in Stage 2; explicit
  product prefixes support full pagination. Report this discovery coverage
  limit. Exact URL capture lists and registered price sampling remain unchanged.
- 2026-10-01: Full-domain CDX queries returned HTTP 504. Switched live discovery
  to the product-path prefix method already allowed in Stage 2, including legacy
  US brand paths. Exclude script/tracking URLs and explicitly wrong versions
  before grouping URL families. Prefix scope omissions are coverage limitations,
  not evidence of an absent price; all price identities still require review.
- 2026-10-02: Corrected CDX pagination metadata requests to contain only scope,
  match type, page size and page-count options; record pages retain the registered
  filters. Corrected the US Dior path filter to admit the already registered
  legacy `/beauty/en_us/` scope. Explicit sets, refills and non-fragrance products
  are excluded before URL-family grouping, consistent with the spray benchmark.
  Added immutable-provenance validation and a stop after exhausted HTTP 429
  retries. These implementation corrections change no sampling, price, wage,
  hypothesis or inference parameter.
- 2026-10-02: Exact-URL capture queries took 5–25 seconds each in the live
  diagnostic run. Overlap at most two such requests, with separate sessions,
  globally spaced request starts of at least one second and a shared HTTP 429
  cooldown. Save completed capture lists deterministically and stop later
  batches on systemic access failure. URL scope, daily capture collapse,
  sampling and all primary parameters are unchanged.
- 2026-10-02: Added complete native-paged CDX capture downloads for registered
  product prefixes, retaining only the candidate URLs' canonical `urlkey`
  records and the first successful capture each day. This replaces repetitive
  wire requests while preserving exact-URL capture histories: a real comparison
  against 39 cached exact queries matched all 376 records. Native page counts,
  response validation and archive runtime-error checks guard completeness;
  failed bulk queries may fall back to the original paced exact queries.
  Sampling, URL-family membership and all analysis parameters are unchanged.
- 2026-10-05: Deferred the optional Creed stretch basket for the first designer
  pass (`collection.collect_stretch = false`). All required retailer histories
  were successfully obtained, but optional discovery hit another HTTP 429 before
  either required bottle had extracted prices. This follows the required-first
  priority and prevents optional acquisition from blocking the designer review
  kit. Existing optional raw downloads are retained. Creed can be enabled in a
  later registered operational update; H4c remains conditional on its collection.
- 2026-10-05: Corrected extraction of a single price paired with an ambiguous
  multi-size field (for example 50 ml / 100 ml). It must remain `needs_review`;
  equivalent 100 ml / 3.4 fl oz labels remain eligible. This enforces the original
  unambiguous size-to-price mapping rule; no real candidate price or primary
  test statistic had been produced when the error was found.
- 2026-10-05: Actual cached HTML revealed Latin-1 decoding loss, unsupported
  reference-encoded JSON, and non-product dollar amounts from navigation,
  shipping thresholds and customer reviews. Correct decoding and product-price
  context checks are applied before the first primary-retailer freeze; pooled
  JSON references must not be mistaken for literal prices. Raw downloads and
  sampled captures are retained unchanged. Reprocessing uses cached pages;
  unrelated price mentions stay in the candidate audit but cannot determine
  the bottle's list price. These repairs enforce the registered extraction
  rules and change no analysis parameter.
- 2026-10-05: Corrected collection cache invalidation. When a stage must rerun,
  its price candidates are reprocessed from immutable cached pages instead of
  reusing derived rows from an earlier extractor. Fully current stages still
  skip. A synthetic tampered-candidate regression confirms source prices are
  restored without changing raw receipts or separate human review work.
- 2026-10-14: Ryder explicitly delegated the price-verification task to Codex
  ("do my part for me"). Codex reviews the immutable archived HTML and its
  structured product/variant evidence rather than attesting that Ryder checked
  the pages. Per-row decisions identify the reviewer, raw checksum, evidence and
  exclusions in ignored local audit files and reviewer notes. Unclear identity,
  concentration, size, USD context or regular-price mapping is excluded. The
  review may resolve extractor identity failures using a specifically named
  product form/card and its explicitly mapped size-price option inside the
  already sampled product URL; a generic page title or nearby price is not
  sufficient. This manual source correction does not expand the URL scope.
  The previously frozen primary retailers, sampling and all analysis parameters
  remain unchanged. Output source notes describe evidence review without
  claiming human verification. This reviewer substitution is a methodological
  deviation and must accompany the results.
