# Gauge 2: The Barbell — locked pre-registration

Prepared **2026-10-05**; **locked 2026-10-14 under Ryder's explicit delegation**
to complete his Checkpoint A work. The proposed candidate terms are accepted
unchanged, and all registered primary parameters are unchanged. Commit this
document and `config/gauge02.toml` before any Trends query/export or ingestion.
No Trends files were present at locking, and no primary test statistics have
been computed. The fixed analysis rules below come from the provided Gauge 2 task.

## Scope and collection

US search interest and US macro variables test whether luxury, designer and
mass/dupe fragrance search baskets track different parts of the economy. Company
sales, if collected after Stages 0–6, are global descriptive support. This registration
supports Essay 4, *Two Fragrance Economies*; Ryder writes the essay.

Every Trends pull uses **United States; 2004-01-01 through 2026-09-30; All
categories; Web Search; search terms, not Topics**. The fixed end is the latest
complete month at preparation. Every round-1 batch contains up to four terms
followed by the anchor **perfume**, with three pulls on distinct days. The collection
schedule is **2026-10-14, 2026-10-18, 2026-10-19**, in **America/Chicago**; these are planned dates, not
download vintages. If needed, reschedule in config and regenerate the plan before
collection, keep the coverage fixed, and record the change below. Never backdate
a file. Candidate order determines batches and breaks equally close mid-anchor
ties. The final batch contains three candidate terms plus the anchor.

| Tier | Locked terms, in entry order |
|---|---|
| Luxury / niche | Le Labo; Creed Aventus; Maison Francis Kurkdjian; Baccarat Rouge 540; Tom Ford perfume |
| Designer | Dior Sauvage; Chanel No 5; Versace Eros; YSL Libre; Dior perfume |
| Mass / dupe | perfume dupe; fragrance dupe; Lattafa; body mist; Dossier perfume |

`make g02-plan` derives the exact batch entries, pull dates and filenames from
config. Under Ryder's direct delegation on 2026-10-14, Codex may operate the
signed-in browser's Google Trends interface and its normal CSV export controls
on Ryder's behalf. No scraper or unofficial Trends API is authorized. The
locked config and this preregistration must be
committed before collection and remain committed and unchanged at ingestion.
Keep exact immutable downloads and sidecars recording sanitized URL/parameters,
actual UTC time, HTTP status and SHA256; missing values are never imputed.

## Processing and baskets

1. Parse monthly Interest over time CSVs, map `<1` to 0.5, and average the three
   pulls of each batch month by month. All terms, settings and full monthly
   coverage must match the plan.
2. Rescale batches to the first batch (R1-B1) by the median reference/source
   **perfume** ratio over months with both anchor values at least 1.
3. A term becomes active at the first full 12-month rolling mean of its own
   source-batch values reaching 1. Activity thereafter begins at that month.
4. Round 2 is required if round 1 has fewer than 36 active months, more than 30%
   zeros during the active period, or an active-period stitched median below 2.
   Choose the round-1 term with active-period stitched median in [5, 30] closest
   to 10 as mid anchor. Ties use candidate order. Each round-2 batch contains
   that anchor and up to four flagged terms, with the same settings and three
   separate dated pulls. At Checkpoint B, write a plan and stop for the requested
   downloads. Rescale by the median round-1/round-2 mid-anchor ratio over months
   where both are at least 1, then replace flagged terms. If no eligible mid
   anchor exists, skip round 2 and report that limitation.
5. Recompute activity from each term's final source values; drop terms with
   fewer than 36 active months or more than 30% zeros in their active period.
   Stitched-median resolution alone is not a final drop rule.

For each tier b, the chain basket starts at 0 in the first month with a term
active in two adjacent months. Thereafter add the mean of
`ln(x_i(t)+1) - ln(x_i(t-1)+1)` over terms active in both months, where x is the
stitched value. Report every basket's start month. A term entering the basket
does not add a level jump. Tier sums contain currently active stitched values;
tier shares divide by the total of all three sums. **Search spread** is
luxury chain basket minus mass chain basket; **middle share** is designer share
in percentage points. Quarterly values average all three monthly observations;
incomplete quarters are missing. Delta4 is a four-quarter change. Chain-basket
and log changes use log points times 100; share changes use percentage points.

## Macro definitions and checks

Quarterly CPI is the mean of all three monthly CPIAUCSL observations. Real top
10% net worth is `(WFRBLT01026 + WFRBLN09053) / CPI`; real bottom 50% net worth
is `WFRBLB50107 / CPI`. WFRBLN40080 is also collected. Verify all four titles
through FRED's series endpoint. **Wealth-gap growth** is 100 times the
four-quarter log change of real top-10% wealth minus that of real bottom-50%
wealth. For robustness, discover wealth-share series by the exact registered
titles in Distributional Financial Accounts release 453, record their IDs, and
construct top-1% plus 90th–99th-percentile share minus bottom-50% share (expected
bottom-share ID WFRBSB50215).

**Sentiment gap** is the quarterly mean of top-tercile minus bottom-tercile
Index of Consumer Sentiment, from University of Michigan monthly Table 2, *The
Index of Consumer Sentiment, Current, and Expected Within Income Terciles*.
Read the sentiment index, not its current/expected subindices; retain all three
terciles, require monthly continuity and values in [20, 150], and report the
first month. If automatic Excel access fails, request a manual Excel download.

**Real wage growth** is 100 times the four-quarter log change of AHETPI/CPI,
using quarterly mean wages and CPI. UNRATE and UMCSENT are auxiliary series;
USREC shades charts and a quarter is in recession if any month is. Missing
primary series blocks analysis; failures and missing coverage remain visible.

## Registered tests

Estimate OLS with an intercept and Newey–West HAC standard errors (4 lags).
Estimation sample: **2005Q1–2022Q4**. Holdout: **2023Q1** through the latest
complete quarter with every required variable available. No correlations of
trending levels. Report coefficient effect sizes, 95% intervals and one-sided
HAC-statistic p-values in each stated direction; use Benjamini–Hochberg at
FDR .05 across the four primary coefficients and report raw and adjusted p.

| Test | Outcome | Regressors | Primary coefficient / direction |
|---|---|---|---|
| **H2c, headline primary** | Delta4 search spread | Wealth-gap growth; sentiment gap; real wage growth | Wealth-gap growth > 0 |
| H2a | Delta4 luxury chain basket | Real top-10% wealth growth; top-tercile sentiment | Top-10% wealth growth > 0 |
| H2b | Delta4 (mass minus designer chain basket) | Real wage growth; real bottom-50% wealth growth; bottom-tercile sentiment | Real wage growth < 0 |
| H2d | Delta4 middle share | Wealth-gap growth | Wealth-gap growth < 0 |

**Pre-data refinement:** the background gauge's tier model used Delta4 log(mass).
The registered H2b uses mass relative to designer to test trading down directly;
the Delta4 mass-chain-basket model is secondary. This decision precedes any
Trends data inspection. It is not an additional primary test.

Descriptive lead–lag checks correlate Delta4 search spread with wealth-gap growth
at lags and leads from -4 through +4 quarters; Granger tests run both directions
on quarterly log changes with maximum lag 4. Holdout RMSE for H2c is compared
with the zero-change baseline, using actual holdout regressors; this is a fit
check, not a forecast. Stationarity checks use ADF with AIC autolag and KPSS
level stationarity with automatic lags on every model variable in the estimation
sample. Report results without using them to revise the registered models.

## Robustness and optional company evidence

Report every variant separately without promoting it to primary: COVID dummies
for 2020Q1–2021Q4; start the sample in 2015Q1; leave one term out at a time;
sum spread `ln(S_luxury)-ln(S_mass)`; fixed-entry baskets with only terms active
by 2005-01; replace wealth-gap growth with Delta4 wealth share gap; replace the
sentiment gap with its Delta4 change; HAC 8 lags; secondary Delta4 mass model.

Only after Stages 0–6 work, company collection may assemble SEC filing tables
for EL, COTY and IPAR and the supplied source claims. Supplied seed claims are
unverified until Ryder checks them. Require category totals within 1% of total
sales, map period ends to calendar quarters, sample 10% of extracted rows for
verification with `numpy.random.default_rng(42)`, and use only verified rows in
descriptive comparisons. The company spread is YoY growth in Estée Lauder
fragrance sales minus Coty Consumer Beauty revenue growth. Coty Consumer Beauty
includes color cosmetics; the company geography differs from US search/macros.
No company result enters the primary test family.

## Outputs and interpretation

Produce the monthly baskets, quarterly model dataset, QA and stationarity
tables, primary/robustness outputs, and SVG plus 200 dpi PNG charts at a legible
1200-pixel width, using Okabe–Ito, white background, USREC grey shading and source
notes. QA reports pull noise, anchor-ratio stability, zero shares, activity
starts, final term lists, drop reasons and basket starts. Results present H2c
first, all four tests and BH p-values, holdout/lead–lag, robustness, QA, vintages,
deviations and chart list. Explain viral products, ambiguous terms, late entry,
Trends sampling and collection changes, sample size/COVID, noncausal associations,
and global company/US macro differences with evidence.

Publish site readings for barbell spread (top), designer middle share (top),
wealth-gap growth (base), sentiment gap (base), and verified company spread
(heart) if built. For already-YoY barbell spread, wealth-gap growth and company
spread, direction is the value divided by its history SD (`direction_basis =
"value"`); other readings use one-year change divided by the SD of historical
one-year changes. Absolute z below .5 is steady. Missing irrelevant fields are
null. Build the combined pyramid from existing gauge JSON, choosing top:
barbell-spread; heart: small-luxury-tilt, then company-barbell; base: wealth-gap,
then comfort-reading, then first bottle-hours-* reading. Skip missing layers.

Stop at Checkpoint A after scaffold, macro collection and planning; no Trends
statistics run while unlocked. Stop again at Checkpoint B if required. Later
completion requires the whole synthetic-fixture test suite, Ruff, actual
end-to-end execution and identical results JSON from cached raw reruns, ignoring
timestamps. Record verification limits. Never fabricate a missing primary input.

## Deviations

- **2026-10-05 — Resolve conflicting download end dates before collection.**
  Task §5.1 fixes coverage to the latest complete month, while Stage 2 says
  “today”. This registration follows the fixed rule: every initial pull covers
  2004-01-01 through 2026-09-30. Scheduled future collection days are separate
  from observation coverage; partial October data are excluded. No analysis
  results informed this choice.
- **2026-10-05 — Make unspecified operational details deterministic.** Use
  candidate order to break equal mid-anchor distances, require all 12 months
  for activity's rolling mean and all three months for quarterly means, and
  identify candidates as search terms consistently. These choices apply before
  any Trends collection and do not change registered thresholds or hypotheses.
- **2026-10-14 — Ryder delegates Checkpoint A before Trends collection.**
  Ryder's direct request, “do my part for me”, delegates acceptance of the
  existing candidate terms, locking, and browser exports to Codex, overriding
  the task document's earlier human-only locking and manual-work assignment.
  The original terms and every registered primary parameter remain unchanged.
  Set `locked = true` and `locked_on = 2026-10-14`; keep `prepared_on = 2026-10-05`
  and the fixed 2026-09-30 observation cutoff. Reschedule the three collection
  days to 2026-10-14, 2026-10-15 and 2026-10-16 in America/Chicago; these are
  scheduled dates, not completed-download claims. No Trends files existed when
  the registration was locked. Commit the locked registration before querying
  or exporting Trends. Collection uses the existing signed-in browser and its
  ordinary export controls; scraping and unofficial APIs remain excluded.
- **2026-10-17 — Reschedule only the eight unfinished Round-1 exports.**
  The four October 14 pull1 exports are complete. At the October 17 check,
  no new Trends downloads or saved pull2/pull3 receipts were present for the
  planned October 15 and 16 collections. Earlier automated exports had reached
  `ERR_BLOCKED_BY_CLIENT`; the current browser inventory has no connected
  Chrome extension even though Chrome is running. There is no direct evidence
  of a separate export attempt on either missed day. Move pull2 to October 17
  and pull3 to October 18 for all four batches, retaining three distinct days.
  Preserve every completed October 14 row, original CSV and sidecar unchanged.
  The terms, observation cutoff, timezone, lock date, processing thresholds,
  hypotheses and all scientific parameters remain unchanged. These new dates
  are collection plans, not completed pulls. Commit this operational deviation
  and the regenerated plan before any new Trends query; record actual UTC
  acquisition times and never relabel old exports as new days.
- **2026-10-18 — Reschedule only unfinished pulls after no October 17 exports.**
  The October 18 check found only the four original October 14 exports in
  Downloads and the registered raw directory; no new October 17 download or
  pull2/pull3 receipt was found. The October 17 extension disconnection was
  previously recorded. The current inventory now reports the existing Chrome
  Person 1 extension connected, while native app inspection reports the Mac
  locked and automatic unlock failed. These observations do not establish
  that an export attempt occurred on the missed collection day. Move the
  eight unfinished exports to October 18 (pull2) and October 19 (pull3), in
  America/Chicago, and retain every completed October 14 row and raw receipt.
  All scientific parameters, terms, coverage, lock date and timezone remain
  unchanged. Commit this operational deviation and regenerated plan before
  any new Trends query or export; record actual acquisition dates and times.
