# Gauge 3 — Comfort Index: frozen exploratory registration

Locked on 2026-10-05 by Codex under Ryder's explicit instruction, “do everything
yourself,” after the first-run checkpoint. Config, this document and the 200-row
label sheet are committed before annual outcomes, statistical tests or holdout
scores are computed. This is an exploratory study of community catalogs, not sales
or a causal design. Ryder writes the essay prose.

## Scope and original specification

**Original concentration-restricted primary analysis: unavailable.** Neither the
Fragrantica v3 CSV nor JSONL has a concentration field. Do not infer concentration
from names or claim the original test was recovered. Separately execute a revised
exploratory all-catalog specification with the supplied H3a–H3c outcomes, lag sums,
directions, HAC and three-test BH family. This adaptation is based on schema QA
only; no outcome or test result informed it. Report the unavailable original result
first, and label the computed all-concentration findings exploratory throughout.

Primary catalog source: Kaggle ledecanteur/fragrantica-perfumes v3, updated
2026-09-20, snapshot year 2026. It contains 140,228 records; use the fuller JSONL
for weighted accords and nested notes. Fragrantica rating-count weights are the
sum of rating.histogram bucket counts (CSV vote_count), not the broader people
aggregate. Replication: Parfumo TidyTuesday release 2024-12-10, snapshot year fixed
to 2024; original scrape/update date unavailable. Macro: exact FRED UMCSENT, UNRATE
and USREC observations and source metadata.

Start in 1980. Each source ends at its latest pre-snapshot year with at least half
the mean launch count of the preceding three calendar years. Source windows are
1980–2025 Fragrantica and 1980–2023 Parfumo. Counts use the whole catalog before
outcome classification. Use raw numeric launch years; missing/invalid years are
excluded without imputation. Decade-only descriptions and source-date conflicts
remain documented error risks; do not repair years from guesses.

## Cleaning, classifications and universes

Remove records that repeat Name/Brand/URL header labels. Exclude every record in
a duplicated nonmissing identifier group, rather than arbitrarily selecting a
conflicting version. This removes 10 headers and 70 Parfumo duplicate-ID rows,
leaving 59,245. Fragrantica IDs are unique. Preserve all raw files unchanged.

Fragrantica revised main universe is all cleaned catalog records with valid years
in its window. Parfumo main retains the exact concentration labels in config;
its all-concentration universe is a separate comparison. Exact vocabulary sets,
concentration labels, popularity columns and frozen house lists are recorded in
config/gauge03.toml. They were selected from whole-catalog frequencies, never
annual outcome shares. Fragrantica gourmand accords: almond, cacao, caramel,
chocolate, coffee, honey, nutty, sweet, vanilla. Fresh: aquatic, citrus, fresh,
green, marine, ozonic. Parfumo gourmand: Gourmand, Sweet; fresh: Aquatic, Citrus,
Fresh, Green. These taxonomies are not equivalent measurements.

Gourmand/fresh flags require any matching accord among the top three, ranked by
strength when supplied and listed order otherwise; preserve order on ties.
Note robustness requires a matching heart/base note from the exact config set.
Zero note-vote weights do not erase listed ingredients. Flat-only notes do not
become invented heart/base notes. Empty fields remain unknown. Shares use all
launches in the selected year as denominator, counting only explicitly recorded
positives; report unknown counts separately. These measure recorded catalog
incidence, not estimated latent preferences. Complete-case flags are a separate
sensitivity. Store fractions; models and charts use percentage points.

Fixed panel: houses with a launch in each of the 1990s, 2000s and 2010s, within
that source's selected universe (309 Fragrantica; 129 Parfumo). Popularity weights
use rating counts; missing/zero counts receive zero weight. Fragrantica gender
labels female/male/unisex map to women/men/unisex. Parfumo lacks gender, so that
variant is unavailable rather than inferred. All-concentration, fixed panel,
popularity, gender and note variants remain separately reported.

## Flankers and evidence labels

Within brand plus nonblank collection, a launch strictly after the line's first
year is a flanker. Missing collection falls back to names; Parfumo always uses
names. Normalize lowercase and punctuation to spaces, collapse whitespace, remove
registered whole-word modifiers (longest phrases first), and group by brand plus
remaining base name. Use full available catalog history, including pre-1980
pillars, before restricting the analysis universe. Missing brand/year/base-name
cases are unknown. Same-year pillar/flanker ordering cannot be distinguished.

Sample 200 eligible Fragrantica records with numpy.random.default_rng(42), sorting
IDs first for row-order stability: first 100 tune, last 100 holdout. Under delegated
autonomy, an independent AI subagent reviewed source-history evidence blind to
rule predictions and split membership. There are 16 pillar, 8 flanker and 176
unsure labels. Only 24 determinate cases remain and they are selected for available
lineage evidence. Report metrics as AI-assisted agreement, not human validation
or representative accuracy. Preserve evidence notes and uncertainty exclusions.

Retain the supplied modifier list unchanged. No tuning against these labels or
outcomes. Record tune agreement, freeze config/labels and their Git commit, then
score holdout once with flanker positive, excluding unsure. Report precision,
recall and F1 with denominator/confusion counts and missing-prediction exclusions.
An immutable receipt binds config, labels, predictions and registration commit;
identical reruns reuse it, and changed inputs cannot silently rescore the holdout.

## Lagged models and multiplicity

Stress is minus the first difference of calendar-year UMCSENT mean. Include every
observed month and report annual observation counts, especially sparse pre-1978
history. Reindex full calendar years before differences and lags. For each outcome
y in percentage points, fit change(y_t) = alpha + beta0*stress_t + beta1*stress_t-1
+ beta2*stress_t-2 + beta3*stress_t-3 + error. Use complete terms within the window,
OLS with HAC/Newey-West covariance, 2 lags.

H3a: gourmand beta1+beta2+beta3 > 0. H3b: flanker beta1+beta2 > 0. H3c: fresh
beta1+beta2+beta3 < 0. Sum variance is w'Vw. One-sided p and 95% intervals use
Student t with residual degrees of freedom. Report each coefficient/interval,
raw p, and Benjamini–Hochberg across exactly these three exploratory Fragrantica
main tests at FDR .05. Also report joint Wald beta1=beta2=beta3=0 using F(3,df_resid).
Lag sums describe cumulative later share change over the indicated lag window
following a one-year, one-point stress impulse; concurrent beta0 is excluded.

Pre-specified refinement supplied before data: the original gauge named a joint
lags-1–3 test primary. Directional hypotheses use summed lag effects instead,
with the joint test alongside. Robustness results never replace these tests.

## Secondary event studies

A recession year has any USREC month equal to one. Events are the first years of
consecutive recession-year runs within the window, expected 1980, 1990, 2001, 2007,
2020; derive from actual USREC. Subtract LOESS trend, frac=.5, robust iterations=3,
over the source window. Average deviations over e…e+3, then across complete events;
report excluded incomplete events. Placebos are non-recession years at least
three years from every event year, with full windows. Draw distinct years, as many
as complete events, 5,000 times using seed42; enumerate all combinations when no
more than 5,000. Tail follows outcome direction (negative fresh responses reverse
sign). p=(1+draws at least as extreme)/(1+draw count).

Report mean deviations, per-event responses, and placebo central 95% null ranges
as null references, not effect confidence intervals. No event-effect CI estimator
was supplied; mark it unavailable. Leave-one-event-out ranges describe sensitivity.
Few episodes and overlapping windows limit inference.

## Robustness, publishing and optional demand

Report note-based gourmand, fixed houses, popularity weighting, gender strata,
all concentrations, recession-dummy stress lags0–3, annual UNRATE change lags0–3,
LOESS-deviation outcomes, complete-case sensitivity, and full Parfumo model/event
replication for selected and all concentrations. Replication and robustness are
outside the three-test BH family. Report null/opposite signs plainly.

Comfort reading: latest Fragrantica gourmand share minus its fixed LOESS trend,
percentage points; full annual history, historical percentile, direction from
one-year change divided by SD of historical one-year changes (|z|<.5 steady).
Use shared validated site writer; Okabe–Ito SVG/200dpi PNG with source notes and
recession context on time-series charts. Heatmap top20 rows are chosen from the
whole-catalog accord vocabulary; cells count top-three presence once per perfume.
QA launch-count charts omit recession shading to avoid pre-lock macro comparisons.

Optional H3d only with three genuine Google Trends comparison CSVs from different
days: United States, 2004 to latest month, All categories, Web Search; vanilla
perfume, gourmand perfume, citrus perfume, fresh perfume, perfume. Average pulls,
geometric means of value+1 for the two-term baskets, monthly difference in their
log ratio on monthly UMCSENT change, HAC12, one-sided beta<0. Treat <1 as .5;
reject unfamiliar formats/mismatched comparison grids. No interpolation or fake
pull dates. H3d is exploratory and outside BH; absent inputs mean unavailable.

## Deviations

- 2026-10-05 — Initial first-pass checkpoint preserved unlocked registration and
  available-source QA; no outcomes or tests were calculated.
- 2026-10-05 — User explicitly delegated remaining work. Official anonymous Kaggle
  API resolved acquisition without credentials; exact v3 archive/schema preserved.
- 2026-10-05 — Original concentration-restricted specification declared unavailable;
  separate all-concentration exploratory analysis fixed before any outcomes.
- 2026-10-05 — Conservative AI-assisted source-history labels replace requested
  human labels: 24 determinate/200, 176 unsure; report agreement limitations, not
  human validation. Original modifiers retained unchanged; score holdout once.
- 2026-10-05 — Fixed header/duplicate exclusion, missing-field denominator policy,
  complete-case sensitivity, Student-t/Wald-F references and LOESS iterations
  before outcomes. Raw numeric years retained with documented uncertainty.
