# Gauge 3 completed-run verification — 2026-10-06 UTC

Stages 0–8 actually executed using the downloaded Fragrantica, Parfumo and FRED
sources. This verifies the delegated **revised exploratory all-concentration
analysis**, not the unavailable original concentration-restricted Fragrantica
test. Optional H3d is unavailable because three separate-day Trends exports do
not exist. No observations or missing attributes were fabricated.

## Commands and outcomes

- `make lint`: passed (`ruff check .`).
- `make test`: **527 passed in 41.06 seconds**, the whole current project suite.
  All test data are synthetic fixtures or mocked responses.
- `make g03`: all stages completed using actual acquired sources.
  A final unchanged `make g03` exited 0 and correctly skipped every stage.
- `.venv/bin/python -m sillage.gauges.g03_comfort --stage all --offline --force`:
  **two consecutive complete runs, each exit 0**, after final code changes.
  Scaffold, data, profile, flankers, series, confirm, robust, charts, demand and
  publish all executed. Demand recorded its unavailable status without blocking
  the catalog study.
- Compared **22 artifacts** across those forced offline runs. Results and site
  JSON match exactly after omitting only top-level `updated_at`; all other
  compared files are byte-identical, including the annual CSV, model outputs,
  replication, robustness, results Markdown, status, validation and all ten
  SVG/PNG files. Checksums and the complete file list are in
  `reports/gauge03-verification.json`.
- The final replay's three statistical outputs also matched their hashes from
  before the final presentation/caveat changes. No estimates changed.
- The immutable holdout receipt remained byte-identical across both replays;
  `score_count` remains **1**. Re-execution reuses that receipt rather than
  scoring the holdout again.
- Inspected all five 1200-pixel PNGs for readable axes, legends and source notes.
  The event inset uses categorical positions so its bars remain visible.
  Final time-series plots use USREC shading; the pre-registration QA chart does
  not introduce macro relationships, and heatmap cells encode accord incidence.
- `git ls-files -- data/raw/fragrantica data/raw/parfumo
  data/processed/gauge03_catalog_rows.parquet` returned no tracked files.
  `git check-ignore` confirmed raw catalogs, processed catalog rows and the
  local manifest remain ignored. The 200-row labeling CSV is the brief's
  explicit exception. Aggregate outputs are committed.

The source/dependency fingerprint checks also rejected an interrupted development
replay after presentation code changed during execution. The two successful
replays above occurred after editing finished; no receipt was bypassed.

## Registration and validation limits

Frozen config, preregistration and AI-assisted labels were committed at
`6c794a66d17aa76dafa232d4f221102bec859344` before real outcome analysis. The CLI
checks their exact committed bytes. Ryder's instruction to do everything
autonomously supplied delegation; the report does not describe the resulting
labels as human ground truth or the revised analysis as the original primary
test. The modifier list was retained unchanged and no post-holdout tuning occurred.

Of 200 sampled rows, 176 remained unsure. Holdout agreement uses only 10 of 100
rows (90 unsure), including three labeled flankers: precision 42.9%, recall 100%,
F1 0.60. These small, selectively supported AI labels cannot establish reliable
classification accuracy. Same-year releases and catalog incompleteness remain
limitations. Missing classification values are reported explicitly; complete-case
sensitivity is separate from the frozen all-launch denominator.

## Actual sources and coverage

- Fragrantica: official public Kaggle dataset `ledecanteur/fragrantica-perfumes`,
  version 3, updated 2026-09-20; 140,228 catalog records. Both CSV and fuller JSONL
  were inspected; nested notes and rating histogram counts are used correctly.
  Neither file has concentration. Frozen exploratory window: **1980–2025**.
  Anonymous official download access made Kaggle credentials unnecessary.
- Parfumo: TidyTuesday 2024-12-10, 59,325 raw rows; 59,245 after the pre-registered
  removal of 10 embedded headers and all 70 rows in duplicate-ID groups. Window:
  **1980–2023**. Concentration, accords and release years have substantial
  missingness; gender and collection are absent. The upstream scrape date is
  unavailable. Selected-concentration flanker shares are all zero; inspected
  names retain brand/year text, undermining the frozen grouping rule. The
  all-concentration variant is also exposed to that encoding problem.
- FRED: UMCSENT pulled 2026-10-05; UNRATE/USREC reuse verified 2026-10-02 caches.
  Exact pull/update dates, raw checksums and software versions are in the results
  JSON. USREC yields recession event years 1980, 1990, 2001, **2008**, 2020 under
  the frozen rule, despite the prompt's expected 2007 onset.

Raw source bytes and sanitized sidecar hashes were validated before and after
processing. Some extracted-file provenance dates are 2026-10-06 UTC while their
local download folder is dated 2026-10-05 in New York; exact timestamps were
preserved. Offline reproducibility depends on retaining these ignored local raw
files and the recorded dependency versions. A fresh online vintage, independent
human labels, unavailable concentrations and future Trends pulls were not verified.

## Findings and repository isolation

None of H3a–H3c passes BH FDR .05. The secondary flanker event result has an
unadjusted p≈.021 and is subject to the classification limitations; it does not
replace the null directional lag-sum test. Parfumo selected-concentration gourmand
and fresh effects have opposite signs to Fragrantica, with wide intervals.
The latest comfort reading is **+3.6 pp**, December 2025, direction **steady**.

Completed stages were committed with explicit Gauge 3 paths and isolated
`git commit --only` operations. Other gauges, frontend and essay prose were not
modified by this work. Shared Gauge 3 execution targets, README documentation,
environment/ignore additions and public statistical source records are the only
project-level supporting changes.
