Which prices, in what order: the inflation was a relay, and shelter ran the last leg

Reachability swept first: yes — every FRED series run through the real connector and its semantics read from FRED's own metadata; the BLS weights fetched from bls.gov. One search result was an instrument failure and is recorded under Open questions so the method is visible. Vintage checked: yes — CPI series via FRED lastUpdated 2026-09-11 (the August 2026 release); Cleveland Fed median CPI 2026-09-12; BLS relative-importance tables for December 2020, 2021, 2022 and 2023. The not-seasonally-adjusted CPI is never revised, which makes the contribution figures permanent — the safest property any sheet in this buffer has.

Segment: Which prices, in what order

• Source theme(s): Batch 54 #5
• Spine role: main story
• The framing people carry: that inflation rose and fell as one thing — prices went up, then the Fed brought them down.
• Hypothesis (from the bank), four sub-claims: (1) the contribution passed between components in sequence, each peaking months apart; (2) shelter kept contributing after the components that started the episode had turned negative, "because of how it is measured"; (3) the headline's peak and the median component's peak arrived at different times; (4) the disinflation came from a few components reversing "while the rest never did".
• Verdict: CONFIRMED on all four as tested — with the mechanism half of (2) NOT tested, and the bank's stated ORDER right by onset and wrong by peak. (1) The relay is real. (2) Energy's year-on-year rate turned negative in March 2023, the very month shelter peaked. (3) The median CPI peaked eight months after the headline. (4) Between 2021 and 2024 food, energy and core goods removed 4.64 points of inflation between them while core services added 0.49 — it did not merely fail to reverse, it grew.

The numbers (every row sourced — no exceptions)

The texture (what narrative flattens)

• It was a relay, not a wave. Goods and energy started it in 2021; by 2023 they were contributing nothing or less than nothing, and core services — mostly shelter — carried almost the whole headline. In 2021 core services were 31% of the inflation; in 2023 92%.
• The handover happened in a single month. Energy's annual rate went negative in March 2023 — the month shelter hit its peak of 8.2%. The component that lit the fire was falling as the one that burned longest reached its top.
• The headline and the typical price told different stories. Headline peaked in June 2022; the median price change kept climbing until February 2023. For eight months the headline said "falling" while the middle of the distribution said "still rising".
• The disinflation was narrow. Every point of it came from food, energy and goods. Core services never came back down — their contribution in 2024 was higher than in 2021.

Datoid artifact (Tier-A only)

• Already seeded: CPIAUCSL (headline) and CPILFESL (core). Every other series here is ingestable through the existing FRED connector, verified by running fetchFredSeries — 115 monthly points for the SA series from 2017-01, 80 for the NSA from 2019-12, each one short of the month count because October 2025 is absent (see Open questions).
• The weights are NOT a series. BLS relative importance is a static annual table on bls.gov, not a FRED series. Contributions are therefore derived, not ingestable, and a dashboard would carry them as a computed line — which the platform can currently only draw as a sum (derived-series.ts), not a product of weight × change.
• The relay is a story about RATES, and the platform plots LEVELS. Every finding above is year-on-year change. The chart has no percent-change transform yet — that is exactly the open fork recorded in #79(b). This sheet cannot become a dashboard without either that transform or pre-computed rate series. It is the largest build gap of any sheet in the buffer.
• Scraping the weights from HTML is fragile. The tables parsed cleanly for all four years, but ri2022/ri2023 are markedly smaller pages than ri2020/ri2021 — the layout changed — and any automated ingest should parse by row label, as this did, never by position.

Story justification (decide here — do NOT defer)

• Does the argument have a reveal that needs sequencing? YES — BUILD.
• Because: the relay is a sequence. Shown as one chart of seven lines, the handover is a tangle; shown in order — goods and energy rising, then falling, then shelter taking over as they go negative — it is the argument. The median peak arriving eight months late is a second reveal that only lands after the headline peak has been shown.

Open questions / where I'm guessing

• THE SHELTER MECHANISM IS NOT TESTED, and "because of how it is measured" must not be spoken as established here. The bank's claim is that the CPI rent measure lags market rents because it samples the stock of existing leases. Testing it needs a market-rent index (the BLS New Tenant Rent Index, or Zillow's ZORI). Neither is reachable through our connectors: a FRED search for "new tenant rent", "tenant rent index" and "Zillow observed rent" returns BEA regional price parities and Consumer Expenditure items, not a monthly market-rent series. The timing is confirmed; the cause is Tier B and untested.
• THE FIRST SEARCH WAS AN INSTRUMENT FAILURE, recorded because the method matters. All three market-rent queries first returned zero — and a control query that had to hit ("consumer price index shelter") also returned no count. The API key had been extracted from .env by a shell one-liner that mangled it; the connector, loading it through dotenv, had worked all along. Re-run with the key loaded correctly, the control returned 270 hits and the market-rent zeros became real. Without the control this sheet would have asserted "no market-rent series exists" on a broken instrument — the births/deaths error in a new form.
• THE ORDER DEPENDS ON WHAT IS MEASURED — say which. The bank's "energy, then goods, then shelter, then services" is exactly right by onset and wrong by peak, where core goods (Feb 2022) led energy (Jun 2022) and core services (Feb 2023) led shelter (Mar 2023). Both orders are true; a script that says "the relay ran energy-goods-shelter-services" must say starting.
• Food's onset is contaminated. Food first exceeded 4% in June 2020 — the pandemic grocery spike, not the 2021 episode. It is excluded from the onset order for that reason.
• Shelter's share is a trajectory, not a single number. 63% in 2023 is the most quotable figure and 57% in 2024 the most recent; quote the rise (20% → 63%) rather than one year. Part of the rise is weight, not inflation: shelter's relative importance grew from 33.3 to 36.2 over the period.
• The identity check confirms the METHOD, not the data. BLS computes the headline as exactly this weighted aggregate, so a perfect close proves the weights were parsed correctly, applied at the right base, and paired with the NSA indices — not that the CPI measures inflation well. It is a stronger check than the FAO or World-Bank-zero identities only in that a parsing or base error would have shown.
• Peaks are SA, contributions are NSA — deliberately: seasonal adjustment is right for dating a monthly peak, and BLS's official December-to-December change is NSA. They are not mixed within any single figure.
• The median CPI is a different producer (Cleveland Fed), computed from BLS components. Comparing its peak date with the headline's is a timing comparison, not an identity, so the producer-mixing that broke two earlier sheets does not apply the same way.
• OCTOBER 2025 IS MISSING FROM EVERY CPI SERIES — SA and NSA, headline and components alike, verified by enumerating the months each series returns. The 2025 US federal government shutdown is the likeliest cause and is not verified here. It affects no finding on this sheet (every peak, contribution and overlap falls outside that month), but it matters downstream: a year-on-year rate for October 2025 or October 2026 cannot be computed, and a dashboard line will carry a gap. Found because a review queried a point count that looked off by one — both counts were right, and the discrepancy was the data, not the arithmetic.
• Four aggregates only. "The rest never did" is established at the level of food / energy / core goods / core services; at finer granularity some sub-components may have reversed.