PULSCAR ANALYSIS — P0172 (SYSTEM TOO RICH, BANK 1) IN NHTSA'S PUBLIC FILES
PulsCar analysis of NHTSA public data, September 2026.
License: CC BY 4.0. Attribution: "PulsCar analysis of NHTSA public data, September 2026."
Published alongside https://www.pulscar.io/blog/p0172-code

------------------------------------------------------------------------------
Prepared 27 September 2026. Source data: NHTSA public flat files (complaints received 2015 to 21 September 2026, 780,071 records; manufacturer communications from January 2015; recalls from 2010) and documents downloaded from static.nhtsa.gov.

1. Complaints naming P0172 (87, all read in full, two coders)
   Every complaint whose narrative contains the string "P0172". 87 complaints; each read in full and coded by one reader to a written codebook (codebook.txt; categories A-G, plus flags for several warning lights at once and for a stall or loss of power). A second reader coded all 87 independently without seeing the first coding. Agreement on the category: 79 of 87 (90.8%), Cohen's kappa 0.86; on the warning-light flag 84 of 87; on the stall flag 83 of 87. 15 complaints had at least one disagreement (8 on the category, 3 on the warning-light flag, 4 on the stall flag); they were resolved by the codebook rules. p0172-87-complaints-read.csv publishes the final code and both coders' original codes (coder1_*, coder2_*) for every complaint, so the agreement figures can be recomputed; the resolution column marks the 15.
   Final categories: injectors diagnosed, replaced or recommended 45; other or undiagnosed 23; fuel pump or lines 6; engine damage 5; fuel in the oil or the 2018-2019 Honda campaign without an injector diagnosis 4; after other work 3; general report 1. Several warning lights at once: 41. Stall or marked loss of power: 31.
   By maker: Honda 63 (CR-V 50, Accord 5, Civic 3, Pilot 3, Fit 2), Kia 10, Chevrolet 4, two each Dodge and BMW, one each Ford, Mazda, Audi, Hyundai, Toyota, Jeep. By year received: 2018 1, 2019 4, 2020 1, 2021 9, 2022 4, 2023 16, 2024 16, 2025 15, 2026 (to 21 September) 21.
   For comparison, complaints naming P0175: 12; P0171: 109; P0174: 22 (string match).

2. Owner-reported amounts
   Every dollar figure in the 87, read in context (p0172-owner-reported-amounts.csv). Injector-set lane: 11 single figures, $1,300 to $2,300, median $1,689; two quoted ranges ($1,200-$2,500 and $1,500-$2,000) listed separately and not used in the median. Figures for unrelated repairs (a brake system, a transmission) were read and left out. Bills, quotes and estimates as owners reported them, not invoices.

3. Documents
   p0172-documents.csv. 120 rows: 93 manufacturer communications found by searching NHTSA's index for P0172 (one family per bulletin, all Honda bulletins individually), one Honda dealer message found through our Honda guides, two recall owner letters, and the other 24 of Honda's 26 parts-request dealer messages (June 2021 to October 2023), added after a second fact-check. One PDF (10223588) is byte-identical to another (10177533). Of the 119 distinct documents, 117 were read in full (two of the parts requests have no text layer and were read from the page images); two Toyota PDFs are images and are known only from NHTSA's summaries. Three Ford PDFs have a garbled text layer and were read from the rendered pages. 108 name P0172 in their own text; Honda's V6 injector campaign documents, Honda 19-072, Ford's 21L04 letter, the two image-only Toyota bulletins and the two recall owner letters (23V-634, 23V-858) do not. Ford TSB 18-2321 (superseded by 19-2091) is known only from NHTSA's summary and is not counted. A search of the recall file (2010 onward) for "P0172" returns no campaigns.
   Quotes: every quotation in the article was checked by script against the PDF text and the complaint texts (quotecheck.py); two — a Honda dealer message whose font drops ligatures and a Ford bulletin with a garbled text layer — were checked against the rendered page images.

Limits: complaints over-represent frightening failures; owners report the dealer's diagnosis as they understood it; two coders reduce reading error but do not verify repairs. Nothing here is a failure rate.