PulsCar P0496 / P0443 analysis: method
PulsCar analysis of NHTSA public data, October 2026.
Built 7 October 2026. License: CC BY 4.0. Credit "PulsCar" and link to https://www.pulscar.io/blog/p0496-code

SOURCES
- NHTSA ODI complaint flat files (COMPLAINTS_RECEIVED), 782,908 distinct vehicle complaints received 1 January 2015 to 4 October 2026 (latest received date in the files downloaded 7 October 2026).
- NHTSA manufacturer communication (TSB) flat files and the PDFs behind them, downloaded one at a time through NHTSA's public API; recall files (Part 573 reports, dealer bulletins, owner letters, chronologies, quarterly reports) for recalls 15V-793, 18V-735, 19V-515, 26V-369, 25V-794 and 25V-796; the NHTSA investigation flat file (RQ23-006, PE13-016, PE12-016).

COMPLAINT POOL (regular expressions are case-insensitive and given in codebook.txt)
1. P0496 or P0443 in any spelling (space, hyphen or period after the P, letter O for zero, suffix 00) or a bare 496 / 443 after a code word.
2. The purge component in owners' words: purge valve / solenoid / control, canister purge / close / vent, CPV, PCSV, CPS, purge flow / line / hose.
3. A fuel tank described as collapsed, deformed, crushed or under vacuum together with a tie word (vapor, vacuum, evap, recall, purge, canister, vent, CPV).
First search 1,672 complaints. An independent completeness audit returned 1,780 further candidates; 985 met the rule written and hashed before reading (purge words in other spellings, EVAP codes, tank collapse or expansion, purge-valve recall numbers); 795 narratives about a stall or no-start after refuelling that name no purge part, no EVAP code and no recall were not added. Final pool 2,657 complaints; 2,409 relevant (not category N).

CODING
Each complaint received one category (E engine harmed, T tank / pump module / sender / canister replaced or quoted, V purge valve replaced, repaired or quoted, W another EVAP part, O another cause, M software only, C checked and cleared, U nothing found or done, N not about the fault on this vehicle) and four yes/no flags (stall, refuel, covered, denied), under a codebook fixed and hashed before coding (hashes in codebook.txt).
Passes: two independent classification passes, blind to each other. A calibration of 200 random complaints (seed fixed before coding) was coded by the same two passes; they are chunk 0 of the pool (flagged calibration_set in the CSV).
Agreement over all 2,657 complaints (Cohen's kappa): category 0.895 (2,459 of 2,657 identical), stall 0.952, refuel 0.975, covered 0.855, denied 0.904. First batch 1,672: category 0.895; audit batch 985: category 0.871.
Every disagreement (376) was read in full and settled under written rules R1-R18 (published in codebook.txt). Two consistency sweeps of the agreed codes, driven by scripted candidate lists that were then read one by one, changed 135 rows (resolution = consistency pass in the CSV).
Both coders' codes and the final code are published per complaint; "rule_applied" names the rule or sweep. A coder sees the narrative (excerpts of up to 900 characters around each match; complaints are at most about 2,000 characters, so almost all were seen whole).

OWNER-REPORTED MONEY
Every dollar amount in a relevant narrative was extracted by script (384 mentions), then read and given one decision under one rule: keep an exact figure tied to the purge or EVAP repair or its diagnosis (classed purge valve job / tank, pump module or canister / diagnosis, test or fee / other EVAP part or fee, and quoted / billed / paid / offered / not stated); exclude lower bounds ("over", "plus"), ranges, approximate ("almost") figures, mixed bills for other repairs, other repairs, rentals, towing, vehicle values, loan balances, deductibles, part prices when the event is unclear and duplicates. 142 figures from 129 complaints kept; 242 mentions excluded, each with a reason held in our working files. These are what owners wrote, not invoices; they are not a price survey.

DOCUMENTS
727 files: 319 from the first search (tsbs summaries naming P0496 / P0443, or a purge valve with code-list entries stripped; plus 6 recall campaigns), 336 added by the audit, 72 recall documents. Byte-identical copies were detected by md5 (79) and near-duplicates by 5-word shingles with Jaccard similarity of at least 0.6 (221 members of clusters). 297 were read in full (including documents whose text layer was empty or garbled, read from page images), 130 were opened and confirmed not to name the codes or a purge valve. Every quotation in the article was checked by script against the document text.
Documents list: p0496-documents.csv (file id, source set, kind of entry, printed number, make, NHTSA date, whether it prints P0496 / P0443 / a purge valve, relevance, duplicate of, link).

LIMITATIONS
- Complaints are self-reported, unverified, and are not a failure rate. 61% of the relevant narratives cite a recall or program and the Ford Focus is 42% of the set: the file measures what reaches a federal form, which recall publicity drives.
- Counts are complaints, not cars. Raw counts across makes reflect very different sales volumes.
- Categories record what a narrative says was found or done; "valve" means the owner or a shop named it, not that a test proved it.
- Coverage status "as of October 2026" is our computation from printed terms; none of the documents prints a car's in-service date.
- Disagreements were settled under the written rules. The codebook, rules and both passes' codes are published so the classification can be checked.

Note on the complaints file: the codes_named column lists codes in standard spelling only; counts in the article of complaints that print P0496 or P0443 (34 and 11) use a spelling-tolerant pattern (for example PO496 with a letter O, p0496, Po496), reproduced in the names_p0496_or_p0443 column.