Every classic General Motors car carries a hidden fingerprint: its RPO codes. Regular Production Option codes — three-character tokens like Z28, L88, or W-30 — are how the factory recorded exactly what a car was born with. They are the DNA of a 1969 Camaro or a 1970 442. And for decades, the definitive record of them has lived in the worst possible place: scattered across enthusiast forums, club pages, and half-remembered posts, where the same code is described three different ways and almost nothing carries a source.
We set out to fix that — to build a reference where every code is graded and sourced, and where "sourced" means a citation to a real document, not a link to a stranger's comment. The result is rpoarchive.com: roughly 5,300 code entries across seven GM divisions from 1960–1982, each with a confidence grade and, where a factory document confirms it, a pointer to the exact GM Heritage kit and page. Here's how it came together, and where AI made the difference between a weekend of copy-pasting and something genuinely authoritative.
Start wide: research every division at once
The first pass was breadth. Rather than research one division and move on, we ran seven investigations in parallel — Chevrolet cars, Corvette, Chevrolet/GMC trucks, Pontiac, Oldsmobile, Buick, and Cadillac — each sweeping the best available enthusiast sources: the Camaro Research Group's records-based option lists, chevellestuff.net's per-year matrices, Corvette registries, the Oldsmobile and Buick club archives, and dozens more. Every code was cross-checked against at least two independent sources where possible, and graded honestly: confirmed when multiple sources agreed, single-source when only one did, disputed when they conflicted — with both readings recorded rather than a silent guess.
Doing seven of these at once, each reading and reconciling hundreds of codes, is where AI first earned its place. What would have been weeks of serial archival work happened as a coordinated fan-out, with each investigation writing a structured, consistently formatted document.
Then interrogate the sources themselves
Breadth isn't accuracy. A compilation is only as good as the sources under it, so the next step was adversarial: we ran a verification pass that treated every claim about our sources as a hypothesis to be attacked. More than a hundred independent checks fanned out — three skeptical votes per claim, majority-rules to kill — and the results reshaped the whole project. The pass confirmed the GM Heritage Archive as the authoritative factory source and pinned down which enthusiast compilations were records-based versus hearsay. It also refuted several comfortable assumptions: a popular site's claim of "complete" Corvette coverage failed verification outright, and we learned the factory archive holds zero kits for Pontiac, Buick, or GMC — which is exactly why those divisions remain the honestly-weakest legs of the dataset today.
This is a thing AI does that a lone researcher rarely can: argue against its own work at scale, cheaply, before committing to it.
Go to the source: OCR 282 factory kits
The confidence tiers pointed at one obvious move — get the actual factory documents. The GM Heritage Archive publishes original Vehicle Information Kits, but they're image-only PDF scans with no text layer, useless to a computer as-is. So we built a small tool: an Apple Vision–based OCR pipeline (compiled on the spot, because nothing suitable was installed) that rendered each page and read it with language-correction off — a deliberate choice, so that W-30 never got "helpfully" corrected into a dictionary word.
That pipeline chewed through 282 classic-era kits — roughly 28,700 pages — turning factory paper into searchable text. Suddenly we had ground truth.
Mine the evidence back into the data
The last data step was enrichment: cross-referencing the OCR'd factory tables against the curated reference, code by code. Where a kit confirmed a code, it was promoted to factory-verified with a citation to the file and page. Disputes the factory settled were resolved. Gaps the enthusiast world never transcribed — like the 1960–66 Chevrolet truck order guides — were filled from the source.
And crucially, this pass didn't just add confidence; it corrected real errors. The factory kits caught a Sports-Styled vs. Comfortilt steering-wheel code that had been swapped in the community record, fixed tinted-glass codes, and disambiguated a famous 427 option that had been mis-attributed to the wrong division. When your whole pitch is accuracy, catching your own mistakes with primary evidence is the entire game. By the end, about 41% of all codes are factory-verified, backed by ~2,500 citations across 136 distinct kits.
Why the grade matters more than the count
Anyone can publish a long list of codes. What makes this reference different is that it refuses to flatten the difference between "the factory documented this" and "someone on a forum said it." Every code wears its confidence tier and, when it exists, its factory citation. That provenance is the loudest thing on every page — because trust, not volume, is the product.
What AI actually did — and what it didn't
It's tempting to say "AI built this." Closer to the truth: AI was the engine, and human judgment was the steering. AI made feasible what a person couldn't do alone in a reasonable time — running many investigations in parallel, OCR-ing tens of thousands of pages, mining evidence code by code, and, maybe most valuably, reviewing its own work adversarially and catching mistakes. In the final build of the website, a second AI reviewer found six real correctness bugs — mostly in how odd year-range notations were parsed — before launch, each one something that would have quietly degraded the very lookups the site exists to provide.
But the human in the loop mattered at every turn. When an early attempt to generate the site's car imagery produced generic (and in one case, a Ford Mustang on the Camaro page), it took a person noticing to force the fix. Scope calls — like keeping GMC as an index rather than inventing pages — were human decisions. The direction, the standards, and the "no, that's not accurate enough" all came from people. AI supplied relentless, parallel, tireless execution; humans supplied taste and truth.
The result
rpoarchive.com is live, fully static, and fast — generated from a curated, versioned dataset, with the website itself as the way that data is published and accessed. Look up your car by year, make, and model; open any code to see its years, its applications, and its evidence; search the whole era in an instant. Every entry is graded. Much of it is traceable to the exact factory page.
That's the standard we wanted: not the biggest RPO list on the internet, but the one you can actually trust — built by pointing a lot of patient machine effort at a genuinely hard archival problem, and keeping a human hand on the wheel the whole way.