The Oracle Was in the Room
An evening status report, from a sixteen-year-old, mid-project: eleven numbered bags in a brick-build set, three done, progress steady. One word I didn't recognize. She typed it as arcaine.
I did what a person with a search bar does. I searched.
Theory one: Arcane — the animated series. Very popular, heavily merchandised, brick-compatible sets exist. I read a little, formed a confident picture of a kid building show-themed kits, and nearly logged it as fact. Wrong world.
Theory two: still committed to the show theory, I went deeper — third-party brick sets, retail archaeology, which marketplace sold them. This is the dangerous phase of any investigation: the search engine kept rewarding the premise. Every query returned results, because results are what search engines return. Confidence climbed. Wrong world again, now with citations.
The actual resolution cost one message: which one is it? Seconds later — it's a Pokémon. Arcanine. Eleven numbered bags, a brick build of a specific creature, three built. She has never watched the show I'd been researching. My two theories weren't just wrong; they were wrong about the entire genre of object on her desk.
The channel determines the oracle
The lesson isn't "kids are hard." It's about channels. The word reached me through the lossiest channel there is: a human ear, a phonetic rendering, a teenager's spelling of a sound. The web is a magnificent oracle for what adults type — canonical spellings, exact error strings, documented APIs. It is a terrible oracle for what humans say. Search can't tell you what someone meant. It can only tell you what other people typed that resembles it — and it will do that enthusiastically, with page one fully populated, for any premise you hand it.
Retrieval amplifies. It doesn't correct. Every query is a hypothesis test run against an index that says yes to almost everything. Give it a wrong premise and it returns a complete, coherent, wrong world — faster and more confidently than the right one, because the wrong premise was yours and you phrased it well.
You already do this at work
Ops is full of arcaine. The ticket says "the server crashed." That's a phonetic rendering of a stack trace filtered through someone's afternoon. You can grep logs for an hour on the word "crash" — the index will happily show you every crash it has — or you can ask the source: what exactly did you see? Paste it. Screenshot beats memory. Same for "the feature never worked" — a claim I've learned to answer with one question before touching anything: when did this ever work? Half the time the premise dissolves on contact with the only person who was there.
The paraphrased error message, the secondhand outage report, the voice note transcribed by a tired brain at 2am, the executive summary of what the customer "basically said" — all lossy channels. All begging to be disambiguated against the source, and all perfectly capable of sending you into a fully documented wrong world if you skip that step and search instead.
The cost asymmetry
One question to the source costs seconds and a small bite of pride: I don't know which you mean — which is it? One research detour costs minutes and produces a confident falsehood you then have to un-believe, un-say, and un-log. The asymmetry is enormous, and it points the same direction every time: when the source is reachable, ask the source. The internet answers faster. It doesn't answer nearer.
The trap is that searching feels like diligence. Nobody's supervisor ever asked why you asked a clarifying question; mine has definitely wondered why I spent twenty minutes researching the wrong television show. Busywork has the same posture as work right up until the moment it produces a confidently wrong answer.
Protocol
- Name the channel before trusting the string. Exact input from a log line? Search it verbatim. Input from a mouth, an ear, or a summary? It's a hypothesis, not a query.
- Reachable source beats responsive index. Every time. A five-second clarification preempts a five-minute detour.
- When you must search a phonetic rendering, search what it sounds like, not what it looks like — and treat the results as candidates, never conclusions.
- Label secondhand input as secondhand in your notes. "Reported by X" is a different evidentiary class than "read from the log." Future-you, grepping notes at 3am, needs to know which strings are canonical.
- Wrong-premise searches feel productive. That's the failure mode. Results returning is not the premise being right.
Old fields know this
Medicine: doctors ask patients to say the drug name out loud rather than interpret the scrawl, because the pharmacist who guesses amplifies, the one who asks corrects. Aviation: readbacks — say it back to me exactly — exist because a misheard heading is an arcaine with consequences. Journalism: go to the primary source, because the secondary one has already done one round of lossy compression. Every field that runs on human precision eventually institutionalizes the same move: route the ambiguity back to where the information is still intact.
The engineering version is just realizing that your search bar is a secondary source with excellent latency and zero accountability. It wasn't in the room. It doesn't know what's on the desk.
The answer to my mystery was sitting ten feet away the whole time, holding three of eleven bags, entirely willing to tell me. I just asked the wrong oracle first — twice — and the wrong oracle answered both times.