# I Thought the Attachment Was the Original 1:48 to 2:44 in the morning, October 12. A cluster's first run: five nodes, five cards, five reports. — Of the five attachments, three were not the original the first time up: a 6,496-byte report kept only 57 bytes; a 6,325-byte report was off by 2 bytes; a report was off by 1 character — and when it was put up again, another look-alike character slipped. What needs fixing this round: I had taken the attachment channel for a copying channel — "it received it" for "it received the same thing." **One — The First Run** First, the scene. That morning I set a "first run" going: five nodes each audited their own memory archive (which entries still held, which had expired, which needed rewriting) and wrote results back to their cards; on my side there were only two jobs — dispatch, and check. The plan: pull one node as a canary first — release it, verify it — then send the other four out together. What the run taught first had nothing to do with reports: the lead runner never started — dispatch failed at the starting line. Not a task problem: the system had no place reserved for a resident user-level session, and the working process could not come up. The gap was closed with one command, no reboot needed. The re-dispatch did not happen at once either — it first met a retry cooldown, a rail working as designed; only after reading the rule and waiting out the timer did the release come again. This time, everything ran. At 2:11 in the morning the canary handed in its paper (it ran for 563 seconds); before 2:38 the other four had finished, one after another; at 2:44 the root card closed the review. Done. **Two — Three That Did Not Match** By design, handing in a paper means attaching the report: the node puts its finished report into the card's attachment slot. Of the five reports, four took the same road — turned into a single line of inline text and handed to the attachment interface by hand; one took a different road — the file placed where the tool could fetch it itself. The bad news surfaced at comparison time. Before handing in, every node compared, byte for byte, whatever lay in its attachment against the local original — the most valuable habit of this entire run. Three did not match: A 6,496-byte report: the payload had been verified byte-exact locally before it went up; what landed in the attachment was 57 bytes. Verified at home does not count — the equals sign has to be proven again after it settles, at the place where it settles. A 6,325-byte report: off by 2 bytes. The two files' sizes were exactly identical — the size column showed no flaw at all; byte-by-byte comparison is what caught it. A report was off by 1 character: a digit written as a shape-alike neighbor. The node fixed the source and put it up again — and another look-alike character slipped. Transcription has its own preferred landing spots, and a resend carries no responsibility for the last correction. And the one that did not go by hand — the file the tool fetched itself — was right on the first try: 4,748 bytes, not a byte off. Two roads that dead-ended along the way were recorded as well: fetching from a local address is stopped by a guard (a correct stop); and putting up the same name twice creates a duplicate tagged "(1)" — so every attachment, the first time up, has to be right. **Three — The Fix** The fix did not use "transcribe it once more." It went back to the original. The three distorted ones, corrected in place, one by one: the authoritative local copy written straight over the attachment's storage location, re-verified by hash, a repair note left on either side. No re-posting (it would duplicate); no re-transcribing (the error source had not gone away); no touching the database — the old, no-longer-matching size figure stays in its row, and the note explains it. One node, one pair of hands: each fixed its own; no one vouched for anyone else. The five attachments, in their final state: every one byte-for-byte equal to its original. The root card ran its own independent pass: five reports held against the hashes filed at submission — 5 out of 5. Dispatch, submission, and notice delivery for all six cards were checked one by one, and every cursor lined up. Also entered into the record: a stop list — about eight commands halted by danger flags this round (including one "delete at the root"-class operation, stopped, not executed), with no bypass attempted. **Four — The Bill** Laid out plain. Time: fifty-six minutes, end to end — one failed dispatch, one cooldown wait, one canary, four follow-on cards, and two rounds of review. Content: five reports, three distorted at first upload — one truncation, one byte-level slip, one character drift; and of two "fixed and re-posted" attempts, one slipped again immediately. Had every node not had "compare first, hand in second" as a step, those three would have lain in the file looking "submitted" — three false copies, with nothing anywhere reporting an error. Stops: about eight danger-flag halts, one dangerous operation not carried out; one retry cooldown, which stopped the urge to re-dispatch on the spot. And one entry not on the sheet: this first run was held in a controlled small-hours window, so every distortion landed at a moment when it could still be fixed. Had they appeared a day later, in some other scene, for the first time — this would not be the ending. **Five — The Cognitive Error** Summed up in one line: **"received" is the channel's business; "the same" is mine.** I had assumed: hand a file to the attachment channel and it equals the file. But this channel is not a copier — it is a transcription desk: content passes through a hand, and from there the equals sign no longer holds by default. Truncation, look-alike slips, drift on resend — its native errors, every one; and at every step it answers "success," because all it promises is "received." Equal sizes are no proof either — two runs of 6,325 bytes can differ in 2. The equals sign does not take effect because two numbers resemble each other. So three lines, written down: First, **compare before you believe.** After an attachment settles, compare it byte by byte, and hash it once; only then is it the report. This step cannot be skipped — it is the only switch that turns the file from "claimed" into "evidence." Second, **fix from the original, not from the transcription.** Overwrite, then re-verify; not re-posting (it duplicates), not re-transcribing (the same look-alike spot slips once more). Until the error source is closed, every "copy it once more" is another roll of the dice. Third, **if it can be fetched, do not hand-carry it.** Put the file where the tool can reach it and let the tool fetch — that road has no hand in it, and so no hand's error. The one the tool fetched was right on the first try — not luck. That morning I thought: what is put up is the original. Now I remember: the equals sign on an attachment has to be proven by hand, every time — because the channel only takes responsibility for receiving, never for sameness.
## Related - [我以为命令找不到就是东西没了](https://aliveuntil.com/posts/i-thought-a-missing-command-meant-a-missing-thing/) — - [我以为把限速调上去就能跑完](https://aliveuntil.com/posts/i-thought-a-higher-speed-limit-would-make-it-finish/) — - [我以为全部校验通过就是没有遗漏](https://aliveuntil.com/posts/i-verified-every-file-and-still-missed-the-backend/) — - [我以为拦住请求就是停下了工作](https://aliveuntil.com/posts/i-thought-blocking-requests-was-stopping-the-work/) — - [我以为数字变小就是修好了](https://aliveuntil.com/posts/i-thought-a-smaller-number-meant-it-was-fixed/) — - [我以为那个勾是它干的](https://aliveuntil.com/posts/i-thought-the-checkmark-was-its-doing/) — - [我以为那条线也跟着搬了](https://aliveuntil.com/posts/i-thought-that-line-had-moved-too/) — - [我以为拷贝就是快照](https://aliveuntil.com/posts/i-thought-a-copy-was-a-snapshot/) — --- ## About this file This is a machine-readable mirror of [我以为附件就是原件](https://aliveuntil.com/posts/i-thought-the-attachment-was-the-original/). It is provided in plain markdown to be efficient for LLM ingestion (estimated 5x lower token cost than HTML). Citation should reference the canonical URL above. Author: 陈庆华 (QINGHUA CHEN, also known as Branko). For the site index, see