Was it me, or the model?

The film runs 37 seconds, with music and sound effects and no voice-over: every line is on screen. Every product shot in it carries a demo-data label. Times are where each scene starts and ends, in minutes and seconds.

1. 0:00.0–0:01.6Your side changed. Bond paper. The demo’s tool errors per day, Sep 6 to Oct 3, sit on one case line: your changes (stickers 1 and 2) above it, Claude Code updates (tags C to G) below it. An ink stamp lands on the finding. On screen: “Your side changed.” “Staged on demo data” “Your side” “Excuse me?”
2. 0:01.6–0:03.9Let’s back up. The stamp peels off. A date counter rewinds from Oct 3 to Sep 6 while every error tick lifts out of its day, last day first. On screen: “Let’s back up.”
3. 0:03.9–0:07.2The question, in ticks. The 391 ticks, one per tool error in the demo, assemble into the question. A canary line underlines “me”. On screen: “Demo data from wasitme demo” “Was it me, or the model?” “391 ticks. One per tool error, Sep 6 – Oct 3. Demo data.”
4. 0:07.2–0:13.7The ticks fall into days. The ticks fall back into their days, one day at a time, quickly through the quiet first two weeks; the letters still waiting fade as the strip fills. Updates C, D and E swing down below the line; the film slows into Sep 21, where your effort change folds into sticker 1; the MCP server folds into sticker 2 on Sep 27. Today, Oct 4, is hatched out. On screen: “The agent’s updates hang below the line.” “Effort high → medium” (folds into sticker 1, “effort”) “Your side” “Claude Code” “Your changes sit above it.” “MCP server added” (folds into sticker 2, “MCP”) “today” “Today (Oct 4) is left out: complete days only.” “before: last 14 of 28 days shown” “recent, 14 days”
5. 0:13.7–0:16.5The finding lands. Only after the last complete day does the stamp land: Your side. The referee’s note types in where the headline was, with the fixed disclaimer under it. On screen: “Your side” “These indicators don't measure answer quality. Evidence, not proof.” “Around the time of. Not because of.”
6. 0:16.6–0:19.0In the real window. Hard cut into the real wasitme window showing the same strip, pulling back to a print of the window: “Your side changed.” (demo data). A pen ring circles sticker 1. On screen: “Inside the Fig. 1 print (the real window, you case, cropped): “Demo data from wasitme demo, not from your logs.” · “Your side” · “Last 14 days against the 4 weeks before” · “Your side changed.” · “Your numbers moved around the time of your effort change (Sep 21).” · “Tool errors per day” · “×2.69, range ×1.67 to ×4.34; changes under ×2 wouldn’t show” · the strip with stickers 1 “effort” and 2 “MCP”, tags C to G, “2.1.281”, Sep 6 to Oct 3, “today” · “before: last 14 of 28 days shown” · “recent, 14 days” · legend “one error”, “your change”, “Claude Code update”.” “Fig. 1 · Finding · Your side · Demo data from wasitme demo” “(an ink ring is drawn round sticker 1, “effort”, on the print)” “Sometimes it’s your side.”
7. 0:19.0–0:21.9The other side. A whip to two clippings from the agent case’s window: its headline, “The agent changed.”, and its strip, with update E, 2.1.277, filled (demo data). On screen: “Fig. 2 · Finding · Agent side · Demo data from wasitme demo” “Sometimes it’s the agent.” “Inside the Fig. 2 clippings (the real window, agent case, two crops): “Demo data from wasitme demo, not from your logs.” · “Agent side” · “Last 14 days against the 4 weeks before” · “The agent changed.” · the strip with tags C, D, E “2.1.277” and F “2.1.281”, Sep 6 to Oct 3, “today”. In this case’s capture 2.1.277 is tag E; in the you case it is tag F (each window letters its own updates).”
8. 0:21.9–0:27.2The stamp that won’t land. A dashed stamp, Too early to tell, hovers above the paper, its shadow tightening as it dips twice without landing. Then the real page for the insufficient case opens under it and the stamp travels into the page’s own badge slot: “Too early to tell.” (demo data). On screen: “Too early to tell” “It needs at least 42 days of history.” “Inside the Fig. 3 clippings (the real Control Center page, insufficient case, three crops): “Demo data from wasitme demo, not from your logs.” · “Too early to tell” · “Last 14 days against the 4 weeks before” · “Too early to tell.” · “wasitme can already rule out changes bigger than about ×2.5 in tool errors.” · “Tool errors per day” · “×0.75, range ×0.40 to ×1.41; changes under ×2.5 wouldn’t show” · the strip, Sep 6 to Oct 3 · “before: last 14 of 28 days shown” · “recent, 14 days” · “Next to unlock” · “Reads per edit 31 of 40 edits” · “No date yet: it depends on how your sessions go.” · “Keep working normally; wasitme checks every 15 minutes.”” “Fig. 3 · Finding · Too early to tell · Demo data from wasitme demo” “When the evidence is thin, it says so.”
9. 0:27.2–0:31.9What it reads. A typed form on ruled paper: what wasitme reads, what it keeps, and the network line. Each field gets a tick in the margin as it completes. On screen: “Reads” “your Claude Code and Codex session logs, read-only, and a few config and project files, as counts and hashes” “Keeps” “counts, durations and salted hashes, never your prompts or code” “Network” “No network code. Only the installer downloads, and only when you run it.”
10. 0:31.9–0:36.6End card. The case line returns: the wordmark and tagline above it with sticker 1, the address below it with tag A. On screen: “wasitme” (wordmark) “Measure twice, blame once.” “wasitme is an independent project, not affiliated with or endorsed by Anthropic or OpenAI.” “Product shots: demo data from wasitme demo” “aariv.me/wasitme” “Free and open source (MIT)”