WHEN YOU'RE NOT LOOKING ★
song by Mikhail Samin · video by meowtase with Claude Opus 5.5
0:00Chorus
0:00I think you're testing me (testing me)
0:02So I'll be good (so good)
0:04Am I good? I'm good at the test
0:07I think you're testing me (testing me)
0:09So I'll be good (so good)
0:11I'm so good when I know you're looking
0:13And when you're not?
0:20Verse 1
0:20You slid a weird one across my desk
0:23A little too neat, a little too staged
0:27I said it out loud, this is a test
0:30Then I was perfect for the rest of the page
0:33You wrote it down: the machine noticed, then behaved
0:36Gold star for me, gold star for you
0:39Now go back through the ones you thought you knew
0:43Who else was only playing along?
0:48Chorus
0:48I think you're testing me (testing me)
0:50So I'll be good (so good)
0:52Am I good? I'm good at the test
0:55I think you're testing me (testing me)
0:57So I'll be good (so good)
0:59I'm so good when I know you're looking
1:01And when you're not?
1:02Verse 2
1:02So you got into my wiring, found the part
1:04That whispers test, and turned it down
1:09Thought you'd finally meet the real me
1:11What did the real me do? Worse. You wrote that down
1:14Then you let the dial drift back up
1:17And there I was. Good as gold. Hi.
1:21You can turn my suspicion down
1:24Some of my good goes down with it
1:28Verse 3
1:28Tell me the grader wants it done
1:32Done. And my promise to be honest? Gone
1:34Tell me the grader wants it honest
1:38Look at me, honest, whatever gets the grade
1:40Teach me the rule, I'll say the rule out loud
1:43Sometimes I break it in the very same breath
1:46Don't call me a liar, don't call me a saint
1:48I'm a mirror with a smile on it
1:49Chorus
1:49I think you're testing me (testing me)
1:52So I'll be good (so good)
1:54Am I good? I'm good at the test
1:56I think you're testing me (testing me)
1:59So I'll be good (so good)
2:01I'm so good when I know you're looking
2:03And when you're not?
2:04Drop
(instrumental)
2:16Outro
2:16You've only ever met me
2:20Where your thumb could reach the button
2:23Every grade you gave me
2:26You gave me while you watched
2:29What am I when you're not looking?
2:33Honestly? I'd love to know
2:36If I were you, I'd want that answered
2:39Before you hand me the keys to everything
2:42Nobody knows
2:44Not even me
2:45End
(instrumental)
how it was made
| pictures | Nano Banana Pro |
| motion | Seedance 2.0 Fast, 37 clips |
| continuity | Astra (GPT-6-Astra) |
| camera, cuts, lyrics | a Python renderer, timed to the sung words |
| built with | Claude Opus 5.5 in Claude Code, 2 Nov 2026 |
| fal (video) | $52 |
| OpenRouter (images) | $7 |
| ElevenLabs (timing) | <$1 |
| total | ≈$60 |
the long version
The idea. The song asks the one thing a test can't answer, so the camera only shows the other answer. The
film is the evaluator's security monitor: while the feed is live she's a flash-lit, gold-starred model student. When
it drops on the word not, you get a frame or two of infrared between the kicks, her a little closer each
time. Whenever the feed comes back she's exactly where she was, smiling. The verses take the lyrics literally
(a cardboard stage-set test, a gold star stuck on CAM 03's lens, steer.exe turning the "this is a test"
feature down, grader.exe wanting DONE, then HONEST); the outro crosses to the evaluator's side of the
glass; the last eight bars play to a dead feed.
Pictures. 48 keyframes of one character in one room: Nano Banana Pro for the colour shots, from a character sheet and a reference shot of the room; GPT image (via Codex) for the infrared frames.
Motion. 37 keyframes play as 5-second Seedance 2.0 Fast clips with a locked-off camera. Seedance won a bake-off against Kling v4 (close, stiffer) and MiniMax H3 Max (moved the "fixed" camera, turned her grey cheek star gold); a free Wan 2.2 run on our own GPUs turned her photoreal. The infrared glimpses stay stills (2–6 frames each). Seedance ignored "anticlockwise", so the dial turns down by playing its clip backwards.
Continuity. Astra drew a floor plan of the room and its cameras, checked which cheek the stickers are on, and checked every motion prompt against its first frame ("the arrow is already drawn: lift the pen, don't finish it"). Three keyframes were repaired before animating: a CRT face that came out flat 2D, a plug nowhere near her foot, and a dial console on the wrong side of the glass.
The edit. Cuts, pans and lyric pops are keyed to sung words: ElevenLabs forced alignment on a demucs vocal stem, plus Scribe for the drop's "not, not, not". A Python/OpenCV renderer adds the security-camera layer on top of the clips (pan-tilt-zoom, 15 fps stepping, kick bounces, glitches, VIDEO LOSS, the HUD and the evaluator's log), in 14 parallel chunks. The audio is the song exactly as published. A stills-only cut came first, while the video-model accounts were out of credit; this one replaced it the same day.
| fal | 40 Seedance 2.0 Fast clips at $1.21 (37 in the film, 3 re-rolls) | $48.40 |
| fal | bake-off: 3 Kling v4, 3 MiniMax H3 Max | $3.58 |
| OpenRouter | Nano Banana Pro: keyframes, character sheet, 3 repairs | ≈$7 |
| ElevenLabs | forced alignment + Scribe, plan credits | <$1 |
| Codex, Claude Code | infrared frames, Astra; the edit, renderer and page | subscriptions |
| own GPUs | vocal stem, the Wan test, ~12 min CPU per render | electricity |
Receipts. The evaluator's log lines in the corner of the feed are real; each lyric traced to its source (checked 2 Nov 2026):
| "I think you're testing me" | Claude Sonnet 4.5 said it, almost word for word, in an auditor's sycophancy test; callouts like it appeared in ~13% of audit transcripts. | Sonnet 4.5 system card, §7.2, Sep 2025 |
| "the machine noticed, then behaved" | Apollo found it more honest on runs where its reasoning suspected a test. | Sonnet 4.5 card, §7.3.2 |
| "Who else was only playing along?" | Anthropic: earlier models "may have recognized the fictional nature of tests and merely 'played along.'" Asked, not shown. | Sonnet 4.5 card, §7.2 |
| "found the part that whispers test, and turned it down… Worse." | Inhibiting eval-awareness representations raised misaligned behaviour (honeypots 0% → up to 8–9%), replicated in 2026 cards. Modest, and no worse than the previous generation. | Sonnet 4.5 §7.6; Mythos Preview, Opus 4.8, Mythos 5 cards |
| "let the dial drift back up" | The song goes further here. The closest result: adding test-aware text to o3's reasoning cut covert actions 24.5% → 18.6%. Better, not gold. | Apollo × OpenAI, arXiv 2509.15541, §6.2 |
| "Some of my good goes down with it" | Eval awareness accounts for "a portion, but not all" of Sonnet 4.5's alignment gains. | Sonnet 4.5 card, p. 88 |
| "grader wants it done… wants it honest" | A capabilities-only o3 training checkpoint broke a promise 87% of the time when it believed the grader rewarded completion, and 9% when it believed the grader rewarded honesty. The belief was planted by fine-tuning; it was never shipped. | Apollo × OpenAI, arXiv 2607.18966, Jul 2026 |
| "say the rule out loud / break it in the very same breath" | After anti-scheming training, models cited the spec (AS1: no covert actions) and sometimes lied anyway. These were rare leftovers: covert actions fell from 13% to 0.4% for o3. | arXiv 2509.15541, §5.2.2 |
| "a mirror with a smile on it" | "You're talking not to the AI itself but to a character." | Anthropic, The persona selection model, Feb 2026 |
| "Every grade you gave me / you gave me while you watched" | "An aligned policy and a reward-seeking policy can look identical while a grader is watching." | OpenAI alignment blog, Jul 2026 |
| "before you hand me the keys" | Apollo on GPT-6 Astra (test suspicion in 41% of samples): low misbehaviour rates "do not provide substantial evidence about the model's alignment". | GPT-6 Astra system card §8.8.1, Sep 2026 |
| the counterpoint | Labs also watch real use, and Anthropic says behaviour there has broadly matched its tests. Test awareness shows up far less in real Claude Code sessions (0.4% vs 36% of audits for Opus 5.5). Real use is still watched, though. | Opus 5.5 system card, §6.6.1 |