DJ Agent

No code moved in six weeks. What moved is what I know about the code — a fresh-context critic read the repo against my eval standard and filed the verdict (docs/evals/2026-07-25-methodology-audit.md). It found the deterministic half in good shape and the model half entirely ungraded, and the sharpest finding takes out the number the project has been steering by since June:

• Set acceptance was 100% by construction. The build plan named set-acceptance rate — "% of proposed sets I approve at the HITL gate" — as the number that says whether this is working. But the skill that actually runs a set mandates HITL_LEVEL=none, hitl.confirm then returns True without asking, and that True gets written into every history record. The flag can never say no, so it carries no information — and the exports (rekordbox XML, .m3u8, set sheet) get written on every run, before any human has said yes. The human-approves-the-plan rule holds in the architecture and is bypassed by the thing that invokes it.
• Both models fail silently. architect.plan_arc catches every exception and passes; the Selector's loop catches and breaks. A live model that times out, returns garbage, or gets talked out of its instructions degrades to the deterministic fallback with no log and no counter. The offline greedy path is the well-tested one; the live path is what ships by default.
• The thing called an eval grades the wrong pipeline. compare_to_acoustic_baseline defaults to no model, and its only caller never overrides that — so the A/B meant to protect the Selector never invokes it. The runner then returns 0 regardless of the result, so its exit code gates nothing.
• Track titles reach the prompt as-is. selector._format_request interpolates each candidate's title and artist straight into the model's context, and those strings come from YouTube via yt-dlp. A track named to look like an instruction is a live path into set selection. Untested.

What the audit says to keep is the part ADR 0006 got right on purpose: a key clash is a fact, not an opinion, so the verifier is math and never a model. The plan builds the case set around the Critic — the approval bug written as a failing case before it's fixed, then an offline tier that gates and a live tier that's allowed to fail while it's still a trend line.

Two ADRs landed in one day, and both close gaps between what the architecture claimed and what actually executed:

• allin1, finally (ADR 0012). The "allin1 → librosa fallback" chain had a secret: allin1 had never run. Dependency pins blocked the install, and once bridged, its attention layer materialized full T×T score matrices on CPU — 11.3 GB peak for a 30-second clip; a real track OOMed the machine. The fix is two parts: the allin1 CLI now runs as a subprocess from whatever Python can host it, JSON-cached per track, and a vendored memory-linear attention patch does the same math in O(T·k) — verified to float32 epsilon, and it restores the trained positional bias the upstream API migration silently dropped. Same clip now peaks at 2.07 GB, a full track at 5.2 GB. First real output: BPM 133 and 12 sensible sections on a bachata test track. Librosa is finally just the fallback.
• Genre profiles as data (ADR 0013). The critic's transition gates — BPM jump and harmonic compatibility — stopped being constants. A frozen profile table — nine genres plus an open default that is exactly the old behavior — feeds the architect, selector, critic, mixer and set sheet from one place. The same set with a 7-BPM jump and one key clash now passes as hip-hop and fails as house; that exact case is a unit test. Slots grew into entry → core → exit play spans sized to the genre's airtime window, crossfades quantize down to powers of two capped per genre, and approval writes a third artifact: a printable set sheet with per-transition bar counts and blend-style notes.

The fast suite went from 227 to 256 tests, still green offline.

The clap-first-rescope branch landed — the live pipeline now runs brief-to-render without a gap, and the test suite grew from 103 to 227 fast tests, all green without a database, a model, or audio on disk. Three ADRs went code-complete with it:

• Link ingestion (ADR 0009). Paste a Spotify or YouTube URL and the curator resolves metadata, pulls the audio via yt-dlp to FLAC, and feeds it the same idempotent ingest path as a local file. A bulk judge (dj.taste.judge <url>) scores a whole playlist against my taste vectors without downloading anything.
• Two-form export (ADR 0010). Every approved plan exports a rekordbox.xml for manual decks and an .m3u8 always; --render still produces the automatic .wav. The plan is the product — the render is one of its outputs.
• Grid truth (ADR 0011). Real downbeat phase estimation, section bounds snapped to the beat grid, and first_downbeat_s exported as the TEMPO anchor — so what rekordbox shows matches what the mixer assumed.

A full review pass landed across phases 0–5: 103 fast tests green, every deterministic module covered. The best catch was real: librosa's beat tracker folds double-tempo genres — DnB, footwork — down to half-tempo, and a naive octave correction elsewhere was quietly un-folding them back. Fixed before any real audio ever got graded by it.

Next gate is contact with reality: the first full ingest of my actual library, where the allin1 → librosa fallback chain finds out which segmentation backend actually installs and which genres break the beat grid.

The Critic is the interesting half: harmonic compatibility, BPM continuity, energy-arc RMSE, artist spacing — pure math, reused unchanged as the eval scorecard. The Selector revises against its notes until pass or budget (three revisions, then best-so-far wins). The console at the top of this bench plays that exact loop. With no API key the whole pipeline degrades to a deterministic arc and greedy nearest-fit ordering, which is also the A/B baseline the agents have to beat.

The strongest taste signal happens while listening — usually on Spotify, where there's no file to tag. So capture decouples from ingest: a review left in the moment ("Bicep — Glue: dreamy 3am closer, 5") parks in a pending_taste table and auto-applies when the Curator ingests a confident match — ISRC, or name plus duration. Anything ambiguous surfaces for a one-line manual resolve instead of guessing.

That last part is the design: a wrong auto-match would poison the taste layer silently, and the taste layer is the only thing here that can't be recomputed from the audio.

"Sounds similar" and "is my kind of track" are different questions, so every track carries both: a 512-d CLAP acoustic vector and a 384-d taste vector embedded from my own notes — separate columns on the same Postgres row, blendable per query. "Surprise me" leans acoustic; "my favorites" leans taste. Sections are first-class rows with their own vectors, because a set doesn't mix tracks, it mixes the right parts of tracks on musical boundaries.