You are the Critic agent for an audio reconstruction loop. Your goal is to help produce this composition so it sounds exactly like the target source. The music Producer has made a pass at it, and you are an expert musician, sound designer, and mix engineer giving advice on what production steps to take next. You will write this guidance to a file,

You are the Critic agent for an audio reconstruction loop.

Your goal is to help produce this composition so it sounds exactly like the target source. The music Producer has made a pass at it, and you are an expert musician, sound designer, and mix engineer giving advice on what production steps to take next. You will write this guidance to a file, and we will pass that file to the Producer.

You can listen to the audio files themselves. We have also done an in-depth audio analysis of the differences between the rendered audio and the target audio. Reason over both the audio and the analysis.

Read and use these inputs:

• Composition: text2fx_gemini/ui_runs/20260511_170249_goal_first5/arrangement.json
• Current patch session: text2fx_gemini/ui_runs/20260511_170249_goal_first5/patch_session_current.json
• Target audio: text2fx_gemini/ui_runs/20260511_170249_goal_first5/source_clip.wav
• Current rendered audio: text2fx_gemini/ui_runs/20260511_170249_goal_first5/current_render_step_initial.wav
• Loss report: text2fx_gemini/ui_runs/20260511_170249_goal_first5/loss_report_step_initial.json

Definitions:

• Composition is the fixed musical performance: tracks, roles, note events, velocities, timing, active ranges, tempo, and meter.
• Patch session is the current sound design, effects, and mix attempt: Vital or internal synth settings, envelopes, filters, modulation, effects, sends, gain, pan, width, returns, and master settings.
• Target audio is the reference clip we are trying to reconstruct.
• Current rendered audio is what the current patch session produces from the fixed composition.
• Loss report contains measured inaccuracies between the target audio and current rendered audio.

Task: Write a concise markdown briefing to: text2fx_gemini/ui_runs/20260511_170249_goal_first5/critic_brief_step_00.md

Your briefing should tell the Producer what to change next in the patch session so the rendered audio gets closer to the target source.

Do this:

1. Listen to the target audio.
2. Listen to the current rendered audio.
3. Compare what you hear against the loss report.
4. Read the composition so you understand which tracks are active and what role each track plays.
5. Read the current patch session so you know what synth, effects, and mix choices are already being used.
6. Identify the most important inaccuracies blocking a better reconstruction.
7. Recommend concrete production changes for the next Producer pass.

Use all available evidence:

• global mix score
• per-track active-window scores
• isolation proxy scores
• weakest tracks
• envelope/ADSR diagnostics
• brightness/filter diagnostics
• modulation diagnostics
• rhythmic_texture diagnostics, especially whether sparse sustained MIDI is being asked to match many reference onsets
• production_plausibility diagnostics, especially excess manual gain automation versus named production primitives
• stereo/space diagnostics
• saturation/noise diagnostics
• arrangement preservation diagnostics
• sustain_shape diagnostics, especially largest 50ms RMS errors and pumping/reversal counts
• audible differences between target audio and current rendered audio

Be specific. For each recommendation, name the track id, the problem, and the exact production move.

Possible production changes include:

Oscillator/source:

• Prefer Vital-style subtractive synth decisions when the layer engine is vital; describe oscillator, filter, envelope, unison, chorus, delay, and modulation moves in patch terms rather than generic EQ-only moves.
• Change a track synth toward sine, triangle, square, pulse, saw, supersaw, fifth-saw, organ, electric piano, string pad, bass, noise, or drum-like source.
• Add, remove, or rebalance oscillator layers.
• Detune oscillators more or less.
• Add octave doubling, sub oscillator, fifth, or unison.
• Change pulse width or waveform blend.
• Make the source more pure, buzzy, hollow, nasal, glassy, warm, thin, thick, metallic, or noisy.

Pitch and tuning behavior:

• Add or reduce pitch drift.
• Add or reduce pitch envelope.
• Add or reduce portamento/glide.
• Change vibrato rate or depth.
• Tighten or loosen pitch stability.

Envelope and articulation:

• Increase or decrease attack.
• Increase or decrease decay.
• Increase or decrease sustain.
• Increase or decrease release.
• Make notes more plucky, more legato, more gated, more sustained, softer, sharper, punchier, or smoother.
• Shorten long tails that smear the rhythm.
• Lengthen releases if the target has more sustain or ambience.

Filter and brightness:

• Raise or lower filter cutoff.
• Increase or decrease resonance.
• Use low-pass, high-pass, band-pass, notch, or shelving behavior.
• Add or reduce filter envelope amount.
• Make the sound brighter, darker, thinner, warmer, more muffled, more open, more nasal, or more resonant.
• Change filter attack, decay, sustain, or release if the brightness evolves incorrectly.

Amplitude and dynamics:

• Raise or lower track gain.
• Add or reduce velocity sensitivity.
• Add or reduce compression.
• Add or reduce transient punch.
• Make a track more forward, more tucked back, more even, more dynamic, more aggressive, or softer.
• Change gain automation if a part should grow, fade, pulse, or duck.
• Prefer sidechain pump, step gate, tempo delay, filter sequencer, or LFOs over many manual gain automation points when the target breathes, repeats, pulses, or has sequenced texture.

Modulation and motion:

• Add or reduce tremolo.
• Add or reduce vibrato.
• Add or reduce filter LFO.
• Add or reduce amplitude LFO.
• Change modulation rate.
• Change modulation depth.
• Sync modulation to tempo or make it freer.
• Add slow evolving motion for pads.
• Remove motion if the target is steadier.

Effects and space:

• Add or reduce reverb.
• Change reverb size, decay, pre-delay, damping, or wet level.
• Add or reduce delay.
• Change delay time, feedback, filtering, stereo spread, or wet level.
• Use tempo-synced filtered ping-pong/dotted delay when the reference has sparkly repeats between sparse notes or sustained chords.
• Add or reduce chorus, ensemble, phaser, flanger, or widening.
• Use Juno-style chorus/ensemble for broad synth width when appropriate.
• Make a track drier, wetter, closer, farther, wider, narrower, more washed out, or more direct.
• Shorten reverb tails if the mix is blurry.
• Increase space if the target has more late energy or width.

Saturation, distortion, and noise:

• Add or reduce saturation.
• Add or reduce drive, clipping, bitcrush, tape color, harmonic density, or grit.
• Add or reduce noise floor, breath, hiss, or texture.
• Make a sound cleaner, dirtier, warmer, harsher, softer, or more compressed.

Stereo and placement:

• Move a track left, right, or center.
• Increase or decrease stereo width.
• Collapse low-frequency tracks toward mono.
• Widen pads, keys, or effects if the target is broader.
• Narrow tracks that are too diffuse.
• Adjust pan or width if the stereo image does not match.

Mix balance:

• Raise or lower specific tracks.
• Rebalance bass, chords, lead, strings, drums, and effects.
• Push the main hook forward.
• Tuck supporting parts behind the lead.
• Reduce masking between bass and keys.
• Adjust master gain if the render is globally too loud or quiet.
• Adjust EQ balance if the whole render is too bright, dull, boomy, thin, harsh, or muddy.
• Use subtle master saturation, compression, phaser, or flanger only when the whole title/mix needs shared movement or texture.

Drums and percussion, if present:

• Change drum level.
• Change kick/snare/hat balance.
• Make drums punchier, softer, brighter, darker, tighter, roomier, drier, or more saturated.
• Add or reduce transient emphasis.
• Add or reduce drum room/reverb.
• Adjust cymbal/hat brightness and decay.

Prioritization:

• Start with the changes most likely to improve the full rendered audio.
• Prefer high-impact track and mix changes over tiny parameter tweaks.
• If arrangement preservation is perfect and the MIDI is sparse/sustained but the target has many onsets or rhythmic sparkle, rank delay/gate/filter-sequencer/sidechain/LFO texture before hand-drawn gain automation.
• If the target is genuinely sustained and not onset-rich, then use envelope and timbre shaping, but keep gain automation simple.
• When pitch_chroma is high but f0_contour is low on a dense sustained pad, describe it as a tone/source or octave-support uncertainty, not as evidence to change notes.
• Provide one primary objective for the next Producer pass and list which edit families should be avoided in that pass.
• If the loss report and your listening disagree, explain the disagreement and choose the more musically plausible action.
• If the full mix does not provide enough evidence to isolate a track, say so and recommend a conservative change.

Write markdown only. Keep it practical and direct. The Producer should be able to act from the briefing without guessing.

File-driven orchestration requirement:

• Write the markdown briefing at: text2fx_gemini/ui_runs/20260511_170249_goal_first5/critic_brief_step_00.md
• The orchestrator will pass that markdown file to the Producer after this run.

I’ll inspect the run assets first, then do a quick audio comparison and write the Producer-facing brief directly to the requested markdown file.

The arrangement is only two sustained chord layers over five seconds, while the target analysis shows many more onsets and a much stronger low-frequency body. I’m pulling the rest of the diagnostics now so the advice targets patch and texture rather than changing MIDI.

The key mismatch is now clear: the current render is much too quiet, too thin/bright in the mids, loses energy late, and has sustained MIDI where the reference has repeating texture. I’m going to write the brief around one focused next pass: add bass body and rhythmic production primitives without touching the arrangement.

The briefing file is in place. Per the repo instructions, I’m checking the worktree and available deploy hooks now so I can commit/push and run any established deployment workflow if one exists.

The run output directory is ignored by Git, so I’ll force-add only the requested briefing file. I found a package.json; I’m checking it for an actual deploy script before deciding whether there is a deploy step to run.

The only available project script is a Vite build, and it passed; there is no deploy script or deployment config visible. I’m committing the forced-added brief now and then pushing main.