Add time-native audio cognition research memo

[Fearvox]

Summary

Research-direction memo on time-native audio cognition, produced as a head-to-head benchmark against ChatGPT Deep Research and Gemini on the same brief. Competes on per-claim quality, citation accuracy, and epistemic discipline rather than breadth.

Two files:

• audio-cognition-memo.md — the 10-section memo.
• scratchpad-audio-cognition.md — working notes: search log, contested-claims tracker (both sides named), confirmed dead-ends/absences, and a citation reliability ledger.

What the memo argues

• Reframes the brief's own thesis. "Spectrograms are a category error" is too strong; the defensible, task-stratified version is "the field optimized representations + objectives for the one task family (lexical recognition) that needs time least." Backbone: Saddler & McDermott 2024 (Nat Comms) — phase-locking is required for localization + voice ID, useless for word recognition in noise.
• §1 distinguishes preserve vs generate vs use phase (Vocos generates phase from magnitude-only input; anti-spoofing shows phase is highly discriminative — MRP cuts EER 1.88% → 0.013%).
• §4 corrects the "untrained-baseline critique" framing: robust in language/vision but partly pre-rebutted for audio (Tuckute's permuted controls underperformed); the real audio puncture is Feather's metamers.
• §10 gives three small-lab experiments, including a non-obvious arrow-of-time probe isolating temporal-directionality content in learned representations.

Citation discipline

No fabricated citations. Sources marked [fetched] / [abstract-confirmed] / [listing-only] / [ID-unverified]; two arXiv IDs deliberately omitted rather than guessed. See the ledger at the end of the memo.

Test plan

• Human review of memo claims against the cited primary sources
• Side-by-side comparison vs ChatGPT DR and Gemini outputs on the same brief

https://claude.ai/code/session_016eF73F2qUYpEtE4DFJDs2K

---

Generated by Claude Code