Algophony v0.3 — Executed Listening Contract

Algophony v0.3 Release Notes

Date: 2026-07-10

v0.3 turns the listening contract from documentation into an executed layer,
aligned with AKOÚŌ v0.6 and Earworm v0.2 across the Listening Stack.

Listening-contract execution (new)

• workers/listening_plan.py: deterministic AKOÚŌ routing plans without any
  LLM. Artifact availability derives the evidence level (prompt-only →
  prompt_only, generation metadata → metadata_only, analysis features →
  measured_signal, everything together → mixed); the evidence level plus
  command overrides derive claim permissions; missing inputs become stop
  conditions instead of imagined listening.
• scripts/generate_reports.py attaches a routing plan and the
  akouo_contract_version pin to every new report and enforces claim
  permissions: blocked claims move to undetermined with an explicit marker,
  never silently dropped.
• scripts/test_listening_plan.py covers evidence derivation, permissions,
  enforcement, schema conformance of built plans, and a drift check against
  the machine-readable AKOÚŌ manifest, so contract copies cannot silently
  diverge from the source of truth.

AKOÚŌ v0.6 contract sync

• 14 listening modes (adds memory-lineage-listening, the sound-memory ear)
  and 17 commands (adds /remember).
• Claims may carry source (audio, dsp, metadata, model, transcript,
  context, memory, human, other) and time_range anchors.
• Mode outputs may declare akouo_version, apparatus (substrate, perception
  sources, known blind spots), listener, and memory links; routing plans
  may carry budget and preset_id; reports pin akouo_contract_version.
• Synced across schemas/listening-report.schema.json,
  apps/web/app/lib/listening-contract.ts, and apps/web/app/lib/types.ts;
  the report page renders contract, budget, and preset when present.

Earworm v0.2 / akousma spec v1.1 batch surface

• workers/akousmata_source.py reads both raw (v1.0) and enveloped (v1.1)
  listening entries, prefers record summaries, and surfaces summary, typed
  relations, consent status, and pipeline effects on prompt records.
• New kinship and maintenance operations: related(), add_relation()
  (e.g. compares_with across an evaluation batch, idempotent),
  find_by_hash() recurrence lookup, and verify_store() integrity reports
  (absence is information). Query filters extend to tag/text/since/until on
  v0.2 stores with graceful fallback on older ones.
• CLI: --related and --verify flags.

Provenance and openness instrumentation

• Provider openness profiles derived on the registry and shown on the
  /providers page: open_source_internal, open_weights_local,
  open_code_hosted, closed_api — the literacy-or-capture axis made
  operational.
• compute_provenance stamped from the registry at generation time in the
  pipeline (local hardware vs hosted endpoint vs cloud API); never guessed
  after the fact.
• Standing review questions documented as the review protocol
  in docs/benchmark-methodology.md, mapped to existing fields.

Corpus discipline (unchanged)

No backfills: the v0.1.1 corpus keeps null routing plans, traces, provenance and consent
axes, and compute provenance. New fields activate only through actual routed,
traced, or scored passes.