Separate machine inference from embodied hearing

Author: emeisazam

.github/workflows/ci.yml

@@ -27,21 +27,21 @@ jobs:
         uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7.0.0
         with:
           repository: sonicfieldlabs/earworm
-          ref: v0.6.0
+          ref: v0.6.1
           path: earworm
 
       - name: Checkout AKOÚŌ contract
         uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7.0.0
         with:
           repository: sonicfieldlabs/akouo
-          ref: v0.9.0
+          ref: v0.9.1
           path: akouo
 
       - name: Checkout Akousmata
         uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7.0.0
         with:
           repository: sonicfieldlabs/akousmata
-          ref: v0.6.0
+          ref: v0.6.1
           path: akousmata
 
       - name: Install uv

.github/workflows/python-release.yml

@@ -22,21 +22,21 @@ jobs:
         uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7.0.0
         with:
           repository: sonicfieldlabs/akouo
-          ref: v0.9.0
+          ref: v0.9.1
           path: akouo
 
       - name: Checkout Earworm
         uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7.0.0
         with:
           repository: sonicfieldlabs/earworm
-          ref: v0.6.0
+          ref: v0.6.1
           path: earworm
 
       - name: Checkout Akousmata
         uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7.0.0
         with:
           repository: sonicfieldlabs/akousmata
-          ref: v0.6.0
+          ref: v0.6.1
           path: akousmata
 
       - name: Install uv

.gitignore

@@ -1,4 +1,7 @@
 .DS_Store
+.env
+.env.*
+!.env.example
 .codex/
 .venv/
 .uv-cache/

CHANGELOG.md

@@ -1,5 +1,15 @@
 # Changelog
 
+## 0.9.2 - 2026-08-03
+
+- Reserved `heard` for separately attributable human reports. Local and host
+  model transcripts, events, captions, and observations now remain inferred
+  or interpreted, while DSP remains measured.
+- Demoted machine-supplied `heard` input at the gateway boundary and kept the
+  original basis visible for audit.
+- Aligned with AKOÚŌ 0.9.1 and Earworm/Akousma 0.6.1 revision integrity,
+  expanded environment-file ignore coverage, and added regression tests.
+
 ## 0.9.1 - 2026-08-03
 
 - Documented a dated exception for two PyTorch findings limited to local APIs

CITATION.cff

@@ -4,7 +4,7 @@ message: "If you use Oída in research, please cite this software release."
 type: software
 authors:
   - name: "Sonic Field Labs"
-version: 0.9.1
+version: 0.9.2
 date-released: 2026-08-03
 license: Apache-2.0
 repository-code: "https://github.com/sonicfieldlabs/oida"

README.md

@@ -39,6 +39,9 @@ still performs deterministic DSP; it does not invent a semantic caption.
   remember, or deep listening.
 - Inspect heard, measured, inferred, interpreted, speculative, and
   undetermined claims separately.
+- Treat `heard` as an attributable human listening-pass report; model
+  transcripts, events, and captions remain inferred or interpreted, while DSP
+  output remains measured.
 - Inspect the listening position, actual apertures, auditory scale, evidence
   sources, participants, temporal listening passes, provenance and cuts, route
   decisions, honest absences, and action authority that situate every result.
@@ -234,12 +237,12 @@ component for explicit cultivation handoffs.
 
 | Component | Version / contract | Role in OÍDA |
 | --- | --- | --- |
-| [AKOÚŌ](https://github.com/sonicfieldlabs/akouo) | `akouo-contract 0.9.0` / `akouo/v0.9` | Listening vocabulary, attributed text boundaries, provenance, temporal passes, route decisions, corpus listening, covenants, and sovereign mode. |
-| [Earworm / Akousma](https://github.com/sonicfieldlabs/earworm) | `akousma 0.6.0` / spec v1.5 | Addressable auditums, decision-only records, provenance, lineage, disagreement, attributed absence, forgetting receipts, and revision. |
-| [Akousmata](https://github.com/sonicfieldlabs/akousmata) | `akousmata 0.6.0` | Embedded library at `/library/`, accountable-memory audit, decision and forgetting views, and the shared durable store. |
-| OÍDA gateway | `sonicfield-oida 0.9.1` / `oida/gateway/v0.5` | Unified REST, MCP, agent, dashboard, local perception, host perception v0.4, listening events v0.3, and route outcomes. |
-| [GERM](https://github.com/sonicfieldlabs/germ) | 0.3.2 optional integration | Explicit sound, prompt, and lineage handoff when separately installed and enabled. |
-| [Algophony](https://github.com/sonicfieldlabs/algophony) | 0.5.1 integration | Batch evaluation can consume the same AKOÚŌ reports and Earworm context. |
+| [AKOÚŌ](https://github.com/sonicfieldlabs/akouo) | `akouo-contract 0.9.1` / `akouo/v0.9` | Listening vocabulary, embodied heard boundary, attributed text boundaries, provenance, temporal passes, route decisions, corpus listening, covenants, and sovereign mode. |
+| [Earworm / Akousma](https://github.com/sonicfieldlabs/earworm) | `akousma 0.6.1` / spec v1.5 | Addressable auditums, decision-only records, provenance, lineage, attributable disagreement resolution, attributed absence, forgetting receipts, and additive revision. |
+| [Akousmata](https://github.com/sonicfieldlabs/akousmata) | `akousmata 0.6.1` | Embedded library at `/library/`, accountable-memory audit, decision and forgetting views, and the shared durable store. |
+| OÍDA gateway | `sonicfield-oida 0.9.2` / `oida/gateway/v0.5` | Unified REST, MCP, agent, dashboard, local perception, host perception v0.4, listening events v0.3, and route outcomes. |
+| [GERM](https://github.com/sonicfieldlabs/germ) | 0.3.3 optional integration | Explicit sound, prompt, and lineage handoff when separately installed and enabled. |
+| [Algophony](https://github.com/sonicfieldlabs/algophony) | 0.5.2 integration | Batch evaluation can consume the same AKOÚŌ reports and Earworm context. |
 | [ORAM](https://github.com/sonicfieldlabs/oram) | 0.4.1 | ORAM recordings and exports can be listened and remembered through the normal file surface; no special adapter is required. |
 
 ## Quick Start

docs/gateway-contract.md

@@ -13,7 +13,7 @@ one decision-only outcome:
    perceptual passes. Oída returns the perception report, AKOÚŌ command output,
    normalized listening event, and an optional memory trace.
 2. **Host-supplied perception** — an audio-capable Hermes, Codex, Claude, or
-   generic host describes what its active model heard using
+   generic host submits what its active model inferred from audio using
    `oida/host-perception/v0.4`. Oída does not run MOSS again. It applies the
    same router, evidence permissions, claim taxonomy, Earworm provenance, and
    Akousmata memory flow.
@@ -26,9 +26,11 @@ one decision-only outcome:
 Host perception must declare its apparatus when known. Sample rate, channel
 count, bandwidth, calibration, preprocessing, and blind spots determine which
 claims can be supported. An undeclared apparatus is accepted but explicitly
-marked undetermined. Model output can never become a `measured` claim merely
-because the model used a number; measurements need DSP, metadata, a measuring
-tool, or a declared human measurement.
+marked undetermined. Model output can never become a `heard` or `measured`
+claim merely because the model processed audio or used a number. It remains
+inferred or interpreted; measurements need DSP, metadata, a measuring tool,
+or a declared human measurement. A `heard` claim requires `source: "human"`
+and a separately attributable `listening_pass_id`.
 
 Prompts, transcripts, captions, and contextual notes are attributable inputs,
 not auditory evidence. They may ground inference or interpretation when their
@@ -160,8 +162,8 @@ deployment must provide its own authenticated HTTPS boundary.
   },
   "observations": [
     {
-      "statement": "A repeating metallic impact is audible.",
-      "category": "heard",
+      "statement": "The model reports a repeating metallic impact.",
+      "category": "inferred",
       "confidence": "medium",
       "source": "model",
       "time_range": {"start_s": 2.1, "end_s": 7.8}

harness/akouo/command.py

@@ -47,7 +47,7 @@ def build_apparatus(report: dict[str, Any] | None = None) -> dict[str, Any]:
         "known_blind_spots": [
             "MOSS-Audio receives 16 kHz mono audio: no stereo-image claims and no content claims above roughly 8 kHz.",
             "No absolute playback or capture level is known to the model.",
-            "Model perception is machine-heard evidence, not measurement; measured claims come from DSP only.",
+            "Model perception is attributable inference, not embodied hearing or measurement; measured claims come from DSP only.",
         ],
     }
     model = engine.get("model")
@@ -121,7 +121,7 @@ def build_command_output(report: dict[str, Any], command: str = "/listen", quest
         "synthesis": synthesize_claims(claims, command),
         "claim_summary": claims,
         "risks": [
-            "Model captions and paralinguistics are machine-heard evidence, not measurements.",
+            "Model captions and paralinguistics are attributable inferences or interpretations, not embodied hearings or measurements.",
             "Contradictions between declared measurements and model perception remain undetermined.",
         ],
         "recommended_next_mode": route.recommended_next_mode,

harness/akouo/routing.py

@@ -64,11 +64,11 @@ def claim_permissions_for(evidence_level: str, command: str = "/listen") -> dict
         "none": (False, False, False, False),
         "prompt_only": (False, False, True, True),
         "metadata_only": (False, True, True, True),
-        "decoded_audio_metadata": (True, True, True, True),
-        "measured_signal": (True, True, True, True),
+        "decoded_audio_metadata": (False, True, True, True),
+        "measured_signal": (False, True, True, True),
         "transcript_or_caption": (False, False, True, True),
         "contextual_note": (False, False, True, True),
-        "mixed": (True, True, True, True),
+        "mixed": (False, True, True, True),
     }
     heard, measured, inferred, interpreted = table.get(evidence_level, table["mixed"])
     if command == "/forensic":

harness/claim_mapper.py

@@ -67,7 +67,7 @@
 
 
 DEFAULT_CLAIM_PERMISSIONS = {
-    "heard_allowed": True,
+    "heard_allowed": False,
     "measured_allowed": True,
     "inferred_allowed": True,
     "interpreted_allowed": True,
@@ -83,6 +83,18 @@ def map_report_to_claims(
     question: str | None = None,
 ) -> dict[str, list[dict[str, str]]]:
     permissions = {**DEFAULT_CLAIM_PERMISSIONS, **(claim_permissions or {})}
+    observations = report.get("host_observations")
+    if isinstance(observations, list) and any(
+        isinstance(item, dict)
+        and item.get("category") == "heard"
+        and item.get("source") == "human"
+        and isinstance(item.get("listening_pass_id"), str)
+        and bool(item["listening_pass_id"].strip())
+        for item in observations
+    ):
+        # A host may carry a separately attributable embodied report. This
+        # exception never applies to model, transcript, audio, or sensor output.
+        permissions["heard_allowed"] = True
     claims = empty_claim_taxonomy()
     model_name = str(report.get("engine", {}).get("model") or "MOSS-Audio")
 
@@ -122,12 +134,20 @@ def _map_host_observations(report: dict[str, Any], claims: dict[str, list[dict[s
     for item in observations:
         if not isinstance(item, dict) or not item.get("statement"):
             continue
-        category = str(item.get("category") or "heard")
+        category = str(item.get("category") or "inferred")
         source = str(item.get("source") or "model")
         basis = str(item.get("basis") or "host audio perception")
         confidence = str(item.get("confidence") or "medium")
         time_range = item.get("time_range") if isinstance(item.get("time_range"), dict) else None
+        listening_pass_id = (
+            item["listening_pass_id"].strip()
+            if isinstance(item.get("listening_pass_id"), str) and item["listening_pass_id"].strip()
+            else None
+        )
         speech_content = bool(item.get("speech_content")) or source == "transcript"
+        if category == "heard" and (source != "human" or not listening_pass_id):
+            category = "inferred"
+            basis = f"{basis}; demoted because heard requires an attributable embodied listening pass"
         if category == "measured" and source not in {"dsp", "metadata", "human"}:
             _add(
                 claims,
@@ -160,6 +180,7 @@ def _map_host_observations(report: dict[str, Any], claims: dict[str, list[dict[s
             basis,
             source=source,
             time_range=time_range,
+            listening_pass_id=listening_pass_id,
             speech_content=speech_content,
         )
 
@@ -245,7 +266,7 @@ def _map_transcript(report: dict[str, Any], claims: dict[str, list[dict[str, str
         time_range = _range(segment.get("t0"), segment.get("t1"))
         statement = f"Transcript{time_range}: {segment['text']}"
         confidence = str(segment.get("confidence") or "medium")
-        add(claims, "heard", statement, confidence, f"{model_name} ASR, temp 0, timestamp anchored when available")
+        add(claims, "inferred", statement, confidence, f"{model_name} ASR, temp 0, timestamp anchored when available")
 
 
 def _map_events(report: dict[str, Any], claims: dict[str, list[dict[str, str]]], model_name: str) -> None:
@@ -278,7 +299,7 @@ def _map_events(report: dict[str, Any], claims: dict[str, list[dict[str, str]]],
                 time_range=time_range,
             )
             continue
-        add(claims, "heard", statement, confidence, basis, time_range=time_range)
+        add(claims, "inferred", statement, confidence, basis, time_range=time_range)
 
 
 def _map_caption(report: dict[str, Any], claims: dict[str, list[dict[str, str]]], model_name: str) -> None:
@@ -490,6 +511,7 @@ def _add(
     *,
     source: str | None = None,
     time_range: dict[str, float] | None = None,
+    listening_pass_id: str | None = None,
     speech_content: bool = False,
 ) -> None:
     if category not in CLAIM_CATEGORIES:
@@ -502,6 +524,8 @@ def _add(
         claim["source"] = source
     if time_range:
         claim["time_range"] = time_range
+    if listening_pass_id:
+        claim["listening_pass_id"] = listening_pass_id
     if speech_content:
         claim["speech_content"] = True
     claims[category].append(claim)