Freeze independent English German speech and boundary validation inputs

.gitattributes

@@ -11,3 +11,5 @@ fixtures/mixed/sources/** -text
 fixtures/music-validation-v1/*.json -text whitespace=blank-at-eol,blank-at-eof,space-before-tab,cr-at-eol
 tools/rerank-eval/*-plan.json -text whitespace=blank-at-eol,blank-at-eof,space-before-tab,cr-at-eol
 tools/rerank-eval/*-manifest.json -text whitespace=blank-at-eol,blank-at-eof,space-before-tab,cr-at-eol
+
+fixtures/speech-validation-v1/*.json -text whitespace=blank-at-eol,blank-at-eof,space-before-tab,cr-at-eol

fixtures/speech-validation-v1/README.md

@@ -0,0 +1,100 @@
+# Independent English/German speech preparation v1
+
+No ASR, VAD or Google inference has run on this fixture. The ten topic/absence
+queries and two source recordings were frozen before preparation. The sources
+are human LibriVox readings of Article 1 of the Universal Declaration of Human
+Rights, unrelated to the earlier NASA topics. They share translated content and
+are correlated; this is not a representative multilingual speech benchmark.
+
+| Source | Human reader | Recording permission | Original bytes |
+| --- | --- | --- | ---: |
+| [English Article 1](https://commons.wikimedia.org/wiki/File:Universal_Declaration_of_Human_Rights_-_eng_-_drb_-_Art1.ogg) | David Barnes | LibriVox public-domain dedication / copyrighted free use | 168186 |
+| [German Article 1](https://commons.wikimedia.org/wiki/File:Universal_Declaration_of_Human_Rights_-_ger_-_robu_-_Art1.ogg) | Roman Buettner | LibriVox public-domain dedication / copyrighted free use | 151136 |
+
+The publisher pages explicitly identify the human readers, document and language.
+The primary reference text is [UN Article 1 in English](https://www.un.org/about-us/universal-declaration-of-human-rights)
+and [UNRIC Article 1 in German](https://unric.org/de/allgemeine-erklaerung-menschenrechte/).
+The manifest records reference keywords and topic paraphrases, never model
+captions or recognition-derived ground truth. Full primary text is externally
+linked; it must not be supplied as an ASR prompt or indexed as if recognized.
+
+Only **319322 new bytes** were downloaded. Sources are checked against official
+Commons API size/SHA-1 and acquired SHA-256. The existing independently licensed
+US Navy drum cadence supplies the instrumental control, with its previously
+locked source bytes. Including that reused source, original inputs total
+1893869 bytes. Generated silence is exact zero PCM, released CC0.
+
+## Four-clip pool and honest timestamp scope
+
+Prepared audio is mono 16-kHz s16 WAV. Each human recording remains complete,
+with five seconds of exact leading silence; no reference words are cropped.
+The resulting English/German durations are 22.512 and 16.219 seconds. Music and
+silence controls each cover exactly thirty seconds. Core WAVs total 3159568 bytes.
+The music waveform is the first thirty seconds of the already locked drum source;
+no descriptive tags are used to gate recognition.
+
+Actual source audio stream starts are verified as zero, and the five-second
+padding/PCM source mapping is exact. **The publishers do not provide phrase-level
+timed captions.** Neither the padding boundary nor the whole article bounds prove
+the first phoneme or each sentence's start. Strict utterance-point scoring is
+explicitly unavailable; do not count a start anywhere in the article as precise
+localization success. Independent listening/timed annotations remain required.
+
+Six positive EN/DE queries describe equal dignity/rights, reason/conscience and
+brotherly treatment. Either language's recording is relevant: translated topic
+equivalence is intentional. Four absent queries ask about Mars funding and salary/
+paid leave. Silence and instrumental music remain in every candidate pool, and
+any recognized hallucination remains an explicit measured failure/competitor.
+Per-language recognized concept coverage is separate from semantic file recall.
+
+## Separate seam primitives
+
+`seam-plan.json` freezes two artificial repeat controls, outside the four-clip
+retrieval pool. English source samples repeat twice and German samples three
+times after the same known padding, producing 40.024 and 38.657 seconds. Their
+2517880 additional WAV bytes and exact piecewise source maps stress a 30-second
+window with two-second overlap. These joins are artificial repetitions, not
+natural uninterrupted speech. Do not extrapolate their quality to long talks.
+
+Independent byte checks verified each derived piece equals the original decoded
+source PCM and the prefix is all zeros. Compare full-source VAD on these short
+controls with bounded independent windows, recording interval disagreement,
+source coverage, seam omissions/duplicates and processing cost. Restarted VAD
+need not reproduce a continuously initialized VAD's probabilities. Preserve
+differences rather than changing thresholds or treating an unverified reference
+as infallible. The same labels must not choose gate intervals or discard a seam.
+
+## Preparation and acceptance
+
+```powershell
+python -m unittest discover -s tools/speech-eval -p test_prepare_validation.py
+python tools/speech-eval/prepare_validation.py fixtures/speech-validation-v1/fixture.json .searchi/speech-validation-v1-seams .searchi/media-runtime/bin/ffmpeg.exe .searchi/media-runtime/bin/ffprobe.exe
+```
+
+Preparation already completed; existing files are preserved, so use a fresh
+owned directory for another preparation. Five structural tests passed. Exact
+padding and every seam's source pieces were checked independently without a
+model. No recognition/quality score follows. Hashes are frozen in `inputs.lock.json`:
+
+- Fixture: `e3a850d7a2385e250408f57fe51f4620ce80a20f8625ced42205f02cc46d517d`.
+- Seam plan: `5da4400814c4965fc12a87bb00c36665300730a702a69acd65a72b9f5201ac9e`.
+- Prepared `.searchi/speech-validation-v1-seams/prepared.json`:
+  `085f34f0af863656a28e7a905746973a54d1123bb07053541e09c8af0a58bb3f`.
+
+The earlier four-clip-only plan
+`4635e772834b0eed847459f5f4d47d7c99a591f0adba2c60ce75712557092279`
+is preserved under its original ignored directory and was never inferred.
+
+Before accepted scoring, require every source/window and all ten queries, strict
+UTF-8 and bounded times, zero excluded failed competitors, unchanged worker/model/
+decoder hashes and exact source-byte provenance. Report genuine German recognition
+and all preregistered reference concepts, source-level recall/MRR, absent-return
+rate and gate false positives separately. Both negative controls must retain
+zero accepted recognized speech segments; raw Google audio search stays available.
+Measure decode/VAD/ASR/alignment/Google/storage, cold model load, foreground impact
+and warm reuse separately. ASR remains bounded to thirty seconds even for seam
+controls. Pause between work units for foreground priority; do not claim native
+preemption during one inference call.
+
+This plan does not adopt a provider, change production search or close fine
+timestamp, long-source, accent, representative-noise or human-listening gates.

fixtures/speech-validation-v1/fixture.json

@@ -0,0 +1,185 @@
+{
+  "schema": 1,
+  "frozen_before_inference": true,
+  "assets": [
+    {
+      "id": "speech-en",
+      "kind": "speech",
+      "language": "en",
+      "source_url": "https://upload.wikimedia.org/wikipedia/commons/e/e7/Universal_Declaration_of_Human_Rights_-_eng_-_drb_-_Art1.ogg",
+      "source_page": "https://commons.wikimedia.org/wiki/File:Universal_Declaration_of_Human_Rights_-_eng_-_drb_-_Art1.ogg",
+      "bytes": 168186,
+      "sha1": "4bfeec5eab9837eff3ca8318f2824e7e6982526d",
+      "license": "LibriVox public-domain dedication and copyrighted free use",
+      "author": "David Barnes",
+      "license_url": "https://librivox.org/pages/public-domain/",
+      "reference_text_url": "https://www.un.org/about-us/universal-declaration-of-human-rights",
+      "reference_section": "Article1",
+      "reference_keywords": [
+        "free",
+        "equal",
+        "dignity",
+        "rights",
+        "reason",
+        "conscience",
+        "brotherhood"
+      ],
+      "prefix_silence_ms": 5000,
+      "fine_utterance_timestamps_available": false,
+      "speech_offset_note": "Exact padding5000ms does not establish first phoneme or individual sentence times."
+    },
+    {
+      "id": "speech-de",
+      "kind": "speech",
+      "language": "de",
+      "source_url": "https://upload.wikimedia.org/wikipedia/commons/d/d8/Universal_Declaration_of_Human_Rights_-_ger_-_robu_-_Art1.ogg",
+      "source_page": "https://commons.wikimedia.org/wiki/File:Universal_Declaration_of_Human_Rights_-_ger_-_robu_-_Art1.ogg",
+      "bytes": 151136,
+      "sha1": "7f5d3620b4939d44849639c10b415920071353a3",
+      "license": "LibriVox public-domain dedication and copyrighted free use",
+      "author": "Roman Buettner",
+      "license_url": "https://librivox.org/pages/public-domain/",
+      "reference_text_url": "https://unric.org/de/allgemeine-erklaerung-menschenrechte/",
+      "reference_section": "Article1",
+      "reference_keywords": [
+        "frei",
+        "gleich",
+        "Würde",
+        "Rechten",
+        "Vernunft",
+        "Gewissen",
+        "Brüderlichkeit"
+      ],
+      "prefix_silence_ms": 5000,
+      "fine_utterance_timestamps_available": false,
+      "speech_offset_note": "Exact padding5000ms does not establish first phoneme or individual sentence times."
+    }
+  ],
+  "controls": [
+    {
+      "id": "music-control",
+      "kind": "instrumental-negative",
+      "source_fixture": "fixtures/music-validation-v1/fixture.json",
+      "source_id": "unseen03",
+      "source_bytes": 1574547,
+      "source_sha256": "5083f07730c6ffb1cd2c3b5cc1a43714a0cfda04fe913935e83d1048466562e6",
+      "license": "PD-USGov USNavy",
+      "clip_duration_ms": 30000
+    },
+    {
+      "id": "silence-control",
+      "kind": "silence-negative",
+      "generation": "480000 exact zero s16 samples at16kHz mono",
+      "clip_duration_ms": 30000,
+      "license": "CC0-1.0"
+    }
+  ],
+  "queries": [
+    {
+      "id": "equal-en",
+      "text": "someone explains that every person has equal dignity and rights",
+      "language": "en",
+      "expected": [
+        "speech-en",
+        "speech-de"
+      ],
+      "category": "spoken-topic",
+      "strict_point_score_eligible": false
+    },
+    {
+      "id": "equal-de",
+      "text": "jemand sagt dass alle Menschen gleiche Würde und Rechte haben",
+      "language": "de",
+      "expected": [
+        "speech-en",
+        "speech-de"
+      ],
+      "category": "spoken-topic",
+      "strict_point_score_eligible": false
+    },
+    {
+      "id": "reason-en",
+      "text": "a speaker mentions reason and conscience",
+      "language": "en",
+      "expected": [
+        "speech-en",
+        "speech-de"
+      ],
+      "category": "spoken-topic",
+      "strict_point_score_eligible": false
+    },
+    {
+      "id": "reason-de",
+      "text": "eine Person spricht über Vernunft und Gewissen",
+      "language": "de",
+      "expected": [
+        "speech-en",
+        "speech-de"
+      ],
+      "category": "spoken-topic",
+      "strict_point_score_eligible": false
+    },
+    {
+      "id": "solidarity-en",
+      "text": "people should treat one another with brotherhood",
+      "language": "en",
+      "expected": [
+        "speech-en",
+        "speech-de"
+      ],
+      "category": "spoken-topic",
+      "strict_point_score_eligible": false
+    },
+    {
+      "id": "solidarity-de",
+      "text": "Menschen sollen einander mit Brüderlichkeit begegnen",
+      "language": "de",
+      "expected": [
+        "speech-en",
+        "speech-de"
+      ],
+      "category": "spoken-topic",
+      "strict_point_score_eligible": false
+    },
+    {
+      "id": "absent-mars-en",
+      "text": "a speaker explains funding for a mission to Mars",
+      "language": "en",
+      "expected": [],
+      "category": "absent-topic",
+      "strict_point_score_eligible": false
+    },
+    {
+      "id": "absent-mars-de",
+      "text": "jemand erklärt die Finanzierung einer Marsmission",
+      "language": "de",
+      "expected": [],
+      "category": "absent-topic",
+      "strict_point_score_eligible": false
+    },
+    {
+      "id": "absent-salary-en",
+      "text": "someone explains their annual salary and paid leave",
+      "language": "en",
+      "expected": [],
+      "category": "absent-topic",
+      "strict_point_score_eligible": false
+    },
+    {
+      "id": "absent-salary-de",
+      "text": "eine Person erklärt ihr Jahresgehalt und den bezahlten Urlaub",
+      "language": "de",
+      "expected": [],
+      "category": "absent-topic",
+      "strict_point_score_eligible": false
+    }
+  ],
+  "acceptance": {
+    "complete_pool": "All4clips and all10queries retained; zero excluded invalid windows; invalid outputs fail accepted scoring",
+    "recognition": "Report per-language transcript/reference concept coverage separately; publisher references never become inference prompts or indexed passages",
+    "retrieval": "ActualGoogle textadapter, allrecognized passages inclnegativehallucinations; positive fileR1/R5+MRR and absentreturn separate",
+    "timing": "No fine utterance publisher captions: precise startscore unavailable, do not credit wholeclip as precise. Independent listening/timedlabels required before fine localization acceptance",
+    "gate": "Zero accepted recognized segments for exactsilence and instrumentalcontrol; retain gatingdecision/source and failures; no endpoint/rawGoogle suppression",
+    "runtime": "CPU1thread on VM; ASRwindowmax30s; repeatedoperation/sourcehash guard/explicitoffsetmap; coldload separately measured"
+  }
+}

fixtures/speech-validation-v1/inputs.lock.json

@@ -0,0 +1,119 @@
+{
+  "fixture_sha256": "e3a850d7a2385e250408f57fe51f4620ce80a20f8625ced42205f02cc46d517d",
+  "seam_plan_sha256": "5da4400814c4965fc12a87bb00c36665300730a702a69acd65a72b9f5201ac9e",
+  "prepared_sha256": "085f34f0af863656a28e7a905746973a54d1123bb07053541e09c8af0a58bb3f",
+  "frozen_before_inference": true,
+  "model_inference_performed": false,
+  "clips": [
+    {
+      "id": "speech-en",
+      "sha256": "0aa1fa97d367d987bfb3bb01d7eca11db77a544552ccb8dc0a78ad74d0b496ea",
+      "samples": 360192,
+      "source_sha256": "d69ddbb1dbf64bae10bb03ff74edc840b08c1cd08c172ca524fa73cb2c98ee48",
+      "source_audio_start_ms": 0.0,
+      "prefix_silence_ms": 5000
+    },
+    {
+      "id": "speech-de",
+      "sha256": "4af09fe7d80056600e4c288161697a094108f343b069ffd581ccabc06a048afe",
+      "samples": 259504,
+      "source_sha256": "fb180307f9d14ced5687c6d4a59d389508d952ead568c2b83b7b8eea65e58f61",
+      "source_audio_start_ms": 0.0,
+      "prefix_silence_ms": 5000
+    },
+    {
+      "id": "music-control",
+      "sha256": "af76cb9e35948fcf81c72f5dcb95165c578a8bcc28d756c96aef0e3c1466b8f5",
+      "samples": 480000,
+      "source_audio_start_ms": 0
+    },
+    {
+      "id": "silence-control",
+      "sha256": "40abc924dc63f75aafb2400bb99fd66741e22eb9ff2450973a18db53a06286ac",
+      "samples": 480000,
+      "source_audio_start_ms": 0
+    }
+  ],
+  "seam_controls": [
+    {
+      "id": "seam-en",
+      "sha256": "d4553fac71e86adbd4eced7a1c4750782c63935501446697cf5bc7b48870b31d",
+      "source_sha256": "d69ddbb1dbf64bae10bb03ff74edc840b08c1cd08c172ca524fa73cb2c98ee48",
+      "samples": 640384,
+      "piecewise_map": [
+        {
+          "derived_samples": [
+            80000,
+            360192
+          ],
+          "source": "speech-en",
+          "source_samples": [
+            0,
+            280192
+          ],
+          "copy_index": 0
+        },
+        {
+          "derived_samples": [
+            360192,
+            640384
+          ],
+          "source": "speech-en",
+          "source_samples": [
+            0,
+            280192
+          ],
+          "copy_index": 1
+        }
+      ],
+      "prefix_silence_samples": 80000
+    },
+    {
+      "id": "seam-de",
+      "sha256": "6b2adaab5fd505dd313b49b8878fc0b372b5141d80c589dcf06fba37ad699c00",
+      "source_sha256": "fb180307f9d14ced5687c6d4a59d389508d952ead568c2b83b7b8eea65e58f61",
+      "samples": 618512,
+      "piecewise_map": [
+        {
+          "derived_samples": [
+            80000,
+            259504
+          ],
+          "source": "speech-de",
+          "source_samples": [
+            0,
+            179504
+          ],
+          "copy_index": 0
+        },
+        {
+          "derived_samples": [
+            259504,
+            439008
+          ],
+          "source": "speech-de",
+          "source_samples": [
+            0,
+            179504
+          ],
+          "copy_index": 1
+        },
+        {
+          "derived_samples": [
+            439008,
+            618512
+          ],
+          "source": "speech-de",
+          "source_samples": [
+            0,
+            179504
+          ],
+          "copy_index": 2
+        }
+      ],
+      "prefix_silence_samples": 80000
+    }
+  ],
+  "ffmpeg_sha256": "b778ca5786ee4b0b4f37f34478617093d2fdb9a532d281fb76cda69d1d579d46",
+  "ffprobe_sha256": "4e1b0afd27bf02d79739f27b2d650137470925afb19d54b5e89a708b7a389236"
+}

fixtures/speech-validation-v1/labels.lock.json

@@ -0,0 +1,7 @@
+{
+  "fixture_sha256": "e3a850d7a2385e250408f57fe51f4620ce80a20f8625ced42205f02cc46d517d",
+  "frozen_before_inference": true,
+  "source_count": 2,
+  "derived_clip_count": 4,
+  "query_count": 10
+}

fixtures/speech-validation-v1/seam-plan.json

@@ -0,0 +1,23 @@
+{
+  "schema": 1,
+  "frozen_before_inference": true,
+  "purpose": "Artificial repeat/offset primitive stress, separate from4clipretrievalpool; no sentence-precise labels",
+  "policy": "ReferencefullsourceVAD versus independent30s VAD windows with2s overlap; no parityassumption or postscore parameter tuning",
+  "recipes": [
+    {
+      "id": "seam-en",
+      "source": "speech-en",
+      "copies": 2,
+      "prefix_silence_ms": 5000,
+      "maximum_duration_ms": 45000
+    },
+    {
+      "id": "seam-de",
+      "source": "speech-de",
+      "copies": 3,
+      "prefix_silence_ms": 5000,
+      "maximum_duration_ms": 45000
+    }
+  ],
+  "acceptance": "Report sourcecoverage/intervalmembership disagreement, boundary omissions/duplicates and cancellation/failure before publication; map padding exactlyonce, preserve known source offsets and no label-selected speech interval. Any whole-file probability difference is a finding, not automaticallya correctnesserror; functional acceptance requires no missing sourcewindows, validbounds and independently assessed speechrecall. Human finephoneme labels unavailable; do not claim precise localizedquality."
+}

fixtures/speech-validation-v1/seams.lock.json

@@ -0,0 +1,4 @@
+{
+  "seam_plan_sha256": "5da4400814c4965fc12a87bb00c36665300730a702a69acd65a72b9f5201ac9e",
+  "frozen_before_inference": true
+}

tools/speech-eval/prepare_validation.py

@@ -0,0 +1,105 @@
+"""Acquire frozen small human speech inputs and prepare controls without inference."""
+import argparse
+import hashlib
+import json
+from pathlib import Path
+import subprocess
+import urllib.request
+import wave
+
+
+def sha(data):
+    return hashlib.sha256(data).hexdigest()
+
+
+def execute(args):
+    return subprocess.run(args, capture_output=True, check=True, timeout=60).stdout
+
+
+def write_wav(path, samples, prefix_ms=0, maximum_ms=30000):
+    if not 0 < maximum_ms <= 45000 or len(samples) % 2 or not samples or len(samples) // 2 + prefix_ms * 16 > maximum_ms * 16:
+        raise ValueError("Invalid or oversized 30-second source")
+    if path.exists():
+        raise ValueError("Preserve existing prepared clip")
+    with wave.open(str(path), "wb") as output:
+        output.setnchannels(1)
+        output.setsampwidth(2)
+        output.setframerate(16000)
+        output.writeframes(bytes(prefix_ms * 32) + samples)
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("fixture", type=Path)
+    parser.add_argument("output", type=Path)
+    parser.add_argument("ffmpeg", type=Path)
+    parser.add_argument("ffprobe", type=Path)
+    args = parser.parse_args()
+    raw = args.fixture.read_bytes()
+    fixture = json.loads(raw)
+    lock = json.loads(args.fixture.with_name("labels.lock.json").read_text())
+    if sha(raw) != lock["fixture_sha256"]:
+        raise ValueError("Frozen plan changed")
+    args.output.mkdir(parents=True, exist_ok=True)
+    plan_path = args.output / "prepared.json"
+    if plan_path.exists():
+        raise ValueError("Preserve existing preparation")
+    prepared = []
+    for asset in fixture["assets"]:
+        path = args.output / (asset["id"] + "-source.ogg")
+        if not path.exists():
+            request = urllib.request.Request(asset["source_url"] + "?download=1", headers={"User-Agent": "SearchiMediaEvaluation/1.0"})
+            with urllib.request.urlopen(request, timeout=60) as response:
+                data = response.read(asset["bytes"] + 1)
+            if len(data) != asset["bytes"] or hashlib.sha1(data).hexdigest() != asset["sha1"]:
+                raise ValueError("Publisher source changed")
+            path.write_bytes(data)
+        original = path.read_bytes()
+        if len(original) != asset["bytes"] or hashlib.sha1(original).hexdigest() != asset["sha1"]:
+            raise ValueError("Cached source changed")
+        source_probe = json.loads(execute([str(args.ffprobe), "-v", "error", "-show_format", "-show_streams", "-of", "json", str(path)]))
+        audio = next(s for s in source_probe["streams"] if s["codec_type"] == "audio")
+        if "duration" not in audio or float(audio["duration"]) + asset["prefix_silence_ms"] / 1000 > 30:
+            raise ValueError("Full primary speech stream must fit 30 seconds; never silently crop reference speech")
+        samples = execute([str(args.ffmpeg), "-v", "error", "-nostdin", "-threads", "1", "-i", str(path), "-ac", "1", "-ar", "16000", "-f", "s16le", "pipe:1"])
+        output = args.output / (asset["id"] + ".wav")
+        write_wav(output, samples, asset["prefix_silence_ms"])
+        prepared.append({"id": asset["id"], "path": str(output.resolve()), "sha256": sha(output.read_bytes()), "samples": len(samples) // 2 + asset["prefix_silence_ms"] * 16, "source_sha256": sha(original), "source_probe": source_probe, "source_audio_start_ms": float(audio.get("start_time", "0")) * 1000, "prefix_silence_ms": asset["prefix_silence_ms"], "fine_utterance_timestamps_available": False})
+    music_source = Path(".searchi/qa/music-validation-v1-canonical/unseen03-source.ogg")
+    expected_music = fixture["controls"][0]
+    if sha(music_source.read_bytes()) != expected_music["source_sha256"]:
+        raise ValueError("Frozen instrumental control changed")
+    samples = execute([str(args.ffmpeg), "-v", "error", "-nostdin", "-threads", "1", "-i", str(music_source), "-t", "30", "-ac", "1", "-ar", "16000", "-af", "aresample=16000,atrim=end_sample=480000,asetpts=PTS-STARTPTS", "-f", "s16le", "pipe:1"])
+    if len(samples) != 960000:
+        raise ValueError("Instrumental control must cover full 30 seconds")
+    for id, data in [("music-control", samples), ("silence-control", bytes(960000))]:
+        output = args.output / (id + ".wav")
+        write_wav(output, data)
+        prepared.append({"id": id, "path": str(output.resolve()), "sha256": sha(output.read_bytes()), "samples": 480000, "source_audio_start_ms": 0, "expected_recognized_segments": 0})
+    for item in prepared:
+        item["actual_probe"] = json.loads(execute([str(args.ffprobe), "-v", "error", "-show_format", "-show_streams", "-of", "json", item["path"]]))
+    seam_raw = args.fixture.with_name("seam-plan.json").read_bytes()
+    seam_plan = json.loads(seam_raw)
+    seam_lock = json.loads(args.fixture.with_name("seams.lock.json").read_text())
+    if sha(seam_raw) != seam_lock["seam_plan_sha256"]:
+        raise ValueError("Frozen seam plan changed")
+    seams = []
+    for recipe in seam_plan["recipes"]:
+        source = next(item for item in prepared if item["id"] == recipe["source"])
+        with wave.open(source["path"], "rb") as clip:
+            clip.readframes(source["prefix_silence_ms"] * 16)
+            samples = clip.readframes(clip.getnframes())
+        path = args.output / (recipe["id"] + ".wav")
+        write_wav(path, samples * recipe["copies"], recipe["prefix_silence_ms"], recipe["maximum_duration_ms"])
+        count = len(samples) // 2
+        prefix = recipe["prefix_silence_ms"] * 16
+        mapping = [{"derived_samples": [prefix + index * count, prefix + (index + 1) * count], "source": recipe["source"], "source_samples": [0, count], "copy_index": index} for index in range(recipe["copies"])]
+        seams.append({"id": recipe["id"], "path": str(path.resolve()), "sha256": sha(path.read_bytes()), "source_sha256": source["source_sha256"], "samples": prefix + count * recipe["copies"], "piecewise_map": mapping, "prefix_silence_samples": prefix, "artificial_repeat_control": True})
+    result = {"fixture_sha256": sha(raw), "seam_plan_sha256": sha(seam_raw), "frozen_before_inference": True, "model_inference_performed": False, "ffmpeg_sha256": sha(args.ffmpeg.read_bytes()), "ffprobe_sha256": sha(args.ffprobe.read_bytes()), "clips": prepared, "seam_controls": seams}
+    with plan_path.open("x", encoding="utf-8") as stream:
+        json.dump(result, stream, ensure_ascii=False, indent=2)
+    print("Prepared four clips without inference; plan SHA256", sha(plan_path.read_bytes()))
+
+
+if __name__ == "__main__":
+    main()

tools/speech-eval/test_prepare_validation.py

@@ -0,0 +1,45 @@
+import tempfile
+import unittest
+import wave
+from pathlib import Path
+
+from prepare_validation import write_wav
+
+
+class PreparationTests(unittest.TestCase):
+    def test_known_padding_does_not_change_source_samples(self):
+        with tempfile.TemporaryDirectory() as directory:
+            path = Path(directory) / "clip.wav"
+            samples = b"\x01\x00" * 16000
+            write_wav(path, samples, 5000)
+            with wave.open(str(path), "rb") as clip:
+                self.assertEqual(clip.getnframes(), 96000)
+                self.assertEqual(clip.readframes(80000), bytes(160000))
+                self.assertEqual(clip.readframes(16000), samples)
+
+    def test_full_source_over_budget_is_rejected(self):
+        with tempfile.TemporaryDirectory() as directory:
+            with self.assertRaises(ValueError):
+                write_wav(Path(directory) / "clip.wav", bytes(960000), 5000)
+
+    def test_existing_preparation_cannot_be_overwritten(self):
+        with tempfile.TemporaryDirectory() as directory:
+            path = Path(directory) / "clip.wav"
+            path.write_bytes(b"original")
+            with self.assertRaises(ValueError):
+                write_wav(path, bytes(32000))
+            self.assertEqual(path.read_bytes(), b"original")
+
+    def test_odd_pcm_is_rejected(self):
+        with tempfile.TemporaryDirectory() as directory:
+            with self.assertRaises(ValueError):
+                write_wav(Path(directory) / "clip.wav", b"\x00")
+
+    def test_explicit_seam_budget_is_bounded(self):
+        with tempfile.TemporaryDirectory() as directory:
+            path = Path(directory) / "seam.wav"
+            write_wav(path, bytes(32000 * 35), maximum_ms=45000)
+            with wave.open(str(path), "rb") as clip:
+                self.assertEqual(clip.getnframes(), 16000 * 35)
+            with self.assertRaises(ValueError):
+                write_wav(Path(directory) / "oversized.wav", bytes(32000), maximum_ms=45001)