Learned audio signature

experiments/military_signature/ANNOTATION_ROUNDTRIP.md

@@ -0,0 +1,39 @@
+# P0 annotation import and export verification
+
+The supplied `annotation.json` passed the frozen P0 import/export contract on
+2026-10-11. All six artificial recordings have an interval or logged rejection.
+The seven intervals comprise one foreground, one background and five rejections.
+
+The input matches `/Users/magi/Downloads/annotations.json` semantically and binds
+to the final synthetic review batch. Its original bytes are retained outside the
+checkout; the supplied file is unchanged. Batch/config/tool-version bindings,
+private mapping commitments, review-audio hashes, connected groups and roles all
+passed validation.
+
+Two mono clips from `fixture-b` were exported at the native 16,000 Hz rate:
+
+| Kind | Interval | Native frames | Samples |
+|---|---|---|---|
+| Foreground | 2–6 seconds | 32000–96000, end exclusive | 64000 |
+| Background | 0–2 seconds | 0–32000, end exclusive | 32000 |
+
+Each exported waveform matches its raw fixture slice sample for sample. All five
+rejection records remain in the manifest and export record. The original 55-file
+P0 freeze still verifies; its historical report and protocol were not rewritten.
+
+The detailed evidence, copied input, manifest, reuse audit and native clips are at
+`artifact:military_signature/p0/browser-roundtrip-20261011T232425Z/`. The main
+record is `verification.json`; clip records are in `native_clips/export.json`.
+
+The user subsequently confirmed the Chrome run: “Yes, playback and display
+worked,” in response to a question covering audio playback, waveform,
+spectrogram and interval-control display. This completes the requested human
+browser check. The exact question, answer and input hash are retained in
+`browser_confirmation.json` beside `verification.json`.
+
+The annotation artifact verifies the synthetic import/export workflow; the
+separate user confirmation supplies run-level playback/display evidence.
+Browser version, observation time and per-recording playback traces remain
+unknown. Vehicle labels remain unknown; these artificial signals provide no
+model-recognition evidence. Real-media processing, fitting and P1 still require
+separate authorization under [PROTOCOL_DRAFT.md](PROTOCOL_DRAFT.md).

experiments/military_signature/CURRENT_STATUS.md

@@ -0,0 +1,103 @@
+# Military acoustic signature current status
+
+Updated 2026-10-12 local date. This additive status supersedes the browser-handoff
+and powerplant-evidence gaps in the dated [P0_STATUS.md](P0_STATUS.md) where the
+new evidence below applies. Frozen v1 artifacts remain unchanged.
+
+All named P0 preparation deliverables have evidence. The user subsequently
+authorized P1 collection and preparation with a 5 GB total storage ceiling and
+20-minute per-asset limit. Scientific feasibility and model recognition remain
+untested; fitting and representation extraction are outside the current scope.
+
+## P1 collection and admission
+
+The [current authorization](P1_AUTHORIZATION.md),
+[collection contract](p1_admission_5gb.json) and
+[P1 report](../../reports/military_signature_P1_admission.md) supersede the earlier
+metadata-only handoff for collection. At the current verified snapshot, 54
+recordings total 8,735.906 seconds and occupy 4,879,449,428 bytes across the new
+data and artifact roots. The longest recording is 505.706 seconds. The 64-asset
+collection target was not reached under the storage/reserve limit.
+
+Seven caption-hidden packets are available at
+[the local review index](http://127.0.0.1:8766/). Real-media human review is pending;
+the earlier user confirmation concerns only the P0 synthetic packet. No real
+foreground segment has been admitted, no independent recording group has been
+verified and all experimental roles remain unassigned. Caption asset counts and
+the 38 named originators are not independent-session counts.
+
+The full 54-recording waveform reuse audit completed at 2026-10-12 07:10:02 UTC.
+Its authoritative output is
+`artifact:military_signature/p1_admission/20261011_5gb/reuse_audit.json`.
+It accounts for 53 fingerprinted assets, 4,175 indexed windows and one entirely
+zero stream with unresolved fingerprint status. One four-second suggestion links
+Bradley assets 897659 and 897660 with correlation 0.988650272254228, agreeing with
+their metadata dependence lead. The frozen original library reproduced that exact
+score. Neither this link nor absent waveform matches establishes independence.
+Historical/protected audio remains excluded.
+
+The earlier `provisional_admission_groups.json` snapshot remains preserved.
+`admission_groups.json` incorporates the completed waveform audit, metadata links
+and zero-stream flag. It binds the 54 unchanged collection records and joins their
+unknown sessions in one quarantine component under the frozen protocol. It asserts
+no known common shoot or vehicle and is not an experimental admission freeze.
+Closing checks verified all source/native hash pairs, all seven review-packet
+commitments and 11 protected repository witnesses. Verified independent groups
+and admitted foreground segments remain zero.
+
+Native-channel numerical integrity was subsequently checked for all 54 streams.
+One decodes entirely to zero in every channel; the other 53 contain nonzero
+samples. No entire mono-display cancellation of nonzero native audio was found.
+The external `native_signal_integrity.json` retains per-asset hashes and counts.
+These observations do not establish vehicle audibility or replace human review.
+
+The latest focused verification passed 50 checks. The CAST test-count canary
+remains red at 720 collected versus 467 expected; baseline-file and frozen-file
+checks passed. Human-only baseline refresh and line-by-line review of the new
+test file remain required before merge. The original 55-file P0 freeze is intact.
+
+## P0 preparation evidence
+
+| Requested deliverable | Evidence | Status and limit |
+|---|---|---|
+| Frozen draft protocol; phase gates; eight-versus-ten distinction | [PROTOCOL_DRAFT.md](PROTOCOL_DRAFT.md), [canonical JSON](p0_protocol.json), [freeze ledger](p0_freeze.json) | Verified preparation. Eight groups/model is development feasibility; ten is a proposal minimum, not a power calculation. |
+| Manifest schema and recording/reuse/known-individual grouping | [manifest_schema.json](manifest_schema.json), [implementation](../../src/abvid/signature/manifest.py) | Verified with artificial cases. Public asset IDs supply no verified session count. |
+| Offline trimming and deduplication tooling | [CLI](../../src/abvid/signature/__main__.py), [round trip](ANNOTATION_ROUNDTRIP.md) | Synthetic import/export passed; user confirmed Chrome playback and display. Real-media behavior remains untested. |
+| Synthetic validation evidence | `artifact:military_signature/p0/validation-20261011-final/validation.json`; round-trip `verification.json` | Native sample equality, rejected intervals and provenance preserved; 37 focused regression checks passed at the recorded run. |
+| Literature, feasibility, confounds and negative findings | [frozen report](../../reports/military_signature_P0.md), [follow-up](../../reports/military_signature_P0_followup.md), [source supplement](source_evidence_supplement.json) | PDF-authoritative selected claims; unsupported variant/engine labels remain unknown. |
+| Public metadata inspection | [candidate_metadata.json](candidate_metadata.json) | Seven leads spanning four model captions; zero verified independent recording groups. Rights and audio suitability are unresolved. |
+| Harness integrity and handoff | Original report and `artifact:military_signature/p0/verification-20261011.json` | Recorded count canary failed: 707 versus 467; baseline-file and frozen-file checks passed. Human baseline refresh and line-by-line test review remain required before merge. |
+
+The user confirmation receipt is
+`artifact:military_signature/p0/browser-roundtrip-20261011T232425Z/browser_confirmation.json`.
+It preserves the exact question and answer linked to the supplied annotation
+hash. The original annotation is unchanged. The follow-up literature contains
+six additional PDFs and eight visually checked pages; it records historical or
+variant-specific Bradley, M109A6, Stryker and M1165A1 evidence without assigning
+engines to recordings.
+
+The supplements are outside the original 55-file freeze. Their separate hashes
+and external-evidence checks are recorded in
+`artifact:military_signature/p0/followup-verification-20261011.json`.
+The original freeze can still be checked with:
+
+```sh
+PYTHONPATH=src .venv/bin/python experiments/military_signature/verify_freeze.py
+```
+
+## Next input needed for scientific execution
+
+Current collection and preparation are authorized. The next input is real-media
+blind interval review and supported shoot/physical-vehicle provenance. Reuse and
+provenance admission must be frozen before allocating experimental roles.
+
+Any subsequent fitting/evaluation decision must separately specify the admitted
+population, identity granularity, exact checkpoint and extraction configuration,
+fit/compute budget and evaluation roles. Representation extraction, model or CAST
+fitting, protected-source evaluation and media publication remain outside the
+current scope. Confirmatory model-wide claims additionally need new unexposed
+groups, physical-vehicle support and an allocation/power plan. The eight-group
+pilot and ten-group model-claim minimum remain distinct and unachieved.
+The retired benchmark is not resumed by this work.
+
+No commit or merge was performed. The human-only harness baseline was not changed.

experiments/military_signature/P0_STATUS.md

@@ -0,0 +1,41 @@
+# Military acoustic signature P0 status
+
+P0 is separately authorized for literature, public asset metadata, protocol design,
+and implementation using synthetic fixtures. Real media collection or processing,
+model fitting, protected-source evaluation, and P1 execution remain unauthorized.
+This side project does not resume the retired military benchmark.
+
+## Deliverables
+
+- [x] Frozen draft [protocol](PROTOCOL_DRAFT.md) and explicit prospective phase gates
+- [x] [Manifest schema](manifest_schema.json) with conservative session/reuse/known-individual grouping
+- [x] Offline caption-hidden trimming and annotation tooling
+- [x] Fingerprint and correlation reuse audit
+- [x] Synthetic validation and CLI verification: 37 focused tests passed
+- [x] PDF-authoritative [source ledger](source_evidence.json), [metadata leads](candidate_metadata.json), and [research report](../../reports/military_signature_P0.md)
+- [x] Integrity and human handoff record; existing canary failure preserved
+
+These checkboxes indicate preparation deliverables, not real-media admission or
+scientific validation. Native browser playback/layout still requires human review:
+Firefox computer-use approval was rejected; DOM event checks used mocked media APIs.
+The current CAST canary collects 707 against 467 and fails only the count check;
+baseline-file/frozen-file checks pass. The baseline is untouched and human-only.
+
+Final artificial evidence:
+`artifact:military_signature/p0/validation-20261011-final/validation.json`.
+Freeze check: `PYTHONPATH=src .venv/bin/python experiments/military_signature/verify_freeze.py`.
+P1 requires separate authorization, per-asset rights and verified recording groups.
+
+## Initial evidence
+
+Before P0 implementation, the CAST harness canary collected 661 tests against
+467 in the existing baseline. The baseline-file and frozen-hash checks passed;
+the count check failed. Existing tests, lockfiles, configs, and baseline remain
+untouched. Additional unrelated work is present in this shared checkout.
+
+Eight sessions per model is a feasibility target, not a model-recognition claim
+gate. The proposal requires at least ten independently verified groups per model
+for such a claim, plus the hierarchy, null, and nuisance checks. A DVIDS-only
+pilot cannot establish cross-outlet invariance. Producer and camera differences
+within DVIDS must still be tracked. Public metadata is not audio-quality or
+independence evidence.

experiments/military_signature/P1_ADMISSION_DRAFT.md

@@ -0,0 +1,88 @@
+# Bounded P1 media-admission draft
+
+2026-10-11 local date; **draft for a separate user decision, not authorization**.
+P0 authorization explicitly excludes real-media collection and processing.
+The [P0 protocol](PROTOCOL_DRAFT.md) and original 55-file freeze remain unchanged.
+This stage prepares a reviewable admitted/rejected media population before a
+separate Tier 0 pilot decision. It does not execute the six-model × eight-group
+pilot or establish any model-recognition result.
+
+## Proposed action and bounds
+
+If explicitly authorized, implement a separately versioned admission path and
+inspect rights/provenance for the ten named DVIDS assets below. Collect and
+process an asset only after its per-asset restrictions and intended research
+use pass a recorded admission decision. Unclear or conflicting rights remain
+quarantined and require a specific resolution; a public-domain banner alone
+does not clear them.
+
+| Asset ID | Public metadata URL |
+|---|---|
+| 1754 | [Abrams b-roll](https://www.dvidshub.net/video/1754/abrams-tank-b-roll) |
+| 572786 | [Abrams package](https://www.dvidshub.net/video/572786/taa-abrams-set-up-b-roll-package) |
+| 855729 | [Finnish Leopard](https://www.dvidshub.net/video/855729/us-soldiers-explore-finnish-leopard-battle-tank) |
+| 776583 | [Bradley b-roll](https://www.dvidshub.net/video/776583/b-roll-stallions-conduct-bradley-live-fire) |
+| 386317 | [Stryker convoy](https://www.dvidshub.net/video/386317/m1126-stryker-combat-vehicle-convoy-ramstein-air-base) |
+| 937004 | [Pinon Canyon package](https://www.dvidshub.net/video/937004/pinon-canyon-maneuver-site-b-roll-package) |
+| 329602 | [M1151 training](https://www.dvidshub.net/video/329602/humvee-training-b-roll) |
+| 607339 | [M109A6 firing](https://www.dvidshub.net/video/607339/arkansas-field-artillery-unit-fires-m109a6-paladins-during-operation-western-strike) |
+| 967175 | [Army birthday parade](https://www.dvidshub.net/video/967175/army-250-parade-celebration-b-roll) |
+| 966208 | [Staged vehicles](https://www.dvidshub.net/video/966208/us-army-vehicles-staged-west-potomac-park-waterfront-b-roll) |
+
+Bounds are ten asset IDs; one selected media rendition per admitted asset; at
+most 3 GB total downloaded media; at most 20 minutes per asset and 45 minutes
+total decoded source duration. The current page durations sum to 42 minutes
+11 seconds, but actual media must satisfy the bounds. Failed, unavailable or
+rejected assets are retained as outcomes. No unlisted substitute is collected.
+The restricted NATO Leopard asset 879491 is excluded from this draft.
+
+Collection, original-media hashes, native audio extraction, blind review packets,
+label-blind reuse checks, provenance grouping and reviewed interval import/export
+are the proposed processing scope. Representation extraction, classifier fits,
+CAST fits, contrastive learning, scientific retrieval/probe metrics and protected
+source access remain outside this stage. All new media/results remain external,
+resolved through `abvid.paths`; no existing raw data, protected roles, historical
+manifests or frozen P0 files are changed. Human-only harness baseline changes
+and merge approval are not part of this authorization.
+
+## Required admission sequence
+
+1. Preserve the source page, rights statement, date/unit/location conflicts,
+   producer attribution, rendition metadata and retrieval failures. Record the
+   intended use as local acoustic research without publication or
+   redistribution of the media. Resolve restrictions before accessing media.
+2. Enforce byte/duration bounds and hash the chosen original rendition. Retain
+   its native audio rate/channels and extraction provenance. No quality label,
+   engine identity or session independence is inferred from captions.
+3. Use a new admission configuration/tool version. The current P0 configuration
+   permits synthetic fixtures only and must continue to reject real media.
+   Verify the new processing mechanics before using admitted files.
+4. Build randomized caption-hidden audio/waveform/spectrogram review packets.
+   A human reviews audible vehicles, foreground/background intervals,
+   contaminants, motion and rejection reasons. Caption shot-log intervals are
+   not used to preselect or accept audio. If no reviewer is available, stop at
+   the packet handoff and report that admission is incomplete.
+5. Audit reuse and retain all known or suspected shoot/population links. Session
+   and individual IDs remain `unknown` unless evidence establishes them.
+   Assets with unresolved independence cannot fill a verified-group quota or
+   be divided across experimental roles. No train/development/test allocation
+   is performed in this admission stage.
+6. Import human annotations, retain every rejection and export accepted clips
+   with native-frame provenance. Attach caption labels separately after blind
+   trimming, with confidence and evidence. Mixed-model event material remains
+   diagnostic unless a new protocol establishes valid model/interval identity.
+
+## Outputs and stopping rule
+
+Deliver a rights/provenance admission ledger, native-media/extraction hashes,
+caption-hidden packets, imported human annotations where available, conservative
+group/reuse audit, rejected/unavailable asset list and an admission report. The
+report must state admitted foreground/background support, verified independent
+groups, known physical-vehicle support and remaining unknowns, including zero
+support where appropriate. Store them under a fresh external P1-admission run.
+
+Stop after these ten assets, at a budget bound, or at a required rights/reviewer
+handoff. This draft does not permit collecting until a positive gate appears.
+The six-class identity definition, eight-group pilot coverage, exact frozen
+representation/checkpoint, fit budgets and Tier 0 evaluation allocation must
+be resolved in a separate pilot protocol before fitting or metric evaluation.

experiments/military_signature/P1_AUTHORIZATION.md

@@ -0,0 +1,27 @@
+# P1 admission authorization and current limits
+
+The user authorized more real-media collection after reviewing the bounded
+P1 admission request: “Get more data,” retaining the per-asset time limit.
+The initial 20 GB allocation was then replaced by “lets start with 5 GB.”
+The [current contract](p1_admission_5gb.json) therefore uses **5,000,000,000 bytes
+total across the new data and artifact roots**, with **1,200 seconds per asset**.
+It expands the original ten-asset discovery list; the old 45-minute aggregate
+draft limit no longer defines this expanded batch. Storage is a ceiling, not a
+requirement to fill the disk. The first collection target is 64 playable assets
+across the six candidate caption classes, subject to the budget and availability.
+
+Collection and preparation are authorized. Rights/provenance review, original
+rendition retention, native audio preparation, caption-hidden human review and
+label-blind reuse auditing are within scope. Fitting, representation extraction,
+protected-source evaluation and media publication remain outside it.
+
+All original P0 files and their 55-file freeze remain unchanged. A separately
+versioned P1 implementation and external run are used; real files are never
+passed through P0's synthetic-only admission guards. Existing archival source
+catalogs are read as metadata solely to exclude known historical assets/filenames.
+Protected media are not opened or used for waveform comparisons.
+
+Metadata diversity across dates, units and sites helps candidate selection but
+does not establish independent sessions or physical vehicles. All collected
+material is pending blind review and conservative provenance admission. Captions
+are retained privately and do not preselect accepted audio intervals.