Measured product lead custody receipt.
github.com/Zer0pa/ZPE-VideoDEVELOPER-READY
LIVE
RECEIPT CORE OPEN
See What AI Models Really See
A byte-identical perception receipt for AI video pipelines · ZPE-Video · PyPI zpe-video v0.1.0 · github.com/Zer0pa/ZPE-Video
AI systems decide what gets flagged in video, who gets identified, what a generation model is trained on — and until now, no one outside the pipeline could check what the detector was actually shown.
ZPE-Video closes that gap. Two independent Python writers, built from the same spec, produce a byte-identical perception receipt for the same frames. The receipt records detector, tracker, and manifest state; it does not reconstruct pixels or audio. It is a re-derivable record, not a video codec.
01 · THE GAPNO RECEIPT
An AI processes video — but no one can check what it was actually shown.
02 · MARKETSADJACENT FORECASTS
Adjacent AI-video and provenance markets · hypothesis only; no adoption, compliance, or legal-sufficiency claim.
03 · VALUE
$39.7B
Content provenance is growing; ZPE-Video is a bounded Python receipt layer beneath that infrastructure, not a replacement for it.
04 · INSIGHT
What AI saw in the video can now be checked.
05.1 · CURRENT TECHNO RECEIPT FORMAT
Perception traces from AI video pipelines scatter across Parquet, JSON, pickle, and MCAP containers. There is no portable record format and no second writer that can independently rebuild the same bytes from the spec.
05.2 · OUR TECHBYTE-IDENTICAL RECEIPT
zpe-video v0.1.0 ships a zero-dependency Python library with a documented wire format, per-frame CRC32, stable receipt hashes, and SHA-256 manifest binding. Two independent writers — one built from the spec by hand — produce byte-identical output across 3 receipt-core cases. The record is re-derivable, not just inspectable.
05.3 · BENCHMARKS3 RECEIPT-CORE CASES
Cross-writer SHA3cases stable
Manifest bind3cases verified
Receipt corePASSnot sovereign
PyPIv0.1.0connected
Benchmark: 961B vs Parquet 5,386B · 0.302ms vs 4.500ms · receipt scope only.
06 · MEASUREMENTRECEIPT VALIDATION
Receipt evidence lives in bytes, hash, CRC, and manifest.
06.1 · COMPARATIVE RECEIPT VALIDATIONCROSS-WRITER SHA STABILITY
Two writers produce the same bytes on 3 receipt-core cases · manifest binding holds · Parquet and raw-struct comparators shown for reference · wider corpus coverage open. Source: proofs/artifacts/
07 · KEY METRICSPHASE 09.4 PROVENANCE
07.1 · CROSS-WRITER SHA
3CASES STABLE
Two independent Python writers · byte-identical output
07.2 · MANIFEST BIND
3CASES VERIFIED
SHA-256 receipt → manifest binding verified
07.3 · RECEIPT CORE
PASS
Receipt scope only · not full-video replay
07.4 · RELEASE
v0.1.0
PyPI zpe-video · released 2026-05-04
07.5 · COMPRESSION
NO
No video-frame compression claim.
08 · DETERMINISMBYTE-EXACT RECEIPT
Two independent Python writers, one byte-identical receipt.
08.1 · WHAT DETERMINISTIC MEANSRECEIPT SCOPE
Deterministic means the same detector and tracker input plus the same wire-format spec produce a byte-identical perception receipt from two independent Python writers, with SHA-256 manifest binding verified on the same three receipt-core cases. Cross-writer SHA-stable measures the receipt only — it is not deterministic computer vision, deterministic LLM output, a legal evidence chain, full-video replay, or a competitor to AV1 or VVC. The scope is the record, and the scope is named.
08.2 · HONEST BLOCKER
Honest Blocker ·
The receipt carries detector and tracker state — boxes, track IDs, timestamps, CRCs, manifest binding. It does not reconstruct pixels, appearance, or audio. Cross-runtime replay is open, C2PA integration is not in scope yet, and the PyPI v0.1.0 README still carries stale private-repo and video-codec wording.
09
WHAT THE MODEL SAW, on the record.
09.1 · THE AMBITION
The ambition is not to compete with video codecs. It is to make the question "what was this AI actually shown?" answerable by anyone who can run Python. When that record exists at intake, AI video systems stop being black boxes that decide on inputs no one else can inspect.
09.2 · WHAT WORKS NOW
Working today: byte-identical perception records, two independent Python writers, manifest-bound on three receipt-core cases.
09.3 · WHAT'S STILL OPEN
Still open: cross-runtime replay, C2PA integration, sovereign closure, and the v0.1.0 PyPI README cleanup.
09.4 · MODERATION · NEAR-TERM (12–24 MO)
Moderation teams gain a re-derivable record
A trust-and-safety lead reviewing a content-removal appeal can ask the pipeline what frames the detector actually saw, and a second engineer can rebuild the writer and confirm the same bytes. Disputes stop hinging on the platform's word alone.
09.5 · STORAGE · NEAR-TERM (12–24 MO)
Detection logs shrink without losing the record
A surveillance archive owner storing months of detector output keeps the same boxes, tracks, and timestamps at roughly a fifth of the storage footprint of default Parquet. The infrastructure bill drops without changing what the auditor can re-read later.
09.6 · PROVENANCE · MID-TERM (24–48 MO)
Provenance standards get a perception layer
Content-authentication standards like C2PA describe what was made; this describes what was seen on the way in. As both layers connect, a published video can be traced back to the exact detector inputs the producing pipeline acknowledged, not just the rendered output.
09.7 · INTEGRITY · MID-TERM (24–48 MO)
Silent corruption stops reaching the model
A video-LLM inference pipeline that accepts a corrupted detector record today produces a quietly wrong answer. With per-frame CRC checked on decode, corruption surfaces as a raised error at the gate, so a downstream operator notices before the bad inference reaches a user.
09.8 · GOVERNANCE · PARADIGM (48 MO+)
AI video input becomes legally answerable
Regulators, insurers, and courts asking what an AI was shown today get a shrug. When pipelines carry byte-identical perception records at intake, the question becomes answerable on demand, and AI-video liability shifts from speculation about model intent to inspection of recorded input.