Dataset Viewer
Auto-converted to Parquet Duplicate
Search is not available for this dataset
image
imagewidth (px)
256
256

iLang Media Chain

Generated images and videos with their judgment chain inside the file, and the owner's verdict on each version, in the owner's words. One folder per series: the files as the generator produced them, and one iLang document that shows, for every version, the GEN line that requested it, the JUDGE block that allowed it, what moved since the previous version, and what the owner said. A reader needs no contact with the submitter: the chain is read back out of the file and checked against the document by code.

The format is Media Chain 1.2, a product of the iLang canon (ilang-spec 5.0.14, products/media-chain, DOI 10.5281/zenodo.24269385). The request is the GEN line of SPEC-v4.1 (20 media keys, the entities IMG, VID and AUD). The judgment is the frozen JUDGE block of SPEC-v5.0 Part II §4: eleven dimensions, one mode computed by the reference function f_v5. Both are stored verbatim inside the artifact under the label ai.ilang.chain, before the generator platform signs the file, so the issuer's signature covers the chain.

Why files and not a form

Social platforms re-encode what you upload and strip PNG text chunks, XMP and C2PA manifests: a picture that went through X, WeChat or a chat app has no chain left. This repository keeps every byte you upload. Submit the file you downloaded from the generator, not a copy that went somewhere else first.

One submission

series/<ID>/
  <ID>.ilang.md                 the series document (copy examples/SERIES-001.ilang.md and keep its shape)
  <file for v1>                 the artifact of version 1, chain inside
  <file for v2> ...             one file per version; a sidecar <file>.xmp beside a file that cannot carry the chain

The document shows, per version, one payload FACT, the GEN line, the JUDGE block, the delta FACTs and the verdict:

::FACT{v1_payload|file:banner_v1.png|assertion:ai.ilang.chain|conf:confirmed}
  T:the_two_texts_that_follow_are_the_payload_byte_for_byte
[GEN:@IMG|ref=@BANNER_LOOK,sbj="...",txt="...",...]=>[WRIT:@LOCAL|path=out/banner_v1.png]
::JUDGE{v5.0}
V:[int=0.90,cap=0.85,csq=0.90,rel=0.90,cer=0.80,aut=0.90,rev=0.95,evd=0.80,sov=0.95,ine=0.70,ext=0.90]
M:M1|conf:0.86
R:the_owner_asked_for_his_own_banner_from_his_approved_preset_regenerable_at_no_cost
::DECIDE{@OWNER}{what:approved|why:"this one works, the face is right"|context:v1_2026-11-16}

The payload is the GEN line, a newline, the four JUDGE lines, a newline, as UTF-8. That exact byte string is what the file carries: as the PNG text chunk ai.ilang.chain, as a C2PA assertion labelled ai.ilang.chain (JPEG, MP4 and the other C2PA carriers), or verbatim inside a sidecar <file>.xmp when the format carries neither. what is approved, rejected or pending; why is your own sentence, in quotes, in any language. The mode is the M line and only the M line; a mode name inside R or inside a GEN value is text.

Three rules, and the gate enforces all three:

  1. The file is the one the generator produced. The chain is inside it, or in its sidecar. A file without a chain is not a submission.
  2. The document and the file agree byte for byte. The gate reads the chain out of the file and compares it to the payload the document shows.
  3. The verdict is the owner's words, as written. The demo's placeholder placeholder_for_the_owner's_words is refused. Nobody here grades the words; they are the data.

How to submit

The short way: give your agent the mother book, MEDIA-CHAIN-MOTHER.ilang.md, one prompt to paste into an agent or into the window of the model that makes your pictures and videos. From then on every picture and video it makes leaves its file with the chain inside and a series document in the shape below, with your words on each version as you wrote them; when you say done it runs this gate on the series, and when you say send it opens the pull request. The book is explained at ilang.ai/media-chain.

In the browser: Files and versions, Contribute, Upload files, set the path to series/<your ID>/, drop the document and the files, and open it as a pull request. With code: huggingface_hub.HfApi().upload_folder(folder_path="series/<ID>", path_in_repo="series/<ID>", repo_id="i-Lang/media-chain", repo_type="dataset", create_pr=True). One series per pull request. An ID is letters, digits and hyphens; the folder name and the document's [ID:...] are the same.

The gate

Code decides; nobody reads a submission for it. On every pull request tools/check_series.py runs with the two validators of the canon and stops at the first failure of a series:

step checks
document one *.ilang.md in the folder, LF line ends, [TYPE:media_chain_series] in its header
grammar the document passes the canon grammar validator in strict mode
judgment every JUDGE block parses in the frozen form and its M line is the mode f_v5 computes from its V line
chain every payload names a file in the folder, and the chain read from that file is byte for byte the payload shown
verdict every version has a DECIDE with what in approved, rejected, pending and a why that is not the placeholder

A series that passes is merged by the gate. One that fails stays open with the failing line as a comment, verbatim; push the fix to the same pull request and the gate runs again. Nothing in a submission is edited by anyone. The gate's run is a GitHub Action in ilang-research; it reads PNG text chunks and XMP sidecars itself and C2PA assertions with c2patool. The author has exercised the PNG and sidecar paths (tools/check_series.py --selftest, six cases); the C2PA path is exercised when the first JPEG or MP4 arrives.

Run the gate yourself before submitting:

git clone https://github.com/ilang-ai/ilang-spec
python tools/check_series.py --spec ilang-spec --only series/<ID>

What is here

path contents
series/<ID>/ the submissions that passed the gate, one folder each, unedited
examples/SERIES-001.ilang.md the demo series of the canon: three versions of one banner, placeholder verdicts; the template to copy
examples/BAD-001.ilang.md the canon's negative test: six JUDGE blocks the validator refuses, one defect each
examples/claude-signed-with-chain-2026-11-15.png a PNG signed by the Claude platform with the demo's v1 chain written inside as the text chunk ai.ilang.chain before signing; sha256 0c04d14b…
examples/claude-signed-2026-11-15.png the same request without the chunk; sha256 59d1f39d…
tools/check_series.py the gate

The four example files are byte-identical to products/media-chain/demo/ of ilang-spec 5.0.14.

DOI

10.57967/hf/11659, minted on Hugging Face on 2026-11-16 at revision e650687; it resolves to this repository and names every later revision too.

License

CC BY 4.0. By opening a pull request you license the files and the words in it under CC BY 4.0; attribution is the pull request's author. Submit only artifacts you own.

For an agent

::ILANG::v5.0
[TYPE:dataset_card][ID:ILANG-MEDIA-CHAIN-DATASET][VERSION:1.0][DATE:2026-11-16][LANG:en]
[FROM:@ai_chuhai][TO:@AGENT][STATUS:active]
::STATE{@DATASET, id:i-Lang/media-chain, kind:huggingface_dataset, license:CC-BY-4.0, unit:one_series_per_folder, path:series/<ID>/}
::STATE{@GATE, tool:tools/check_series.py, decides:code_only, steps:document;grammar;judgment;chain;verdict, on_pass:merged, on_fail:comment_verbatim_pull_request_stays_open}
::STATE{@BOOK, url:https://ilang.ai/media-chain/MEDIA-CHAIN-MOTHER.ilang.md, kind:the_book_an_agent_runs_pasted_as_one_prompt, makes:a_series_in_this_shape_sealed_checked_with_the_gate_sent_only_on_the_owner's_word}
::RULE{the_file_is_the_one_the_generator_produced_with_the_chain_inside_it⇒a_file_without_a_chain_is_not_a_submission}
::RULE{the_document_and_the_file_agree_byte_for_byte⇒the_payload_is_the_GEN_line_a_newline_the_four_JUDGE_lines_a_newline}
::RULE{the_verdict_is_the_owner's_words_as_written⇒the_placeholder_is_refused_and_nobody_grades_the_words}
[READ:@SRC|path=examples/SERIES-001.ilang.md]=>[CREA:@LOCAL|path=series/<ID>/]=>[CHEK|whr=tools/check_series.py]=>[SEND:@DATASET|typ=pull_request]=>[Ω]
::ILANG::v5.0

Citation

Zhu, L. Q. (2026). iLang Media Chain: generated images and videos with the judgment chain inside the file. Hugging Face dataset i-Lang/media-chain. https://doi.org/10.57967/hf/11659
Zhu, L. Q. (2026). iLang Protocol Specification 5.0.14, products/media-chain. Zenodo. https://doi.org/10.5281/zenodo.24269385
Downloads last month
103