AI media can look finished and still be unproven.
Collaiber Ecosystem is an evidence and QC layer for AI-era post-production. Its evaluation engine, CTAME, records what AI did to an asset, measures where the result is at risk, and keeps a qualified human in the loop. It flags problems instead of rubber-stamping them.
Noisy generative signal in. A measured, deterministic and verifiable record out.
The industry measures quality. The AI process often disappears.
Generative tools are judged on how the output looks. Delivery pipelines are judged on whether the file plays. For AI-assisted media there is rarely a reliable record of what was generated, what was restored, what a human approved, and where the result still carries risk.
Output looks done
A generated clip can pass a glance and still hide oversharpening, crushed shadows, or temporal drift.
Process is invisible
Which models touched it, with what settings, in what order, and who signed off. Little of it travels with the file.
Risk is undocumented
There is no shared record of what was checked, what was flagged, and what still needs a human eye before delivery.
Collaiber closes that gap through CTAME, at the layer everyone skips: the evidence of how AI-era media was actually made.
It does not make AI media prettier.
It makes it provable.
Cleaner is not always safer.
Drag across this real generative frame to move from source to restored, and keep going into oversharpened. Watch the evidence panel. Past a point CTAME stops calling it ready. This is the judgment the engine makes on real footage.
source
drag →
The honest twist: the restored pass looks better, but pushed too far CTAME measures a five-fold sharpness jump and flags oversharpening. The engine reports the point where restoration starts to damage the asset, instead of mistaking clean-looking for correct.
The same call, on a real delivered clip.
A Kling-generated shot, evaluated as source and after a manual restoration pass. CTAME did not restore it. It registered both files, compared them, and reported what changed and what to check. These are the actual measured values.
Color-space was not explicit enough for trusted metric comparison, so those deltas were reported as untrusted and were not used to claim improvement. Collaiber states what the evidence supports and stops there.
QC categories
Technical fingerprint
The restored pass looked cleaner, but CTAME measured a five-fold sharpness rise and up to 21% of pixels crushed to black. It routed the asset to a human instead of marking it ready.
Upstream provenance does not survive restoration.
Embedded provenance is set by the generator. In our tests, validated with the C2PA reference tool, a single restoration step removed the embedded manifest, and it never returned through the rest of the chain. The restored frames stayed structurally close to the source (SSIM 0.93 to 0.96), so the loss comes from re-encoding, not from changing the picture.
Measured on embedded manifests only, with the C2PA reference tool. Hidden watermark and fingerprint survival is a separate, ongoing measurement and is not claimed here.
What comes out of the system today.
For each evaluated asset, the working engine generates a concrete set of artifacts. Not a vision, a deliverable package you can open, read, and check.
From one asset to a sealed project record.
The engine that proves a single asset is the first slice. The full ecosystem captures every action across a project, commits it to a verifiable archive, and seals that archive on delivery.
Evaluate & document
Each asset is registered, QC’d, restored where needed, and compared. CTAME produces a passport, a QC report, and a hashed evidence record.
Working todayCommit to a .clb archive
Every record, log, and hash composes into one project container with a single root hash, a commitment over the whole project, not just one file.
In developmentSeal at delivery
The final master carries a standard C2PA credential pointing to the archive root, so the delivered file stays interoperable while the full history stays verifiable underneath.
On the roadmapProject-level provenance, not file-level. A standard credential on the delivered file for regulatory interoperability, with a deeper, hashed audit record underneath. Surface standard, depth proprietary.
Built on published measurement.
Collaiber Determinism Index
The first published benchmark for how generative AI video survives Tier-1 broadcast QC. Independent reviewer validation, methodology and data in the open.
zenodo · doi 10.5281/zenodo.19394381A framework, not a verdict machine
Collaiber separates measured facts from delivery decisions. It reports what the evidence supports. The human, the contract, and the institution decide what it means.
orcid 0009-0002-2524-2068For teams that need AI media to be measurable, reviewable, and safe enough for professional workflows.
Scope: Collaiber Ecosystem is a technical validation aid for AI-era media production. It documents integrity and process and flags risk. It is not certification, not final approval, and not a delivery decision. Manual QC by a qualified operator is required.
Put a real clip through it.
Have an AI-generated or AI-restored asset that needs production validation? Collaiber Ecosystem is in active development from Babelsberg. Send a note and a real pilot clip.