cohenta
Get the Runner
HomeHow it worksCapabilitiesAsset typesProduct surfacesPricingCreativeVisualVideoResearchDistributionIntegrationsQuality & TrustSolutionsSecurityThe engineAboutEnterpriseBlogNewsContactFAQSign inGet the Runner

Quality, Trust & Governance

Trust enforced in code, not prompts.

Brand safety isn't a hopeful instruction to a model. Deterministic gates every asset must clear — and a record of why.

Get the RunnerHow it works
Deterministic checksGrounded in your factsWCAG-checkedManifest per runZero custody
Brand safety

A gate, not a prompt

Banned terms, absolute claims, missing disclaimers — caught first.

A severity and a reason on every finding — nothing silently rewritten.

✓ pass
Source grounding

Every number traces to a source

Every statistic checked against the facts and KPIs you provided.

Unverified figures flagged in the trail, never silently kept.

⚑ flagged

unverified figure — flagged and annotated, held for human review

Quality scoring

Three judges that prove the best

Three independent reviewers score each asset against the same rubric.

Scores recorded per model and job — routing proven from results.

✓ pass
Legibility

Accessibility checked first, for free

Real WCAG contrast, blank renders, off-policy text — caught first.

Checked on the real render — body to 7:1, accents to 4.5:1.

✓ pass
Audit trail

Traceable and reproducible

Every run emits a manifest, so assets re-render pixel-identically.

A rejected run resumes from its checkpoint, not the whole campaign.

✓ pass
The governance ledgerSame spec · same pixels · every run

The gate, live

One draft, stamped line by line.

Every claim checked against your facts, flagged with a reason you can act on. The run holds until you decide.

Defense in depth

Every layer leaves a record.

Cheap kernelsdeterministic · runs firstCopy criticscored before shipJudge panelone rubric · parallelVision gateverdict + reasonsPublished & loggedmanifest writtenhuman approval
Five layers, one spineA human holds the last gate

Cheap kernels

Blur, blanks, low contrast, off-policy text — no model call needed.

Copy critic

Coverage, citations, CTA, and length — scored deterministically.

Judge panel

Three reviewers, one rubric. A bounded loop re-renders and re-judges.

Vision gate

Every visual meets an editorial bar. A publish verdict, with reasons.

Published & logged

Each finding logged to the trail. Every asset proves itself later.

Explore capabilities

Common questions

Asked and answered.

What does "brand safety enforced in code" actually mean?
Brand safety runs as deterministic gates between generation and publication — a prompt can be ignored, a gate cannot. Banned terms, unverified absolutes, and missing disclaimers are flagged with a severity; a human makes the final call.
How does the content governance prevent fabricated statistics?
Every statistic, percentage, and multi-digit number is checked against your facts; untraced figures are flagged in the audit trail, never silently kept. Blogs expect citations, social posts don't; localization preserves every figure.
How do you know the output is actually high quality?
Three independent models score each artifact against identical rubrics; below target, a bounded loop re-renders and re-judges. Copy and visuals also clear a deterministic critic and a vision gate — a publish verdict, with reasons.
Which model does Cohenta use, and how is it chosen?
Judge scores are recorded per model and task; design, copy, research, and deck work then routes to whichever has performed best — proven, not preset. The model inventory is queried live, so the router never calls a deprecated model.
Can I configure my own banned terms and disclaimers?
Yes — set voice, banned-term lists, and disclaimers once and they apply across every pipeline and modality, every hit flagged for review. An optional reasoning layer adds tone and regulatory-risk scoring; config stays in your environment.
How is accessibility handled before anything ships?
Cheap kernels catch blur, blanks, monochrome mush, and off-policy text before any model call. Contrast is held to real WCAG math on the render — body AAA at 7:1, accents AA at 4.5:1 — and OCR checks every image against its text policy.
What happens if a governance component fails or is offline?
Fail-safe by construction: a missing checker yields empty annotations, absent judges complete the panel, an unreachable store falls back to file logs, then memory. Coverage narrows, but output never drops to zero and copy is never blocked.
How is AI content made auditable and defensible?
Every run emits a manifest — objective, channel, spec and context hashes, per-modality lineage — stored on your disk, never Cohenta's. HTML/CSS assets re-render pixel-identically from the same spec, proving who generated what and why.
Can Cohenta publish anything on its own?
No — the run pauses at the gate, surviving restarts, until you press Approve or Request changes in the thread. Approval is a button press, never inferred; publishing takes a separate explicit confirm, and ad creative always lands paused.

Next step

See governance you can defend

Start with one brief. Watch every claim trace back to its source.

Get the RunnerExplore capabilities