Corpus AI by Verifyco · Research preview

Detection signals, turned into evidence you can question.

Corpus AI is a conversational media-analysis assistant in development. It does not replace Verifyco's checks; it turns their available findings into cited answers, counterevidence, uncertainty and practical next steps.

In development Private research preview
Inspect the sample case Read the model card
Illustrative interface · Sample data · No live model call
CASE VC-042 CORPUS AI · RESEARCH BUILD ILLUSTRATIVE
00:41 press_statement.mp41080 × 1920 · shared copy · SHA-256 4fd8…a21c EXAMINATION READY
Evidence ledger6 layers
E-01C2PA provenanceNo credentials found · inconclusive E-02Metadata contextRe-encoded twice · weak alone E-03Visual region findingTexture boundary · frames 112–178 E-04Temporal findingLip-sync drift in same segment E-05Audio counterevidenceNo splice artefact detected
Corpus AI conversationModerate confidence

The strongest thread is agreement between two independent findings. The visual boundary appears in [E-03] during the same segment as the audio-visual drift [E-04].

Supporting[E-03] [E-04]
Counterevidence[E-05]
UnknownOriginal source unavailable
Recommended next check

Find the earliest available source and compare the flagged segment directly.

Sample values are illustrative and uncalibrated.Evidence, not an oracle.
01 · Product boundary

A detector score is the beginning, not the explanation.

Corpus AI is the conversational examination layer after detection. The boundary is explicit so a polished answer never pretends to be stronger than its underlying evidence.

  1. 01Verifyco detectors

    Produce layer-level findings from the iOS and web verification workflows.

    Live foundation
  2. 02Evidence normalisation

    Maps each finding to a source, region, time range, quality and limitation.

    In development
  3. 03Corpus AI conversation

    Answers questions from available findings, counterevidence and uncertainty, then cites what it used.

    In development
  4. 04Human verification

    A reviewer inspects the cited regions, uncertainty and recommended next checks.

    Product principle
Execution boundary

Verifyco's iOS analysis can run on-device. Verifyco Web uses its documented web workflow. Corpus AI's final deployment, retention and device/server split are still being evaluated and will be published before preview access.

02 · Sample examination

Inspect the evidence trail. Then ask.

One static-first case, four useful questions. Every answer names what supports it, what argues against it and what remains unknown.

CASE VC-042 · press_statement.mp4 Illustrative sample · no live inference
Source region00:11.2–00:17.8

Flagged region shown for orientation. The source media and all values on this page are illustrative.

Evidence ledgerselectable context
E-01C2PA provenanceInconclusive E-02MetadataContext only E-03Visual regionSupporting E-04Temporal alignmentSupporting E-05Audio integrityCounterevidence E-06Frequency patternContextual
Corpus AI · cited answerConfidence: moderate

Two independent findings point to the same segment. The localized texture boundary [E-03] overlaps the audio-visual timing drift [E-04]. Metadata [E-02] only shows re-encoding, while clean audio [E-05] is counterevidence. The original source is still missing.

Evidence used
[E-03] · [E-04] · [E-02]
Counterevidence / uncertainty
[E-05] is clean; no original source for comparison.
Suggested next check
Find the earliest source and compare frames 112–178.
03 · Architecture

How Corpus AI handles evidence.

A visible evidence path, not a magic orb: detector outputs become normalised claims, cross-modal relationships and a cited examination for human review.

Research position

Corpus AI is informed by the broader explainable multimodal-forensics frontier. Related work includes FakeShield (ICLR 2025) ↗; that work does not validate Corpus AI's wider media or deployment claims. Corpus AI surfaces evidence summaries and citations—not hidden internal reasoning.

04 · Living model card

Trust requires published boundaries.

No invented benchmark score and no vague promise. These are the release gates Verifyco intends to document before a Corpus AI preview is treated as a product.

Defined

Current scope

Reason over structured outputs from Verifyco's forensic layers; cite supporting evidence, counterevidence and unknowns.

Not a standalone detector · Not autonomous adjudication
Evaluation target

Protocol & calibration

Separate held-out sources, manipulation families and compression conditions; measure citation correctness and calibrated uncertainty.

Metrics will be published after measurement
Evaluation target

Known failure modes

Weak detector inputs, missing originals, heavy recompression, domain shift and conflicts between modalities can all produce inconclusive results.

“Inconclusive” remains a valid outcome
Open boundary

Privacy & deployment

Device/server split, retention, training-data use and deletion controls are release-gated decisions—not assumptions hidden behind a UI.

Boundary to be published before preview
  1. Foundation · live

    Verifyco iOS and Web verification surfaces provide the evidence workflow Corpus AI will build upon.

  2. Corpus AI · in development

    Case conversations, evidence citations, counterevidence and uncertainty language.

  3. Preview gates · required

    Calibration, generalisation, red-team evaluation, privacy boundary and versioned model card.

05 · Research access

Request a place in the private preview.

The form prepares an email to Verifyco; it does not silently subscribe you or claim your request was stored.

Your email app opens with a prepared request. Send it to complete the request. · Privacy Building in media integrity? Talk to Verifyco →