The mindset

• Triage, not proof. Lab detectors hit 95%+ but collapse to ~50–65% on real-world, re-compressed, platform-processed media. Combine the signals; never rest a verdict on one tell.
• Corroborate. A "tell" is a hypothesis. Confirm it with the model ensemble, the forensic panel, provenance, and a second region before you commit.
• Document everything. Mark the regions and write the reasoning — your rationale and cited regions are printed into the signed, verifiable report.
• Modern models fix old tells. Absence of artifacts is not proof of authenticity. Weight the ensemble + provenance accordingly.

Spotting AI-generated images

Scan these zones with the loupe at 100%+. Any one is a flag; two or more is strong.

Cross-check: does the OMNI-FORGE ensemble agree? Does the Grad-CAM heatmap light up the same region your eye flagged? Any C2PA "AI-declared" manifest is a hard call.

Spotting deepfake / AI video

Step frame-by-frame, then watch at speed. Faces are where most fakes break.

• Face boundary — shimmer/warping at the hairline, jaw, and where the face meets neck/glasses; a faint mask edge.
• Temporal flicker — skin tone, lighting, or teeth that jitter between frames; identity "wobble."
• Lip-sync — mouth shapes that lag or don't match phonemes; teeth that blur during speech.
• Blink & micro-motion — too few/too many blinks; frozen or unnatural eye movement.
• Lighting drift — face lit differently from the scene; reflections in glasses/eyes that don't track.
• Audio — robotic prosody, missing breaths, clipped consonants, room tone that doesn't match.

The heatmap shows the center frame; corroborate with several frames. Reused/recycled clips also show up in cross-image and provenance checks.

Spotting edited / Photoshopped photos

For a real photo with one altered region (the classic claim-fraud case), lean on the Manipulation panel + your eye.

• Splice edges — halos, soft "cut-out" outlines, mismatched grain/noise inside vs. outside a region.
• Cloning — repeated textures (foliage, debris, crowd) that shouldn't repeat.
• Light/shadow mismatch — an added object whose shadow direction or softness disagrees with the scene.
• Perspective — objects at the wrong scale or vanishing point.
• Metadata — EXIF stripped, edit-software tags, or capture time/GPS that contradicts the claim.
• Forensic maps — ELA, JPEG-ghost, and double-compression highlighting a region (triage ~65–78%; blind to generative fill — corroborate).

Using the workspace

• Zoom / pan / loupe — inspect at 100%+; the loupe magnifies under the cursor.
• Mark regions — draw a box over a suspect area and label it (AI / Photoshop / other). Marked regions auto-log as evidence and print into the report.
• Automated panel — OMNI-FORGE score + per-model bars, C2PA/EXIF/GPS, manipulation signals, and the heatmap overlay.
• Human-tells checklist — tick what you actually observed; it structures your rationale.

Filling out the verdict form

Pick the verdict that matches the evidence, calibrate confidence honestly, and justify it.

• Confidence — map it to evidence strength: High = multiple independent signals agree; Medium = one strong or several weak; Low / inconclusive = conflicting or compression-degraded. Don't inflate.
• Rationale — name the specific signals: "fused fingers on left hand (region A) + ensemble 0.91 + heatmap concentrated on hand; no C2PA." Cite the regions you marked.
• Inconclusive is a valid answer — if real-world degradation makes it a coin-flip, say so rather than guessing.
• Rush cases — same standard, faster turnaround; don't cut the corroboration step.
• Escalate — flag for a second reviewer when stakes are high, signals conflict, or the media is novel/edge-case.

Your verdict, confidence, rationale, and cited evidence are sealed into the cryptographically signed report and validated at /verify. Write it to be read by an adjuster, a lawyer, or a court.