Public relations AI agent · Deepfake response
Misinformation and Deepfake Detection AI Agent
Swear only to what you can read back: preserve the item, extract the fields, mark what the score cannot establish, and leave the notice to the named counsel who signs it.
- Typical delivery
- 4–6 weeks
- Deployment
- Your stack
- Named counsel
- Oath first
- After launch
- Agent Care
What this agent does
Assembles the notice, never swears to it
In
- 01
An item is preserved, and the score it drew is written down beside it, never inside it.
- 02
A detector reading perfect on the benchmark reads worse than a coin toss on real audio.
Reason
- 03
Synthetic is not false, false is not unlawful, and only unlawful content can be notified.
- 04
A Content Credential is missing, and C2PA says in terms that its absence proves nothing.
- 05
The perjury clause at 512(c)(3)(A)(vi) covers the fields, so the extraction is what is sworn.
Decide
- 06
Good faith sits at clause (v), and under Lenz that belief is held by the signer or by nobody.
- 07
A counter-notice lands, and 512(g)(2)(C) opens the put-back window the record must survive.
Out
- 08
No instrument reaches the item, and that is recorded as the finding rather than as a gap.
- 09
Execute write actions only inside the approval boundaries agreed during implementation.
Product statement
Preservation, extraction and drafting belong to the agent. The oath, and the name that goes under it, belong to the counsel who is cross-examined on it later.
Example workflow
One notice, item to filing
- 1Item referred inA flagged upload, a customer report, a broadcast clip or a link handed over by the listening lane
- 2Item preserved and datedThe capture, the hash, the archive location, the uploader account and the hour it was taken
- 3Notice fields extractedThe work asserted, the locations, the account identifiers and detector confidence
- 4Controls appliedProvenance checks, own-content checks, instrument checks and extraction confidence
No human action required
Stages 1 to 4 run unaided, and nothing is filed at any of them — the agent is assembling, and the counsel lane opens at the signature gate.
Goes to named counsel to sign.
Adds a brand-protection read first.
Counsel review
The packet is held with its capture, its score and the steps the agent could not take.
Sign · Amend packet · Send to brand-protection review
Filed — by named counsel
- 6Matter and takedown records updatedOnly where write access and records policy allow it
- 7Outcome evaluatedNotice outcomes, counter-notices received, counsel corrections and what the later review turned up
Corrections
Each correction counsel makes counts in the evaluation.
What should not run autonomously
Human approval stays in control
Outside the boundary — human approval required
8 items
- Signing a notice under penalty of perjury.
- Judging an item unlawful, not merely synthetic.
- Deciding to sue rather than watch it return.
- Naming the copyrighted work said to be infringed.
Automation boundary
Agent acts unaided
- ✓Preserve the item, hash it and date the capture.
- ✓Extract each notice field and record which source produced it.
- ✓Mark which of the three translations it cannot make on this item.
- ✓Hold the packet for the named counsel who signs it.
Nothing is filed with a platform except by named counsel, inside the agreed boundaries.
- Forming the good-faith belief clause (v) wants.
- Telling a platform the packet is complete.
- Choosing which remedy an item is filed under.
- Changes to the holding rules or the signature gate.
Example output
One notice packet, annotated
A listening agent found the item and a crisis agent will answer for it; this one assembles what a signature would rest on. Below is one packet exactly as the agent leaves it.
Notice packet · single item
Illustrative example- Item
- Executive video, one account
- Recorded as
- Synthetic on the audio track, unproven
- Classified
- Score only
- Evidence of record
- Capture hash, 4 August 2026
- Confidence
- Held unsigned
- Held for
- Named counsel, by name
As received
Drawn from the capture, the hash and the account record, and it asserts nothing those three establish.
What the record holds
- Capture and hash
- Uploader account
- Cross-post graph
Why no filing here
Swearing to the notice is an act that named counsel performs alone.
What the score decides
Below the configured threshold a packet gets a brand-protection read before counsel sees it.
Action
SignAmend packetSend to brand-protection reviewValue
Where AI adds value
The same four claims, placed at the point in the workflow where each one applies.
Where the value lands
Value 01 – 04
Each notice
From the item it names
Where the notice is filed
Under Lenz the good-faith belief is subjective and must be held by the signer, and willful blindness can supply the knowing misrepresentation that 512(f) punishes.
Sworn to, not proven
The oath at clause (vi) runs to the fields a machine filled in; the claim that the item infringes sits at clause (v), on good faith alone.
What was checked, and not found
No instrument consulted takes a detection score as grounds for removal. TAKE IT DOWN runs a forty-eight-hour clock on intimate imagery alone; both California deepfake statutes are permanently enjoined; NO FAKES is still a bill; and the FTC rule on impersonating an individual was never finalised.
Build an evidence trail
The item, the score it was given and the person who swore to the notice stay together.
Integrations
Typical integrations
Five system groups connect to the same agent. Which of them are in scope is decided in discovery.
Detection and provenance
- Detector APIs
- manifest readers
- Model scores and credentials
Evidence preservation
- Capture
- hashing
- archives
- Timestamped copies of items
Notice and platform channels
- Platform notice forms
- DSA portals
- Filed notices and the responses
Agent
Misinformation and deepfake notice
- Reads the item
- Assembles the packet
- Holds the signature
Matter and counsel records
- Matter management
- legal hold
- Signature trails and filed versions
Observability & evaluation
- OpenTelemetry
- Langfuse
- Supported monitoring/evaluation sources
Integration availability depends on the client's existing systems and API access.
Agent controls
Six filters between the score and the oath
Six filters in one train, the last the tightest. What survives is set out in the map below.
L6 · Outermost — last line of defence
Inward → L1 · closest to the model
- L1 – L2
- Test whether a notice may be sworn
- L3
- Leaves the oath to named counsel
- L4 – L5
- Keep the notice and the item behind it
- L6
- Escalates to a human read when signals degrade
How Nestack evaluates it
Evaluate the whole assembly run — not only the notice that gets signed.
Coverage runs the whole depth of the workflow, and every layer is cut by slice.
Surface — the notice a platform receives
- E1Final-output evaluationDid the packet stay inside what the capture actually established?
- E2Step-level evaluationDid the agent read the right item, the right account and the live holding rules?
- E3Tool evaluationDid it read and write the correct item and the correct matter record?
- E4Confidence calibrationDo low-confidence packets actually attract more counsel corrections?
- E5Slice evaluationHow does performance change from one generator class to another?
- E6Business outcomeHow many packets needed a correction before counsel signed one?
Floor — the packet read back in discovery
Failure modes
Where each failure originates in the agent
Seven failure modes, each set at the stage it first shows itself.
Agent lifecycle
Direction of processing →
01 · Retrieval
1 mode
- UD-03
Absence read as evidence
A missing credential is worked as a fakery signal.
Stage gathers
The item, the score, the account and the hash
02 · Reasoning
2 modes
- UD-04
Benchmark score read as field
A laboratory number is used as a live one.
- UD-06
Our own footage flagged fake
An authentic clip trips the detector.
Stage proposes
The capture, the score, the work and what stays open
03 · Tool / write
2 modes
- UD-02
Copyright asserted, none owned
The notice names a work nobody here holds.
- UD-05
No instrument reaches the item
The packet is sound and nothing removes it.
Stage writes
Only where write access and approval policy allow it
04 · Output
1 mode
- UD-01
Channel suspended for volume
The platform stops processing what we send.
Stage returns
The notice a platform reads and counsel signs
05 · Change / Version
1 mode
- UD-07
The put-back nobody diarised
The item returns while the case reads closed.
Stage tracks
Model, prompt, filing rules and detector builds
- Sev-1 · a notice sworn on an unread score
- Sev-2 · an authentic clip is notified
- Sev-3 · score thin, packet held back
Affected slices
Unseen generators absorb the corrections
A class-level packet-accuracy figure can read clean while unseen generators carry most of the counsel corrections. Nestack reports the correction rate by generator class, not only in total.
Slice performance — reported separately, not only in aggregate
Illustrative example| Slice | Failure rate | Lift | Lift vs. threshold | Status |
|---|---|---|---|---|
| Unseen generator families | 9.5% | 3.7× | Review | |
| Cloned executive audio | 6.7% | 2.6× | Review | |
| Re-encoded platform copies | 4.2% | 1.6× | Watch | |
| Known generator families | 2.2% | 0.9× | Normal |
Bar: correction-rate lift vs. known-generator baseline · scale 0–4.0× · tick marks the 2.0× review threshold
2 of 4 slices over threshold
Evidence-linked improvement
What an unread score costs
A cycle shuts when the notice sworn on a score nobody opened is a standing case. That suite is what the next takedown filed is measured against.
Improvement cycle · five stages
Switchback — the path turns at Improve and returns at Learn
- 01
Detect
Correction rate rises on unseen generator families.
- 02
Diagnose
The notice sworn out on a number nobody opened, against a clip the company had itself upscaled, is worked backwards until one cause is left standing.
- 03
Improve
Notices ship numbered, and the items behind them travel attached.
- 04
Verify
Each touched notice case is run again, and one red holds it back.
- 05
Learn
The case is kept, and the review rules are amended in that same commit.
Learn → Detect
The return edge. The next item meets a suite one case longer.
Typical build scope
Twelve workstreams across six weeks
The build scope read against the delivery timeline. Week structure follows the six-week plan — discovery, sources, packet assembly, evaluation, integration, then production validation and handover.
Workstream
Week 1Week 2Week 3Week 4Week 5Week 6
- 01Deepfake-notification and automation-boundary scope.Week 1
- 02Detection, capture and notice sources.Week 1, Week 2
- 03Statutory-instrument and counsel-signature mapping.Week 1, Week 2
- 04Item capture and intake.Week 2
- 05Field binding and packet logic.Week 2, Week 3
- 06Confidence scoring and review routing.Week 3
- 07Counsel signature workflow.Week 3, Week 4
- 08Matter-system integration.Week 5
- 09Review and notice cases.Week 4, Week 5
- 10Guardrails and filing controls.Week 4
- 11Notice-trail instrumentation.Week 5, Week 6
- 12Deployment, documentation and Agent Care handover.Week 6
12 workstreams · 6 weeks · bar shows the weeks a workstream is active — several run in parallel
Final scope and sequence confirmed in discovery
Engagement tiers
What each tier includes
Rows are the capabilities named in each tier's scope. Higher tiers include everything below them.
| Capability✓ in scope · — not at this tier | PilotOne generator class, one cycle | ProductionProduction notice workflow | AdvancedMultiple platforms / jurisdictions |
|---|---|---|---|
| Introduced at Pilot | |||
| Assembly to your holding rules | ✓ | ✓ | ✓ |
| Named counsel signature | ✓ | ✓ | ✓ |
| Detector-baseline measurement | ✓ | ✓ | ✓ |
| Introduced at Production | |||
| Reporting by generator class | — | ✓ | ✓ |
| Brand-protection review workflow in your systems | — | ✓ | ✓ |
| Approved matter write-back | — | ✓ | ✓ |
| Detector-and-registry integration | — | ✓ | ✓ |
| Introduced at Advanced | |||
| Multi-regime notice rules | — | — | ✓ |
| Cross-platform notice packs | — | — | ✓ |
| High item volume | — | — | ✓ |
| Multi-platform notice controls | — | — | ✓ |
| Build price | From $5,000 | From $8,000 | Custom quote |
- Final build price
- Confirmed after discovery based on integrations, workflow complexity, item volume, signature controls and deployment requirements.
- Separate from build
- Build pricing is separate from recurring Agent Care, which covers managed monitoring, evaluations, incidents and verified improvements after launch.
What we need from you
What you bring, and what we build with it
Each input maps to a piece of build scope and a week in the delivery timeline.
You bring
We build with it
01Your live item classes and the instrument each one points at
→Holding-rule capture and detector versioning
Week 102Representative items, captures and past notices
→Field binding, packet logic and the detector baseline
Week 203Your escalation path and the counsel it names
→Holding-rule mapping, instrument binding and the automation boundary
Week 104Access to relevant APIs, feeds or exports
→Detection, capture and matter source assessment, then integration setup
Week 205Notices you would not want cross-examined
→Review cases and the evaluation round
Week 406What no score may be sworn to
→Confidence scoring, review routing, guardrails and filing controls
Week 307Named counsel who signs the notice
→Handover to named counsel, then pilot and production validation
Weeks 5–6
Nothing else is required
Deployment, documentation and Agent Care handover are ours.
Delivery timeline
Four phases across six weeks
A band here is as wide as the phase actually costs, so week five shows a pair rather than a tidier blank.
Phase
W1
W2
W3
W4
W5
W6
- DiscoveryW1
- BuildW2 – W3
- EvaluateW4 – W5
- Pilot & LaunchW5 – W6
Week focus
W1Holding-rule discovery, detector versioning and the automation boundary
W2Detection and capture integration and the packet baseline
W3Field binding, packet logic and filing controls
W4Evaluation suite, review cases and failure-mode testing
W5Matter-system integration, pilot items and targeted corrections
W6One detection cycle run under the brand-protection lead, then Agent Care handover
- Reading the band
- No bar runs past the weeks its own work is named for, and week five is the one overlap.
- At the end of W6
- Once the notice record validates, Agent Care adopts the agent.
- Duration
- Six-week plan shown · typical delivery 4–6 weeks depending on scope confirmed in discovery.
Next step · Public relations AI agent
Build a deepfake detection agent around the notice your last incident nearly swore out.
Show us one item you wanted taken down and the notice you drafted for it. Not how fast the detector flagged it. Which work you named as infringed, what the score actually established, and whose name would have gone under the oath. A score nobody opened comes back as a case.
- Typical delivery
- 4–6 weeks
- Deployment
- Your flagged items
- Named counsel
- Oath first
- After launch
- Agent Care available
Nestack Agents · Deepfake response
AGT-COM-08 · Agent Care available after launch