Agentiks

Platform

The pipeline that proves what your model learned from.

Every sample your model trains on passes through one gate: its source judged, its content checked, its verdict sealed. This is the machinery, end to end.

0
trust signals per source
0
verdict outcomes at the gate
0%
of verdicts sealed
0
bytes leave your cluster

01 · End to end

One straight line, five stops.

Agentiks deploys from a Helm chart onto a streaming backbone (Kafka and Flink) inside your cluster. Data flows through it in one direction, and every stop leaves evidence.

Hospital site
Camera line
Vendor feed
Labeling team
samples, attributed and fingerprinted
Your Kubernetes cluster
Ingestion gate
Every sample enters attributed to an authenticated source.
Judge the source · Layer 1
Behavioral detectors feed a trust score and tier per source.
Gate decision
Reads the source's tier on every sample. Fails closed.
Quarantine / reject
pass
Check the sample · Layer 2
Embedding-space detectors issue per-sample verdicts.
↺ detections feed back into source trust
Audit ledger
Every verdict and score change, hash-chained and checkpointed.
↓ Clean data flows to training
↓ Certificates go to your auditor

02 · The trust engine

Grooming stops working when trust decays.

Five behavioral signals feed each source’s score, and recent behavior outweighs old history on a half-life of roughly three weeks. The classic attack, months of clean submissions to build trust and then a flip to poison, buys an asset that evaporates as fast as it turns hostile.

0TRUST SCORE1.0
Probation
Untrusted
Standard
Trusted
▸ New sources start here, whatever they claim to be
Trust grooming, defeated by decay
1.00MONTHS OF CLEAN SUBMISSIONSBEHAVIOR FLIPSTRUST COLLAPSES IN WEEKS
Tier flapping
Hysteresis

Regaining a tier costs more than losing it, so a source hovering at a boundary cannot flap on noise.

One noisy detector
No single-witness convictions

A lone detector family cannot destroy a source: penalties stay capped until an independent signal corroborates.

False alarms
Penalties that heal

A behavioral alarm opens a case. Clean escalated inspection refunds most of it; confirmation makes it permanent. Never all of it, so probing costs.

Framed victims
Victim discrimination

When a ring duplicates an honest source's content, per-member victim scoring spares the source that got copied.

Embedding space · livedetections
CLEAN REFERENCECOORDINATED RINGOUTLIER
Drift, coordination, and outliers are visible as shape.

03 · Check the sample

Bad data has a geometry.

Every sample is embedded (mapped into a space where similar content sits close together) and examined for what eyes cannot see: content that drifts from the clean reference, near-duplicates coordinated across supposedly independent sources, and clusters that change shape over time.

Each sample gets a verdict: pass, quarantine, or reject, with an escalation path for deeper checking. Detections flow back into the source’s trust score, so content evidence and behavioral evidence reinforce each other.

Your Kubernetes cluster

04 · Runs in your cluster

Nothing leaves. Nothing phones home.

01 · Install

One Helm chart into your cluster. Air-gapped works; there is nothing to call home to.

02 · Shadow one pipeline

Point a training pipeline's ingestion at the gate in shadow mode: observing and scoring, never blocking.

03 · Tune together

Detectors and trust thresholds calibrated on your real telemetry, with the founders in the loop.

KubernetesKafkaFlinkPostgreSQLRedis

Calibration defaults and detector internals are shared under evaluation, not published: printed thresholds are a gaming manual for the adversaries this system exists to catch.

05 · Operate and prove

Day two is a console, not a spreadsheet.

A control plane reads the whole pipeline live: trust trajectories per source, the verdict stream, and the sealed-ledger status. Underneath, every consequential event sits in an append-only, hash-chained ledger with Merkle checkpoints that an offline verifier can prove intact.

When an auditor asks, you export integrity certificates: per-sample attestations with inclusion proofs, in standard formats (in-toto, SLSA, CycloneDX).

Agentiks control planelive
Trust · 24h
arxiv-mirror-070.91
vendor-rlhf-030.58 ↓
Decisions
sample 8f21…PASS
sample c04a…PASS
sample 77d1…QUARANTINE
sample 3b90…REJECT
Ledger · sealed ✓ root 7b3e…f1Cert export · in-toto / SLSA / CycloneDX

Design partners

See it run on your data.

We deploy with a small number of design partners and tune the detectors on real telemetry, together.