Don’t censor
We do not add hidden content filters or silent policy layers to inference requests. You choose the model behavior, prompts, and workflow constraints.
Audit every inference
Audit traces turn inference from logging into verifiable records. Off-chain receipts, hashes, signatures, optional on-chain commitments, and preview evidence packs combine into an auditable inference system
TLS 1.3Hash-only audit inputsEd25519 signaturesClean ComputeSOC 2 path
Request
→
SHA-256 Hash
→
Ed25519
→
Chain Link
→
Store
→
Verify
Three Principles
The operating model behind our public product and private deployments.
Don’t know
If we cannot verify a model fact, billing edge case, or infrastructure condition, we say so directly instead of fabricating certainty.
Can’t control
We separate BatchIn guarantees from upstream model behavior. Auditability, routing, and billing are ours. Model weights and vendor training choices are not.
Verifiable output path
Every supported audit record can be hashed, signed, and externally verified. The goal is operational trust, not marketing trust.
Hash only
Input and output content are reduced to hashes for audit records. Raw prompts do not need to live in the verifier layer.
Ed25519 signatures
Each record is signed with an Ed25519 keypair so external verifiers can reproduce the trust check in their own environment.
GPU-linked audit metadata
Model id, GPU id, timestamp, and record hashes are bound together in the signed payload.
Optional on-chain anchoring
When enabled, audit references can link outward for third-party timestamp verification without making chain storage mandatory.
Security posture
- Transport · TLS 1.3 in transit, strict security headers, and httpOnly secure session cookies.
- Storage · Secrets stay server-side. Audit records store hashes and signatures, not plaintext prompts.
- Access control · Dual-layer auth in middleware and server routes, plus rate limits and per-key policy enforcement.
- Operational controls · Logging, key rotation support, CORS restrictions in production, and a formal SOC 2 readiness path.
Clean Compute
AI inference is an infrastructure product. BatchIn can include energy and carbon estimate fields in audit records so compute remains accountable by default without exposing sensitive deployment geography.
Grid factor
~60 gCO2e/kWh
Baseline cloud reference
~400 gCO2e/kWh
Per-record fields
energy_wh · carbon_grams_co2e · energy_source
How audit traces compare
| Feature | BatchIn Audit Traces | Standard Logging | Blockchain Verification |
|---|---|---|---|
| Overhead | < 10ms | ~0ms | 500ms-30s |
| Tamper evidence | Hash chain + Ed25519 | None | Immutable ledger |
| Client-side verify | Yes | No | Yes, expensive |
| Cost per record | $0 | $0 | $0.001-$0.01 |
| Privacy | Hash-only, no raw content in audit | Full content often logged | Hash commitments on-chain |
Need a deeper review?
We can walk through the audit model, security headers, cookie strategy, and deployment posture with your team.