Architecture for durable, governed AI execution.
Olympus is the control plane layer between enterprise applications and model vendors. It combines provider orchestration, policy enforcement, approvals, continuity, and replay into a single operating model. Delta is the replay and continuity engine underneath it.
Core technical claims
- Provider-independent sessions instead of vendor-owned working state.
- Delta replay lineage for reconstructable history and replayable context.
- Policy packs and approval thresholds enforced before execution.
- KV-cache compression benchmarks: 10.6× single-snapshot, 32–113× serving-level on validated workloads.
What sits in the Olympus layer
Provider orchestration
Route across OpenAI, Anthropic, Google, Microsoft, or local infrastructure without letting any one provider own the session state.
Continuity engine
Preserve session state, memory, branches, and working context across retries, failures, and provider changes. This is where Delta sits underneath Olympus.
Governance layer
Apply policy packs, approvals, rate limits, and identity-aware controls before requests leave the control plane.
The replay engine underneath Olympus has measured numbers
Delta is the continuity and replay engine that Olympus uses to preserve session state, reconstruct history, and keep audit trails inspectable. These are not marketing numbers — they come from the Delta benchmarking harness in the repo.
10.6×
Single-snapshot KV-cache compression via eviction + quantization, full round-trip through transformer decoding.
7.1×
Compression at the acceptable-quality threshold — where downstream accuracy stays inside acceptance bands.
32–113×
Effective compression at the serving layer across workloads — the number that matters for audit-trail storage economics at scale.
What this means operationally: a buyer storing 90 days of replayable AI sessions pays for a fraction of the memory footprint that a naive cache would require, without losing the ability to reconstruct decisions.
Where technical diligence usually goes next
Delta replay lineage
Replay Studio exists to answer a hard technical question quickly: what did the system know, what changed, and how did the execution path evolve? Delta is the engine that makes that reconstruction possible.
Policy packs and approvals
Olympus treats risk controls as versioned system state, not chat behavior. That matters when teams need repeatability and defensible operations.
Compression math that holds up under diligence
10.6× single-snapshot, 7.1× at the acceptable-quality threshold, 32–113× at the serving layer. Backed by a reproducible benchmark harness rather than marketing claims.
Deployment model
Support managed, hybrid, and on-prem patterns depending on where the data plane and control surfaces need to live.
Recommended reading order
TSC Research Series
Overview page for the paper set and the right first stop for technical evaluators.
Modern Architecture Paper
Use this when the conversation moves from product claims into system design and deployment logic.
Scale Validation
Use this when the evaluator wants evidence around behavior beyond a single demo path.