Principals
LiveEvery human and agent acts as a named Principal, with passkey and DID-based identity. Delegated agents are scoped to named ledgers and surfaces.
Dual Substrate System
One operating system for identities, data lineage and autonomous agents.
DSS works like an operating system for governed AI memory. Every person and every agent is a named Principal, and every answer traces back to the ledger entries it drew on.
Complete infrastructure to run governed AI
Like an operating system, DSS gives you identities, memory, permissions, interfaces and an audit trail in one layer, rather than scattered across applications.
| In an operating system | In DSS |
|---|---|
| Users and processes | Principals — people, runtime models, organisations and delegated agents |
| Memory and the filesystem | Ledgers — append-only, versioned, git-backed memory boundaries |
| Applications and interfaces | Surfaces — Chat, Decode and the Control Plane |
| File permissions and access control | Grants — managed in the Control Plane and stored as versioned state |
| System log and audit trail | Lineage — every entry keeps the entries it derives from |
| Drivers | Model routing — many providers behind one interface, using your own keys |
An analogy: it is not a kernel, it does not schedule processes, and it makes no real-time guarantee.
- Principalswho
- Grantswhat they may do
- Ledgerswhere memory lives
Grants
LiveGrants are managed in the Control Plane and stored as versioned state.
Lineage
LiveEntries record the entries they derive from. Walk the lineage of any entry — ancestors and descendants — with bounded, fast reads, and see what an answer drew on. A grounding guard checks numeric claims in metric answers against the retrieved context.
Coordinates
LiveRecords sit at prime-indexed coordinates, and the relationships between them — shared primes, parents, children — are derived from the coordinates. Retrieval runs against an index rebuilt from the git ledger, so the ledger stays the source of truth. A p-adic-inspired design, not a claim about p-adic numbers.
Ledgers
LiveLedgers are versioned and git-backed. New entries are added; prior versions stay readable; writes never force-overwrite history. The trust anchor is the git repository that holds them.
Product surfaces
Each surface is a way into the same governed memory.
Chat
Start a governed conversation on the DSS Chat surface.
Open ChatDecode
Explore coordinates, decodings, and benchmark results.
Open DecodeControl Plane
Manage Ledgers, Principals, Grants and Surfaces.
Open Control PlaneThe advantage of a unified, structural layer
Identity, permissions, memory and agent workflows run as one layer, so what an agent did is a matter of record rather than of log-grepping.
Traceable by construction
Every entry keeps its parents and children. You don't guess where an answer came from; you walk back to the entries it used.
Grants you can manage
Principals, grants and access windows are managed in the Control Plane, and the models each principal can pick follow those grants. Ledger-level enforcement is a configurable mode.
Cross-checked, not trusted
Two independent services verify decode reports against a shared invariant checker: the middleware re-verifies every decode report it catalogues. Mismatches are refused, and a failing write is quarantined rather than served.
Compounding structure
Each write adds relationships to the ledger. In one small controlled experiment on a single model, history-aware retrieval beat a no-history baseline, and because the index is rebuilt from the ledger, what the system accumulates is auditable.
Operational today. Honest about what isn't.
| Operational control plane | Live | Principals, grants, model bindings, surfaces and ledger connections are managed in the Control Plane, running in production on Fly.io. |
|---|---|---|
| Bring your own models | Live | Routes to OpenRouter, OpenAI, DeepSeek, Azure, Google Vertex and others using keys you supply (or operator-provisioned keys for approved accounts). A guardian model performs intent checks, and assessment checks when retrieved context exists, and logs telemetry. |
| Fast where it counts | Benchmarked | Retrieval-router lookups measured under 1 ms in published micro-benchmarks on small ledgers (hundreds of entries). The first progress event arrives in under a second; the first words of an answer take several seconds and a full answer about 20 seconds. |
| Public benchmarks | Benchmarked | DSS-EVAL is open: perfect abstention precision (1.000 on 300 queries per seed), and needle and multi-hop recall well above a vector baseline on small suites. The report also lists the cells we don't pass: two integrity gates, a 0.14 recall@1 ranking track, and an ablation where the full configuration scores 0.167. It is in the repository. These benchmarks test structure and abstention, not general retrieval quality. |
| Limits today | In progress | Ledger-level authorisation is configurable but not yet enforced by default. No compliance certification, SSO or SLA yet. Ledgers are anchored to a git repository whose owner can still rewrite history. We'd rather you hear that here. |