Loreflow
An AI product for enterprise knowledge retrieval (an enterprise brain) that reads from organizational sources such as Slack, GitHub and documents and answers questions based on cited evidence. Its core selling point is preventing cross-tenant leakage and prompt injection architecturally, through pure-function evidence projection, evidence budgets and markup isolation.
Timeline (last 90 days)
Loreflow is an "enterprise brain" product that reads a company's Slack, GitHub, and documents, then answers questions based on cited evidence.
Loreflow's evidence projection module is a pure function with no I/O and deliberately holds no database handle, request context, environment variables, logger, or model runtime.
Loreflow has a dedicated isolation test suite that asserts that one workspace's data never appears in another workspace's output.
Every Loreflow answer is assembled under a fixed evidence budget: at most 8 pieces of evidence, each 1000 code points, totaling 8000 code points, counted by code point rather than UTF-16 units.
Loreflow has a stricter sub-budget for live/federated evidence (the highest-risk third-party text): at most 3 of the 8 slots.
Loreflow's evidence budget is fixed by tests; relaxing them is a deliberate decision and will never be accidental.
Loreflow prohibits object spread syntax; every evidence object is explicitly constructed field by field, making the set of fields visible to the model a compile-time contract.
Loreflow's design covers only the answer pipeline; the retrieval layer, authentication layer, and connector layer have their own defenses (permission-mirrored reads, Clerk-scoped operations, fail-closed configuration).
Loreflow's budget, purity, and field list are all asserted by the test suite rather than written as comments.
Loreflow's design principle: in AI systems, the most reliable security properties are those guaranteed by construction, not those guaranteed by instructions.
1 more timeline events
All Facts (11)
Loreflow is an "enterprise brain" product that reads a company's Slack, GitHub, and documents, then answers questions based on cited evidence.
50%UnverifiedLoreflow's evidence projection module is a pure function with no I/O and deliberately holds no database handle, request context, environment variables, logger, or model runtime.
50%UnverifiedLoreflow has a dedicated isolation test suite that asserts that one workspace's data never appears in another workspace's output.
50%UnverifiedEvery Loreflow answer is assembled under a fixed evidence budget: at most 8 pieces of evidence, each 1000 code points, totaling 8000 code points, counted by code point rather than UTF-16 units.
50%UnverifiedLoreflow has a stricter sub-budget for live/federated evidence (the highest-risk third-party text): at most 3 of the 8 slots.
50%UnverifiedLoreflow's evidence budget is fixed by tests; relaxing them is a deliberate decision and will never be accidental.
50%UnverifiedLoreflow prohibits object spread syntax; every evidence object is explicitly constructed field by field, making the set of fields visible to the model a compile-time contract.
50%UnverifiedLoreflow's design covers only the answer pipeline; the retrieval layer, authentication layer, and connector layer have their own defenses (permission-mirrored reads, Clerk-scoped operations, fail-closed configuration).
50%UnverifiedLoreflow's budget, purity, and field list are all asserted by the test suite rather than written as comments.
50%UnverifiedLoreflow's design principle: in AI systems, the most reliable security properties are those guaranteed by construction, not those guaranteed by instructions.
50%UnverifiedInjection within Loreflow's evidence budget can still influence the answer within its 1000 code points; the budget limits the blast radius, and output validation is the next layer.
50%