> ## Documentation Index
> Fetch the complete documentation index at: https://docs.tensorpro.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Core concepts

> Provenance, deterministic retrieval, honest abstention, and the SIP.

## Provenance-aware memory

Every record EDN stores carries a provenance class that records where the
record came from. Provenance is not a label bolted on afterwards: it drives
how, and whether, a record is allowed to reach your model.

| Provenance class    | Origin                                                                                                 |
| ------------------- | ------------------------------------------------------------------------------------------------------ |
| `USER_STATED`       | Stated directly by the user                                                                            |
| `USER_CONFIRMED`    | Confirmed by the user                                                                                  |
| `DOCUMENT_VERIFIED` | Verified from an ingested document                                                                     |
| `AI_INFERRED`       | Inferred by a model; usable as context but always marked as an inference, never as an established fact |
| `AI_GENERATED`      | Generated by a model; confined to the uncertain section and marked uncertain                           |
| `SYSTEM_DERIVED`    | Derived internally by EDN; excluded from retrieval                                                     |
| `ROUND`             | Tied to a conversational round; surfaced as the primary context for the latest turn in the Workspace   |

## Confirmed fact versus excluded inference

EDN draws a hard line between what may be presented as fact and what may not.

* Confirmed facts (`USER_STATED`, `USER_CONFIRMED`, `DOCUMENT_VERIFIED`, and
  round records) can enter the fact-bearing sections of the context package.
* `AI_INFERRED` records may be included, but they are always flagged as
  inferences. They are never presented as established fact.
* `AI_GENERATED` records are structurally excluded from the identity, context,
  and preference sections. They may appear only in the uncertain section, always
  marked as uncertain.

This classification is enforced when the context package is assembled, not
filtered out afterwards.

## Deterministic, LLM-free retrieval

When your application asks for context, EDN ranks stored records by cosine
similarity between the query and pre-computed record embeddings. A fixed
similarity floor (0.45) separates genuine matches from loosely related noise,
and a provenance gate excludes classes that are not eligible for retrieval.

No generative model runs in the retrieval path. The same query over the same
records returns the same records. Retrieval is a deterministic database query,
not a model decision.

## Honest abstention

When a request is scoped to a project and retrieval returns no matching records,
EDN does not quietly fall back to general knowledge. It emits an explicit
grounded instruction telling the downstream model to answer only from the
project's sources, and to state plainly that the project has no sources on the
topic if none apply. Abstentions are recorded in telemetry so you can see how
often memory honestly declined.

## Citations

Because every retrieved record keeps its provenance and source, answers built on
EDN context can be cited back to their origin. The Workspace renders these as
inline citation markers with a provenance tag; your own application can use the
same provenance data to attribute answers.

## The SIP

The Session Initialisation Package (SIP) is the compact, governed context
package EDN assembles for a request. It has a fixed set of five sections and is
held to a fixed token budget (a target of roughly 800 tokens and a hard cap of
1200\), so what reaches your model stays small and predictable.

| Section             | Contents                                                                    |
| ------------------- | --------------------------------------------------------------------------- |
| `identity`          | Who the user is, built from `USER_STATED` and `USER_CONFIRMED` records      |
| `active_context`    | The facts in play, including `AI_INFERRED` records marked as inference only |
| `preferences`       | The user's stored preferences                                               |
| `calibration_notes` | Active calibration channels                                                 |
| `uncertain_facts`   | `AI_GENERATED` content, marked uncertain                                    |

<Note>
  If assembly would exceed the token budget the build fails rather than
  silently truncating, so the package your model receives is never quietly cut
  short.
</Note>