Data Model

Status: Designed, not yet implemented. Object store and search exist from the build system; event log and policy engine are new.

The data model has four components: objects, an event log, reactive policies, and procedures.

Objects

Immutable, content-addressed, with provenance:

Field	Type	Description
id	hash	Content hash (SHA256)
type	enum	transcript, episode, observation, pattern, policy, procedure, …
author	string	Which agent or principal created this
source	list[hash]	Input object IDs consumed
created	timestamp	Creation time
region	enum	ephemeral, operational, structural, identity, glacier
confidence	float	Decays over time
content	text	Payload

Five regions, matched to information timescales: ephemeral (hours), operational (days–weeks), structural (months–quarters), identity (permanent until changed), glacier (archived, indexed, retrievable).

Event log

Append-only. Every mutation is a log entry:

seq_id | operation | object_id | author | causal_deps | timestamp

Operations: WRITE, ARCHIVE, PROMOTE, INVALIDATE, CONVENTION_CHANGE

Materialized views — search index, context documents, conventions directory — are derived from the log and can be rebuilt by replaying.

Tradeoff: Full replay capability requires retaining the complete log. Storage cost scales linearly with mutation count. Compaction strategies are not yet specified.

Reactive policies

Governance is expressed as policy documents stored as objects:

# Policy: Episode Summarization

trigger: new object of type `transcript`
procedure: summarize_episode
author: any agent with role `synthesizer`
output_type: episode
output_region: operational
evidence_requirement: minimum 1 source object

An agent reads the policy, checks if the trigger condition is met, executes the referenced procedure, writes the result. Dependency structure is implicit in the trigger conditions rather than declared as a DAG.

Procedures

Instead of Python transform classes, procedures are markdown documents:

# Procedure: Summarize Episode

Given one or more transcript objects, produce a single episode summary.

## Inputs
- 1+ objects of type `transcript`

## Prompt
[prompt template with {{placeholders}}]

## Output schema
- type: episode
- region: operational
- confidence: 0.8
- source: [input object ids]

The procedure is a versioned object in the store. Changing the procedure changes the behavior of any agent that reads it.

Open question: How reliably can LLM agents parse and execute arbitrary procedure documents? This is an empirical question the implementation needs to answer.

Confidence

Objects carry confidence scores based on provenance:

Source	Initial confidence
Human-provided	1.0
Raw ingestion	0.9
Cross-agent corroborated	0.7
Single-agent synthesis	0.5

Confidence decays over time. The rate is region-specific. These initial values are defaults — whether they are well-calibrated is an empirical question.