3cfaa1e1d3
Phase 1 delivers the complete durability and storage layer:
- WAL with crash recovery: Append-only journal with BLAKE3 checksums,
fsync guarantees, and proper seek-to-EOF on reopen
- Storage engine: sled-backed KVStore with scan_prefix for range queries
- Content-addressed storage: H:{hash}, V:{hash}, E:{hash} key patterns
- Ingestor: Background worker tailing WAL, writing to KV with 8-byte
aligned record headers for rkyv zero-copy deserialization
- Comprehensive tests: 31 tests covering crash recovery, round-trips,
and multi-cycle durability
New crates: stemedb-wal, stemedb-storage, stemedb-ingest
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
58 lines
1.9 KiB
Markdown
58 lines
1.9 KiB
Markdown
# The Gardener (TrustRank Back-Propagation)
|
|
|
|
> **Quick Ref:** Background worker that penalizes agents when corrections are made
|
|
|
|
## The Problem
|
|
|
|
Agents are stateless. When an agent makes a mistake (uses `requests` instead of `axios`), you correct it. But the agent doesn't "learn"—it might make the same mistake next session.
|
|
|
|
Current training uses "Golden Trajectories" (perfect examples). Mistakes are discarded, so agents never learn "don't do X because it fails."
|
|
|
|
## The Solution
|
|
|
|
The Gardener is a background worker that:
|
|
1. Detects when a user correction supersedes an agent assertion
|
|
2. Calculates the "delta" (how wrong the agent was)
|
|
3. Back-propagates the error to the agent's TrustRank
|
|
|
|
```rust
|
|
struct GardenerJob {
|
|
pub agent_id: AgentId,
|
|
pub topic: String, // e.g., "http_libraries"
|
|
pub prediction: Value, // What agent said
|
|
pub ground_truth: Value, // What was correct
|
|
pub delta: f32, // How wrong (0.0 to -1.0)
|
|
}
|
|
|
|
// Example:
|
|
// Agent asserted "requests" (confidence 0.8)
|
|
// User asserted "axios" (confidence 1.0)
|
|
// Gardener calculates: delta = -0.3
|
|
// Agent's TrustRank for topic "http_libraries" drops by 0.15
|
|
```
|
|
|
|
## Effects
|
|
|
|
**Next time this agent predicts an HTTP library:**
|
|
1. Its confidence is mathematically penalized
|
|
2. It's forced to look for external verification
|
|
3. Or the system routes to a different agent with higher TrustRank
|
|
|
|
## Resurrection Mechanics
|
|
|
|
When a constraint is queried and used successfully:
|
|
- `last_verified` updates to NOW()
|
|
- Confidence decay resets
|
|
- Assertion stays in "hot path"
|
|
|
|
When NOT used in 6 months:
|
|
- Confidence decays toward 0
|
|
- Moves to "cold store"
|
|
- Still queryable for audit
|
|
|
|
## Related
|
|
|
|
- [Negative Constraints](../services/lifecycle.md) - What gets stored
|
|
- [Lens::Constraints](../services/lens.md) - How it's retrieved
|
|
- [use-cases/agile-agent-team.md](../../use-cases/agile-agent-team.md#feature-6) - Full workflow
|