Find nodes where investigation would most reduce graph uncertainty.
Optional parameters: $ARGUMENTS
retrieve(action: "frontier", limit: 5, min_gap: 0.3)Response:
{
"status": "ok",
"count": 5,
"frontier": [
{
"node_id": "node_abc123",
"content": "BM25 hybrid retrieval improves SHR by 12pp...",
"confidence": 0.625,
"evidence_count": 1,
"interval_lower": 0.0833,
"interval_upper": 0.9683,
"width": 0.885,
"information_gain": 0.0828,
"access_count": 0
}
]
}Key fields:
interval_lower / interval_upper — Wilson score 95% confidence bounds
width — interval width (upper - lower); wider = more uncertain
information_gain — expected interval narrowing from one more evidence point
evidence_count — number of outcome observations backing this node
Only nodes with evidence_count > 0 appear on the frontier. Nodes without evidence are either current or stale, not "uncertain" in the statistical sense.
Query frontier — retrieve(action: "frontier", limit: 5, min_gap: 0.3)
Pick highest info_gain node — investigate it (run tests, check docs, verify claims)
Report outcome — learn(action: "from_outcome", causal_node_ids: ["<frontier_node>"], status: "success", ...)
Re-query frontier — the investigated node should have narrowed or dropped off
Repeat — each cycle reduces graph uncertainty where it matters most
The frontier uses Wilson score confidence intervals (z=1.96 for 95% confidence):
1 evidence point: very wide intervals (~0.89 width) regardless of confidence
2 evidence points: significantly narrower (~0.58 width)
5+ evidence points: reasonably tight bounds
10+ evidence points: converging on true confidence
This is why the frontier is powerful: it tells you exactly where one more data point would have the biggest impact.
Every learn(action: "from_outcome", ...) call increments evidence_count and updates confidence on causal nodes. This naturally narrows their Wilson intervals and may drop them off the frontier.
Use frontier data to inform coverage decisions:
route(action: "review_goal", goal_id: "...", signal: "frontier shows 3 high-uncertainty nodes in this area")If key nodes are on the frontier, coverage review may route to learn instead of act.
The attention engine can incorporate frontier data when prioritizing goals — goals with high-uncertainty supporting knowledge may need investigation before execution.
Don't ignore high-information-gain nodes — they're where learning has most value
Don't assume evidence_count=0 nodes are "fine" — they're just not measurable yet
Don't set min_gap too low — very narrow intervals aren't worth investigating
Always report outcomes after investigating frontier nodes — that's how the frontier shrinks