Architecture · Intent & ECO
Dynamic Reclassification
After the initial ECO is produced, an analyst agent may reclassify the ECO dimensions once (maximum twice for Lane C tasks). This is the only LLM-gated step before the Task Orchestrator.
Why Reclassification Exists
Heuristic intent detection and structural retrieval can produce a technically valid ECO that is contextually wrong. The analyst provides a lightweight reasoning pass without replacing the structural evidence.
Hard limit
The analyst may reclassify at most twice per task. A third reclassification attempt is rejected — the task is either escalated to Lane C with the current ECO or halted for human review.
What the Analyst Can Change
| Field | Can reclassify? | Notes |
|---|---|---|
intents[] | ✅ Yes | Replace or add a second intent |
| Dimension severity | ✅ Yes (escalate only) | Can raise severity, never lower |
lane_forced_min | ✅ Yes | Can raise, never lower |
confidence_score | ❌ No | Deterministic — not overridable |
eco_snapshot raw data | ❌ No | Immutable once written |
Reclassification Trace
Every reclassification is recorded in the Decision Ledger:
{
"event": "eco_reclassification",
"attempt": 1,
"analyst": "agent:v2.1",
"changes": {
"intents": { "from": ["FEATURE_ADD"], "to": ["FEATURE_ADD", "REFACTOR"] },
"lane_forced_min": { "from": "A", "to": "B" }
},
"rationale": "Spec references restructuring of auth module boundary",
"timestamp": "2026-03-22T10:14:22Z"
}