01 What NRP is
The architecture behind "does Naro Assist actually understand this, or is it guessing?"
NRP (Naro Request Resolution Pipeline) is the architecture that decides how a chat request to Naro Assist gets handled: resolved deterministically by code wherever a request shape is common and well-understood, or handed to the full AI model loop when it isn't. AIOL (AI-Oriented Language) is the validation contract used at some of NRP's boundaries — it isn't the whole pipeline, just the envelope format that guards a write before it happens.
This matrix exists because "how many request types are actually deterministic" and "is X really guarded before it writes" kept needing to be re-derived from memory, session after session. Every row below was re-confirmed directly against the running code on the date above, with file/line references so the next check is fast — it's an evidence snapshot, not a permanent spec.
02 The six recognized routes
Everything else falls through to the full model loop — not a bug, the intended fallback (Flow B).
| Intent | Resolution | Execution | AIOL-validated | Live Capsule |
|---|---|---|---|---|
| task.complete | Deterministic (resolve_task_reference) |
Immediate write | Yes | No |
| task.reschedule | Deterministic (resolve_task_reference) |
Immediate write | Yes | No |
| task.create | Fast path: title + today/tomorrow only. Full tool: model picks fields | Immediate write, no confirm | Fast path only | No |
| event.create | None — always via the model | Confirm-gated | Yes | No |
| grocery.add | None on the direct path | Confirm-gated | Yes | No |
| grocery.restock | Deterministic, fuzzy-scored (resolve_restock_target) |
Resolves only — hands off to grocery.add's confirm gate | Yes | No |
"Fast path only" on task.create: the deterministic fast path builds and validates an AIOL envelope before writing. The full tool's model-loop call site still writes with no envelope/validation step — a known, still-open gap, not something the fast-path fix closed automatically.
03 Two execution shapes, not one
"Guarded write" means two structurally different things depending on the intent.
- Immediate write, precondition-guarded (
task.complete,task.reschedule) — no human confirmation step exists; safety comes entirely from resolving the target deterministically first (refusing on ambiguous/none) and re-checking the row hasn't changed since resolution, right before writing. - Confirm-gated proposal (
event.create,grocery.add, andgrocery.restock's eventual write) — the model, or a deterministic resolver for restock, only ever produces a pending proposal; the write happens in exactly one place, gated on an explicit human tap.
task.create is immediate-write like the first group, but has neither that group's precondition-recheck (there's no existing entity to race against on a pure create) nor the second group's human confirm gate — its own, third shape.
04 Context Capsules: one FACT producer live, RECALL still isn't
A Capsule is a structured, code-computed context payload handed to the model instead of raw data rows.
suggest_next_task's FACT capsule (deterministic ranking, no model judgment in the ranking itself) is now wired into real chat requests — triggered by narrow phrasing like "what should I work on next," read-only, with no write of its own. It's the first Context Capsule any real request actually reaches.
Everything else is unchanged: RECALL (journal-candidate retrieval) still has zero live call sites, and nothing that could feed a write is wired live until a review/confirm step exists to protect it. The "Live Capsule" column above still reads No for all six intents in the table — the FACT capsule is its own new, seventh, read-only route, not a delivery mechanism added to an existing intent.
05 What this implies for the next steps
As of the snapshot date above.
- task.create's fast path is closed. The fast path now builds and validates an envelope before writing, mirroring the pattern task.complete/task.reschedule already used. The full tool's model-loop call site remains a separate, still-open gap.
- Live FACT Capsule wired. suggest_next_task now delivers its capsule to a real request, per Section 4.
- Measured, not just claimed. A real benchmark run (local model, zero cost) found the live Capsule route both faster (2.685s vs. 4.174s) and more compact (23 vs. 49 tokens) than the old model-tool-loop path for the one case that needed a real model call — a single cold-start data point, not a large-N result, and the harness's correctness scoring can't actually judge a read-only recommendation either way.
- Next, not started. A real cross-app consumer (Bldman/Kanso/ACoP — none exists yet) to test AIOL's actual claimed benefit against; DIL work stays gated behind that.
For the fully current version of this table — re-verified against the code, with exact file/line references for every cell — see the raw evidence file: nrp-status-matrix.md.