Add a local label filter for presence silence and no-cue intent

CHANGELOG.md

@@ -7,6 +7,15 @@ loosely while pre-1.0 (breaking changes can land on minor bumps).
 
 ## [Unreleased]
 
+- **Local label filter for presence silence and no-cue intent.**
+  With `ARBITER_DECISION_MODEL` set to an `ollama/…` id, a closed-set
+  next-token distribution can skip a presence review on an extreme
+  `silent`, or route a no-cue utterance without the advisor completion.
+  Both margins default to off. The filter cannot write a note, a brief
+  beyond the canned route line, or a halt. See
+  [Presence](docs/concepts/presence.md#silence-filter) and
+  [Intent](docs/concepts/intent.md#hybrid-classify).
+
 ## [0.13.11] — 2026-09-21
 
 - **Fleet dashboard pane (Phase C1).** The TUI consumes fleet SSE (`stream_id` +

CMakeLists.txt

@@ -195,6 +195,7 @@ set(ARBITER_SOURCES
     src/constitution.cpp
     src/orchestrator.cpp
     src/intent.cpp
+    src/label_score.cpp
     src/reconcile.cpp
     src/advisor_gate.cpp
     src/advisor.cpp
@@ -560,6 +561,7 @@ if(INDEX_BUILD_TESTS)
     add_executable(unit_api_client
         tests/test_api_client.cpp
         src/api_client.cpp
+        src/label_score.cpp
         src/circuit_breaker.cpp
         src/metrics.cpp
         src/json.cpp
@@ -582,6 +584,7 @@ if(INDEX_BUILD_TESTS)
     add_executable(unit_api_client_pool
         tests/test_api_client_pool.cpp
         src/api_client.cpp
+        src/label_score.cpp
         src/circuit_breaker.cpp
         src/metrics.cpp
         src/json.cpp
@@ -603,6 +606,7 @@ if(INDEX_BUILD_TESTS)
     add_executable(unit_api_client_pool_cancel
         tests/test_api_client_pool_cancel.cpp
         src/api_client.cpp
+        src/label_score.cpp
         src/circuit_breaker.cpp
         src/metrics.cpp
         src/json.cpp
@@ -846,6 +850,7 @@ if(INDEX_BUILD_TESTS)
         src/atomic_file.cpp
         src/json.cpp
         src/api_client.cpp
+        src/label_score.cpp
         src/circuit_breaker.cpp
         src/metrics.cpp
         src/tui/opentui/pane_input_editor.cpp
@@ -969,6 +974,7 @@ if(INDEX_BUILD_TESTS)
         src/workspace_root.cpp
         src/json.cpp
         src/api_client.cpp
+        src/label_score.cpp
         src/circuit_breaker.cpp
         src/metrics.cpp
         src/styled_text.cpp
@@ -1015,6 +1021,7 @@ if(INDEX_BUILD_TESTS)
         src/workspace_root.cpp
         src/json.cpp
         src/api_client.cpp
+        src/label_score.cpp
         src/circuit_breaker.cpp
         src/metrics.cpp
         src/styled_text.cpp
@@ -1048,6 +1055,7 @@ if(INDEX_BUILD_TESTS)
         src/workspace_root.cpp
         src/json.cpp
         src/api_client.cpp
+        src/label_score.cpp
         src/circuit_breaker.cpp
         src/metrics.cpp
         src/styled_text.cpp
@@ -1086,6 +1094,7 @@ if(INDEX_BUILD_TESTS)
         src/workspace_root.cpp
         src/json.cpp
         src/api_client.cpp
+        src/label_score.cpp
         src/circuit_breaker.cpp
         src/metrics.cpp
         src/styled_text.cpp
@@ -1390,6 +1399,7 @@ if(INDEX_BUILD_TESTS)
         src/event_routing.cpp
         src/constitution.cpp
         src/api_client.cpp
+        src/label_score.cpp
         src/circuit_breaker.cpp
         src/metrics.cpp
         src/json.cpp
@@ -1614,6 +1624,7 @@ if(INDEX_BUILD_TESTS)
         src/a2a/sse_reader.cpp
         src/constitution.cpp
         src/api_client.cpp
+        src/label_score.cpp
         src/circuit_breaker.cpp
         src/metrics.cpp
         src/json.cpp
@@ -1655,6 +1666,7 @@ if(INDEX_BUILD_TESTS)
         tests/test_constitution.cpp
         src/constitution.cpp
         src/api_client.cpp
+        src/label_score.cpp
         src/circuit_breaker.cpp
         src/metrics.cpp
         src/json.cpp
@@ -1680,6 +1692,7 @@ if(INDEX_BUILD_TESTS)
         src/model_context.cpp
         src/constitution.cpp
         src/api_client.cpp
+        src/label_score.cpp
         src/circuit_breaker.cpp
         src/metrics.cpp
         src/json.cpp
@@ -1707,6 +1720,7 @@ if(INDEX_BUILD_TESTS)
         src/model_context.cpp
         src/constitution.cpp
         src/api_client.cpp
+        src/label_score.cpp
         src/circuit_breaker.cpp
         src/metrics.cpp
         src/json.cpp
@@ -1731,6 +1745,7 @@ if(INDEX_BUILD_TESTS)
         src/model_catalog.cpp
         src/model_context.cpp
         src/api_client.cpp
+        src/label_score.cpp
         src/circuit_breaker.cpp
         src/metrics.cpp
         src/json.cpp
@@ -1772,6 +1787,7 @@ if(INDEX_BUILD_TESTS)
         src/advisor_gate.cpp
         src/advisor.cpp
         src/api_client.cpp
+        src/label_score.cpp
         src/circuit_breaker.cpp
         src/metrics.cpp
         src/json.cpp
@@ -1794,6 +1810,7 @@ if(INDEX_BUILD_TESTS)
         tests/test_presence.cpp
         src/presence.cpp
         src/api_client.cpp
+        src/label_score.cpp
         src/circuit_breaker.cpp
         src/metrics.cpp
         src/json.cpp
@@ -1815,13 +1832,28 @@ if(INDEX_BUILD_TESTS)
     add_executable(unit_intent
         tests/test_intent.cpp
         src/intent.cpp
+        src/label_score.cpp
+        src/json.cpp
     )
     target_include_directories(unit_intent PRIVATE
         ${CMAKE_SOURCE_DIR}/include
         ${CMAKE_SOURCE_DIR}/third_party/doctest
     )
     add_test(NAME unit_intent COMMAND unit_intent)
 
+    # Closed-set label distribution: renormalize, missing-label fail-open,
+    # OpenAI logprob body parse, margin floors. No provider client.
+    add_executable(unit_label_score
+        tests/test_label_score.cpp
+        src/label_score.cpp
+        src/json.cpp
+    )
+    target_include_directories(unit_label_score PRIVATE
+        ${CMAKE_SOURCE_DIR}/include
+        ${CMAKE_SOURCE_DIR}/third_party/doctest
+    )
+    add_test(NAME unit_label_score COMMAND unit_label_score)
+
     # Intent reconcile: admit, contract compile, observe ΔS, JIT ensure
     # waves, verification mandate, implement stub, snapshot rollback.
     # No provider client.

ROADMAP.md

@@ -66,6 +66,7 @@ calendar commitments.
 - [ ] **Delegation policies-** Consitutions declare max depth, allowed callees, budget caps (tokens/$)
 - [ ] ~~**Workflow recipes-** Checked-in “crews” (JSON): ordered/parallel graphs of agents + shared todo board~~
 - [ ] **Advisor policy packs-** Reusable gate profiles (strict / coding / research)
+- [ ] **Decision-filter measurement-** Refit `ARBITER_PRESENCE_SILENCE_MARGIN` and `ARBITER_INTENT_ROUTE_MARGIN` from `filter_*` / `decision_*` traces. The advisor gate stays off this path until those floors exist.
 
 **Acceptance criteria:** 
 - [x] A user can watch and steer a 5-agent job.

docs/api/intent.md

@@ -55,7 +55,8 @@ curl http://127.0.0.1:8080/v1/intent \
 | Field | Meaning |
 |-------|---------|
 | `kind` | Closed taxonomy: `research` \| `review` \| `write` \| `ops` \| `frontend` \| `backend` \| `plan` \| `market` \| `social` \| `multi` \| `unknown`. |
-| `source` | `heuristic` \| `llm` \| `explicit` \| `event` \| `none`. |
+| `source` | `heuristic` \| `llm` \| `decision` \| `explicit` \| `event` \| `none`. `decision` is a peaked local label distribution on a no-cue utterance. It carries the canned brief and no seeds. |
+| `decision_peak`, `decision_margin`, `decision_label` | Present when the local filter ran, including when it fell through to the LLM. Omitted when the filter is off or the heuristic already routed. |
 | `target_agent` | Suggested specialist, or `""` if index should handle it. |
 | `applied` | True when `apply_routing` would rewrite an `index` ingress to `target_agent` **and** that id is in the roster used for this request. This endpoint never dispatches. [`POST /v1/orchestrate`](orchestrate.md) additionally requires the agent to be loaded on the request orch (API: tenant catalog; TUI: `agents_dir`). |
 | `todo_seeds` / `plan_seeds` | Optional decomposition hints. Not persisted. |

docs/cli/environment.md

@@ -37,6 +37,9 @@ See [`docs/cli/connect.md`](connect.md). Local provider keys (`OPENROUTER_API_KE
 | `ARBITER_CIRCUIT_COOLDOWN_SECONDS`  | Cooldown in Open before a half-open probe. Default `30` (clamped ≥0). |
 | `ARBITER_INTENT_LLM_RATE_PER_MIN` | Extra per-tenant token bucket for `POST /v1/intent` hybrid/llm classify. Default `20`. `0` disables. Independent of `ARBITER_TENANT_RATE_PER_MIN`. See [Operations → Per-tenant rate / concurrency limiting](../concepts/operations.md#per-tenant-rate--concurrency-limiting). |
 | `ARBITER_INTENT_LLM_RATE_BURST` | Burst for the intent LLM bucket. Default `5`. |
+| `ARBITER_DECISION_MODEL` | Local `ollama/…` id for the closed-set label filter. Unset, empty, or any non-`ollama/` id leaves the filter off. Hosted models are not called. |
+| `ARBITER_PRESENCE_SILENCE_MARGIN` | Margin (peak minus runner-up) required before an extreme `silent` skips a presence review. `0` or unset disables that skip. Clamped to `[0, 1]`. A skip also requires peak ≥ 0.9. See [Presence](../concepts/presence.md#silence-filter). |
+| `ARBITER_INTENT_ROUTE_MARGIN` | Margin required before a no-cue utterance routes from the label filter without the advisor completion. `0` or unset disables that route. Clamped to `[0, 1]`. A route also requires peak ≥ 0.9. See [Intent](../concepts/intent.md#hybrid-classify). |
 
 ## TUI session durability
 

docs/concepts/intent.md

@@ -19,10 +19,14 @@ The executor still speaks [writs](writ.md). Intent never becomes a writ.
 ```text
 explicit non-index agent  → source=explicit (no reroute)
 heuristic unique match    → source=heuristic, maybe reroute
+else if no cue and a local label distribution is peaked
+                          → source=decision, canned brief, no seeds
 else if mode=hybrid|llm   → one LLM call, tagged <intent> reply
 else                      → fall through to index
 ```
 
+The decision step is off unless `ARBITER_DECISION_MODEL` is an `ollama/…` id and `ARBITER_INTENT_ROUTE_MARGIN` is greater than 0. It runs only when the heuristic kind is `unknown` (no cue hit). A multi-cue result, a confident cue, and `mode=heuristic` never consult it. The model scores single-letter labels, including `unknown`, and the runtime renormalizes over those letters alone. A route is taken only when the peak is at least 0.9, the margin (peak minus the runner-up) clears the env floor, the winner is a real kind, and that kind maps to an agent on the roster. The brief is the canned `Route to <agent>` line. Todo and phase seeds still come only from the tagged completion, which runs for every other case. A missing label in the provider's top logprobs, a flat distribution, or a transport error leaves routing on the existing path. The margin is a local gate to refit from `decision_peak` / `decision_margin` / `decision_label` on the intent event. It is not a calibrated probability. See [Environment](../cli/environment.md).
+
 Closed `kind` values: `research`, `review`, `write`, `ops`, `frontend`, `backend`, `plan`, `market`, `social`, `multi`, `unknown`. They map onto the personal starter callsigns under `agents/*.json` (`scout`, `vera`, `quill`, `forge`, `loom`, `nexus`, `compass`, `beacon`, `echo`). `multi` means "needs decomposition"; seed slots may be filled.
 
 LLM replies use a tagged grammar (same idea as the [advisor gate](advisor.md)):

docs/concepts/presence.md

@@ -67,6 +67,12 @@ The review is history-less — one snapshot in, one signal out — matching the
 
 `SILENT` is the default. Surrounding prose is tolerated. Missing `<signal>`, an unknown token, or `CONTEXT` without `<note>` is **malformed Silent** (fail-open). Presence cannot `HALT` or `REDIRECT` — that remains the advisor's job.
 
+## Silence filter
+
+When `ARBITER_DECISION_MODEL` is an `ollama/…` id and `ARBITER_PRESENCE_SILENCE_MARGIN` is greater than 0, each checkpoint scores a short snapshot before `run_presence_review`. The snapshot is the first line of the task (240 bytes) plus tool names. Arguments, result bodies, recent assistant text, and watcher rules are not in it.
+
+The labels are `silent` and `context`. An extreme `silent` — peak at least 0.9 and margin at least the env floor — skips the review and emits `presence` with `kind=silent`, `detail=filter`, plus `filter_peak`, `filter_margin`, and `filter_label`. A peaked `context`, a flat distribution, a missing label, or a transport error still runs the review, which is the only place a note can be written. The filter cannot inject `[PRESENCE: …]`. The margin is a local gate to refit from those fields. It is not a calibrated probability. See [Environment](../cli/environment.md).
+
 ## Runtime control flow
 
 Where presence fires inside the dispatch loop:

docs/concepts/sse-events.md

@@ -7,7 +7,7 @@ Every event on the `/v1/orchestrate` stream has an `event:` line and a `data:` l
 | Event | When | Fields |
 |-------|------|--------|
 | `request_received` | Exactly once, first event on the stream. | `agent`, `tenant`, `tenant_id`, `message` (first 200 chars, ellipsis added if truncated), `channel?` (`"voice"` when the request asked for spoken output). |
-| `intent` | Pre-dispatch classify/route on **fresh** ingress (`original_query` omitted) when the ingress agent's `intent.mode` is not `off`. After `request_received`, before `stream_start`. See [Intent](intent.md). | `kind`, `confidence`, `source`, `target_agent`, `applied`, `requested_agent`, `applied_agent`, `brief?`, `todo_seed_count`, `plan_seed_count`, `llm_used?`, `malformed?`. |
+| `intent` | Pre-dispatch classify/route on **fresh** ingress (`original_query` omitted) when the ingress agent's `intent.mode` is not `off`. After `request_received`, before `stream_start`. See [Intent](intent.md). | `kind`, `confidence`, `source`, `target_agent`, `applied`, `requested_agent`, `applied_agent`, `brief?`, `todo_seed_count`, `plan_seed_count`, `llm_used?`, `malformed?`, `decision_peak?`, `decision_margin?`, `decision_label?`. |
 | `reconcile.progress` | Phase note on [`POST /v1/reconcile`](../api/reconcile.md). | `request_id`, `phase`, `detail?`. |
 | `reconcile.delta` | Contract observation: residual vs held clauses. Emitted after the initial observe and after every ensure wave. | `residual`, `held`, `empty`, `wave`. |
 | `agent.spawned` | JIT ephemeral clone created for a ΔS cover (`mode=ensure` waves). | `agent`, `clone_id`, `clauses`. |
@@ -23,7 +23,7 @@ Every event on the `/v1/orchestrate` stream has an `event:` line and a `data:` l
 | `sub_agent_response` | After a delegated turn completes (depth > 0). The full turn body in one payload — useful for consumers that don't want to reconstruct from deltas. | `agent`, `stream_id`, `depth`, `content`. |
 | `token_usage` | After each turn completes. | `agent`, `stream_id`, `depth`, `model`, `input_tokens`, `output_tokens`, `cache_read_tokens?`, `cache_create_tokens?`. |
 | `advisor` | Every advisor interaction. Independent of `tool_call` — fires for the runtime's gate decisions (which the executor never invokes directly) and for the executor's own `/advise` consults. See [Advisor concept](advisor.md). | `agent`, `stream_id`, `kind`, `detail?`, `preview?`, `malformed?`. |
-| `presence` | Always-on watcher consult after a peer's tool batch. Fires for `SILENT` and `CONTEXT`. See [Presence](presence.md). | `watcher`, `agent` (the working agent), `stream_id`, `kind` (`silent` \| `context`), `detail?`, `malformed?`. |
+| `presence` | Always-on watcher consult after a peer's tool batch. Fires for `SILENT` and `CONTEXT`. See [Presence](presence.md). | `watcher`, `agent` (the working agent), `stream_id`, `kind` (`silent` \| `context`), `detail?`, `malformed?`, `filter_action?`, `filter_peak?`, `filter_margin?`, `filter_label?`. |
 | `escalation` | Out-of-band advisor halt. Fires before the corresponding `stream_end` (which arrives with `ok: false`). Only fires at the originating depth — sub-agent halts bubble up via the parent's response, not via duplicate escalation events. | `agent`, `stream_id`, `reason`. |
 | `stream_end` | Closes each turn. Any remaining partial line is flushed before this fires, so no `text` events arrive with this `stream_id` after. | `agent`, `stream_id`, `ok`. |
 | `error` | Recoverable errors during the request (transient upstream issue). The stream continues or terminates depending on severity. | `message`, plus optional context fields (e.g. `reason`). |

include/api_client.h

@@ -132,6 +132,11 @@ struct ApiRequest {
     int max_tokens = 1024;
     double temperature = 0.3;
     bool include_temperature = true;
+    // Decision-filter only. Ollama's OpenAI-compatible endpoint. Left
+    // false on every generative call so older Ollama builds never see
+    // the fields.
+    bool logprobs = false;
+    int top_logprobs = 0;
 };
 
 struct ApiResponse {

include/intent.h

@@ -16,6 +16,8 @@
 // Distinct from Constitution::MemoryConfig::intent_routing, which only
 // boosts /mem search types.
 
+#include "label_score.h"
+
 #include <cstddef>
 #include <functional>
 #include <string>
@@ -65,13 +67,20 @@ struct Intent {
     // Closed taxonomy aligned with starter agents, plus multi/unknown.
     std::string kind;            // research|review|write|ops|frontend|backend|plan|market|social|multi|unknown
     double      confidence = 0;  // 0..1
-    std::string source;          // heuristic|llm|explicit|event|none
+    std::string source;          // heuristic|llm|decision|explicit|event|none
     std::string target_agent;    // empty = do not reroute
     std::string brief;
     std::vector<IntentSeedTodo>  todo_seeds;
     std::vector<IntentSeedPhase> plan_seeds;
     bool llm_used = false;
     bool malformed = false;
+    // Set when the local label scorer ran on a no-cue utterance. peak and
+    // margin are the renormalized distribution, even when the route was
+    // not taken and source stayed heuristic/llm.
+    bool decision_consulted = false;
+    double decision_peak = 0;
+    double decision_margin = 0;
+    std::string decision_label;
 };
 
 struct IntentInput {
@@ -105,11 +114,15 @@ bool intent_should_apply(const IntentConfig& cfg,
                          bool fresh_ingress = true);
 
 // Hybrid classify: explicit specialist short-circuits; heuristic when
-// confident; LLM when mode is hybrid/llm and heuristic is below threshold.
-// llm may be null — hybrid then degrades to heuristic-only.
+// confident; a peaked local label distribution when the heuristic found
+// no cue and `decision_margin` > 0; otherwise the LLM. llm and scorer
+// may be null — hybrid then degrades to heuristic-only. decision_margin
+// <= 0 leaves the scorer unconsulted.
 Intent resolve_intent(const IntentInput& in,
                       const IntentConfig& cfg,
-                      const IntentLlmFn& llm = nullptr);
+                      const IntentLlmFn& llm = nullptr,
+                      const LabelScorer& scorer = nullptr,
+                      double decision_margin = 0);
 
 // One-line (plus optional GOAL) preamble injected ahead of the user text.
 // Empty when there is nothing useful to show the executor.