From ad9ee2962413581112084971d2165ceaa6dd985e Mon Sep 17 00:00:00 2001
From: kert <kert@rack.localdomain>
Date: Wed, 11 Nov 2026 15:40:12 -0500
Subject: [PATCH] feat(llm): tagger shortlists 12 themes with a 0.45 floor
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Golden v2 (200 comments, 8 dockets) showed recall, not precision, was
the problem: stored tags scored P 0.96 R 0.55 on the held-out half.
Live tune-split sweep on the rig: 14b/top 8/0.60 F1 0.72; /0.50 0.76;
top 12/0.45 0.78; qwen2.5:32b was more conservative (top 8/0.50: P 0.97
R 0.59 F1 0.73). Held-out with top 12/0.45: P 0.95 R 0.61 F1 0.74 —
PFS dockets 0.67-0.91, IPPS/OPPS ~0.28 (topics outside vocab v2).
---
 stack.toml               | 4 ++--
 tests/llm/test_config.py | 4 ++--
 2 files changed, 4 insertions(+), 4 deletions(-)

diff --git a/stack.toml b/stack.toml
index 6bf51aa..15aeb93 100644
--- a/stack.toml
+++ b/stack.toml
@@ -114,9 +114,9 @@ instruct_model = "qwen2.5:14b"         # baseline, pulled on every fleet host
 instruct_model_large = "qwen2.5:32b"   # used when the chosen host declares >= large_min_vram_gb
 large_min_vram_gb = 20
 chat_num_ctx = 8192                    # passed per request; no Modelfile ctx variants
-tag_top = 8                            # theme tagger: candidates judged per comment (P35)
+tag_top = 12                           # theme tagger: candidates judged per comment (P35; golden v2 sweep 2026-11-11)
 tag_max_tags = 3                       # most themes written per comment
-tag_min_confidence = 0.60              # shortlist similarity a judged yes must reach; gated with eval-tags
+tag_min_confidence = 0.45              # shortlist similarity a judged yes must reach (v2 tune: F1 0.72 at 0.60 → 0.78)
 chat_temperature = 0.2                 # grounded answers cite labels verbatim; low temperature keeps them literal
 chat_max_tokens = 1200                 # num_predict: a history answer ran to 5,000 tokens (3 min) without a cap
 chat_chars_per_token = 2.4             # measured 2.5–2.9 on FR/regulatory text; sizes the excerpt budget so prompt + reply fit chat_num_ctx (overflow is caught and rebuilt)
diff --git a/tests/llm/test_config.py b/tests/llm/test_config.py
index ebf401b..e367e94 100644
--- a/tests/llm/test_config.py
+++ b/tests/llm/test_config.py
@@ -161,6 +161,6 @@ def test_tagging_parameters_load_from_config():
     from llm.config import load
 
     cfg = load()
-    assert cfg.tag_top == 8
+    assert cfg.tag_top == 12
     assert cfg.tag_max_tags == 3
-    assert cfg.tag_min_confidence == 0.60
+    assert cfg.tag_min_confidence == 0.45