Calibrate fit and taste per embedding model instead of min-max per run

scale() maps a cosine to clip((cos - lo) / (hi - lo), 0, 1) with constants
from CALIBRATION in app/config.py, keyed by model and replaceable under
embed.calibration. multilingual-e5-small: fit 0.75..0.88, taste -0.14..0.19
(taste from synthetic votes). An uncalibrated model falls back to min-max
and logs a warning once per run.

The extract and summarise bars default to 0.55 and 0.65, about 300 and 35
roles for the starter agent on the measured data; stored limits are kept.
An agent without an intent is left out of the model-stage scope, and each
score run names such agents in one score.no_intent notice.

Co-Authored-By: Claude Opus <noreply@anthropic.com>
This commit is contained in:
WolfandClaude Opus committed 2026-09-29 14:59:34 +02:00
1 parent 706b127bc0
commit b021838959
12 files changed
+336 -76

No files matched your search

+10 -6
View File
@@ -62,7 +62,7 @@ def ranked(store: Store, *items: Item) -> None:
"""Score the roles for one enabled agent, best first, so they are in the
top extract reads. Every role given must already be stored."""
agents = store.agents()
agent_id = agents[0].id if agents else store.save_agent("Backend").id
agent_id = agents[0].id if agents else store.save_agent("Backend", "python backend").id
store.replace_scores(agent_id, [(item.uid, 1.0 - n / 10, "")
for n, item in enumerate(items)])
@@ -171,8 +171,9 @@ def test_only_the_best_of_each_agent_above_the_floor_is_read(
uid finishes it."""
for sid in ("best", "second", "low", "unscored", "gated"):
store.upsert_role(posting(sid))
one = store.save_agent("One", limits=Limits(extract_top=1, extract_min=0.5))
two = store.save_agent("Two", enabled=False)
one = store.save_agent("One", "python backend",
limits=Limits(extract_top=1, extract_min=0.5))
two = store.save_agent("Two", "python backend", enabled=False)
store.replace_scores(one.id, [(posting("best").uid, 0.9, ""),
(posting("second").uid, 0.8, ""),
(posting("low").uid, 0.2, ""),
@@ -182,11 +183,14 @@ def test_only_the_best_of_each_agent_above_the_floor_is_read(
model = stub(monkeypatch, lambda roles: [answer(r["uid"]) for r in roles])
assert store.extract_pending() == 1
store.save_agent("One", agent_id=one.id, limits=Limits(extract_top=5, extract_min=0.5))
store.save_agent("One", "python backend", agent_id=one.id,
limits=Limits(extract_top=5, extract_min=0.5))
assert store.extract_pending() == 2, "the floor holds the low one back"
store.save_agent("One", agent_id=one.id, limits=Limits(extract_top=5, extract_min=0))
store.save_agent("One", "python backend", agent_id=one.id,
limits=Limits(extract_top=5, extract_min=0))
assert store.extract_pending() == 3, "the disabled agent's pick never counts"
store.save_agent("One", agent_id=one.id, limits=Limits(extract_top=1, extract_min=0.5))
store.save_agent("One", "python backend", agent_id=one.id,
limits=Limits(extract_top=1, extract_min=0.5))
modeltask.run(extract.TASK, config, store)
assert [[r["uid"] for r in batch] for batch in model.asked] == [[posting("best").uid]]