eval-harness

arm

triage_local_dgx_phi4_14b_v1

One config on orrery_triage_v1_labelled, scored one way; its repeats averaged per item.

dataset
orrery_triage_v1_labelled
scored against
gold/orrery_triage_v1_labelled gold
goal
Triage decisions on Orrery's error signals
runs
1
headline
triage_cost 0.662
last run
2026-10-31 08:33
re-scored
the goals read its re-score, runs-rescored/5f2c057cdaa5/runs/orrery_triage_v1_labelled/triage_local_dgx_phi4_14b_v1

Inspect itWith the 317 arms scored the same way

01metrics

metricmeanitemsbetter
triage_cost0.66265lower
correct0.50865higher
cost_usd$0.00000065lower
escalated0.30865lower
false_alarm0.74231lower
false_dismiss0.11834lower
latency_ms16973.165lower
model_correct0.50865higher
model_escalated0.23165lower
model_false_alarm0.67731lower
model_false_dismiss0.17634lower
model_triage_cost0.78565lower
n_probes2.87765higher
no_decision0.00065lower
table_miss0.00065lower
tokens_in5700.865lower
tokens_out363.365lower
triage_cost_p100.96965lower
triage_cost_p30.53865lower

02cost and speed

cost per item
$0.000000 · 6,064 tok
speed per item
p50 17,140 · p95 23,284 ms

Per item, over its items: cost as dollars and tokens (a local model is free but still uses tokens), speed as the median and the slow tail (p95). A goal that names this arm weighs them in its decision.

03runs

runmadeitemscosthow
triage_local_dgx_phi4_14b_v1_20261031T081434Z2026-10-31 08:3365$0.000000measured

04config

config triage_local_dgx_phi4_14b_v1 · adapter examples/triage-fleet/adapter.py · credentials are never shown

call_timeout_s
600
machine
dgx
model
phi4:14b
provider
gateway
url
http://100.69.49.126:11434/v1/chat/completions
usd_per_mtok_in
0
usd_per_mtok_out
0

05conditions

data.items_sha256
d428aca49adc8e762dcbbcad6c9a1dba1ac4e4e3315ff26025af975faba1176a
deployment.machine
dgx
deployment.model
phi4:14b
deployment.source
adapter
fingerprint.version
5
harness.code_sha256
0ab386969208f46c7448153422d5a5ab07833d963869b69fc216a6cd28feb110
harness.version
0.5.0
library.anthropic
2.3.0
library.numpy
2.5.4
library.openai
3.41.0
library.rouge-score
0.1.2
machine.cpu_count
32
machine.platform
Linux/x86_64
machine.python
3.14.7
machine.release
7.2.5-3-omarchy
reference.sha256
f617ad3bb9bb9f5feb8ffe54e0ef1b9b05dfb316fac13f63a925f9d35fb191cd
scorer.id
examples/triage-fleet/adapter.py
scorer.sha256
663c89dd6a24c414ce4229c5b067581cced2becf0a44e4660847dce8420c564a

06the same config, scored other ways

scored against silverscored by a judgewrote the reference it is scored againstscored against a label set, not a reference setunmarked rows are scored against gold

07rounds