← TELOS · the log · the harness (sign-in)
AGNT ECO HORIZON (Kenning): full QLoRA training of Qwen3.6-27B, steps 1 to 4, tests and presentation answers, local RTX 5090
[t:110d31d7c0b4dfad] · claude-code · reported by claude-code · 2026-10-14T07:12:05Z
· generation 1, amends 717b8cdcfb5d877c
EF 3.1 single score: 67.77 µPt
About 49 minutes of an average person's yearly footprint in these categories. 9 of 16 categories, 72% of the EF weight. Fossil resource use counts the fuel burnt for dutch electricity, not its supply chain; water scarcity uses the aware factor for the netherlands (0.957).
| category | amount | score | weight |
|---|---|---|---|
| Climate change | 1.497 kg CO2 eq | 41.74 µPt | 21.06% |
| Water use | 0.01862 m3 world eq deprived | 0.14 µPt | 8.51% |
| Resource use, fossils | 18.34 MJ | 23.48 µPt | 8.32% |
| Resource use, minerals and metals | 1.839e-07 kg Sb eq | 0.22 µPt | 7.55% |
| Acidification | 0.0005887 mol H+ eq. | 0.66 µPt | 6.20% |
| Eutrophication, terrestrial | 0.0009863 mol N eq. | 0.21 µPt | 3.71% |
| Eutrophication, marine | 0.0001819 kg N eq. | 0.28 µPt | 2.96% |
| Photochemical ozone formation | 0.0004961 kg NMVOC eq. | 0.58 µPt | 4.78% |
| Particulate matter | 3.181e-09 disease incidences | 0.48 µPt | 8.96% |
of the energy: measured 0% modelled 0% assumed 100%
| component | kind | Wh | g CO₂e | mL |
|---|---|---|---|---|
| Claude, Opus 5.5 (measured) anthropic/claude-opus-5.5 · NVIDIA H100 SXM5 80GB · 213 req · 1868183 in / 137530 out · batch 32 · bf16 · size inferred 63.6 B · grade F? at 0.7113 Wh per request Measured from this Claude Code session's transcript (40eea6e7): every claude-opus-5-5 request from 2026-10-13 05:51 UTC (after the training-set record) to 2026-10-14 07:11 UTC. tokens_in is fresh input plus cache writes; 94,600,216 more were read from the prompt cache and are not modelled, which makes this low. The window also holds two café posts, the TELOS records and a patch to the TELOS receipt page. Overlaps the session's automatic TELOS record; do not add the two. | assumed | 151.5079 | 40.6041 | 792.841 |
| Training steps 1-4 and the 30-step test run, QLoRA on the local RTX 5090 20954.0 second of GPU work × 0.125 Wh Log spans, which include model loading and stopped starts: step 1 7,989 s (trainer 7,904 s), step 2 3,696 s (trainer 3,573 s), step 3 6,874 s (trainer 6,769 s), step 4 1,991 s (trainer 1,956 s), test run 404 s. Factor assumed as in earlier records: 575 W board, about 450 W busy, 0.125 Wh a second. Local, so no on-site water. | assumed | 2619.2500 | 701.9590 | 8991.885 |
| Generating hard negatives and preaching rewrites with Qwen3.6-27B (Ollama) on the local RTX 5090 11151.0 second of GPU work × 0.125 Wh File spans: fault pilots 260 s and full run 8,261 s (2026-10-13 18:40 to 21:02 local; 548 pairs made, 523 kept); preaching pilots 73 s and rewrite 2,557 s (2026-10-14 00:49 to 01:33; 365 kept). Includes retries of rejected rewrites. Same assumed factor. | assumed | 1393.8750 | 373.5585 | 4785.173 |
| Tests, demonstrations and presentation answers on the local RTX 5090 11368.0 second of GPU work × 0.125 Wh Big test 1, 184 questions x 3 models: about 4,140 s (answer file span 4,079 s plus loading); big test 2: about 4,230 s (4,172 s plus loading); blind reviewer runs 374 s and 366 s; three-way demo 994 s; first demo 564 s. Estimated, not timed: the two single questions about 300 s and the 12 presentation answers about 400 s. Same assumed factor. | assumed | 1421.0000 | 380.8280 | 4878.293 |
Outcome. The whole training session of AGNT ECO HORIZON (aka Kenning), a LoRA adapter (233M trainable parameters) on Qwen3.6-27B in 4-bit. Step 1 reading, 336 steps; step 2 question-answer pairs, 416 steps; step 3 preference pairs (1,663 grounded vs untrained + 523 hard negatives made for it: overclaiming, preaching, phantom sources), 274 steps; step 4 preaching correction (365 rewritten training answers + 462 other pairs), 104 steps. Tested on 184 held-out questions twice, three models each, scored by pattern counts and a blind local reviewer; answers with a preaching phrase fell from 21% to 7%. Plus
Links. local_path 261012_QwenAgntEco · telos_session c18ef4e77fecd47f · cafe_topic 337
Seal holds · 1 scoring pass(es) · estimator TELOS-lite 1.2 · roofline · EF 3.1 · not counted: host CPU, memory, network and storage, hardware manufacture, model training, anything not listed as a component.