schema_only - flag_everything = -114.85, 95% CI [-119.77, -110.26]. Synthetic, not production-validated. https://github.com/caiotheodoro/proctor
Caio Theodoro
caiotheodoro
·
AI & ML interests
None yet
Recent Activity
updated a collection 3 days ago
Proctor: oracle-priced agent actions updated a collection 3 days ago
Proctor: oracle-priced agent actions updated a model 3 days ago
caiotheodoro/proctor-qlora-grpoOrganizations
None yet
Dense optical flow asked for a hand's speed reports the background above a displacement knee. Synthetic, no corpus or GPU needed.
Five of six were my own harness. Three pre-registered hypotheses since falsified, two of them mine.
qwen3.7-plus beats qwen3.6-plus on accuracy and loses 3.7x more. Code: https://github.com/caiotheodoro/lossbench
Loses accuracy to DeepSeek v4-flash, wins severity-weighted recall 0.901 vs 0.872 at 1.000 on HIGH. Code: https://github.com/caiotheodoro/reconforge
Synthetic NFS-e invoices and the pre-registered pinned-DiffusionGemma study. Real invoices stay private.
28 judge errors on 153 human-labelled frames, 26 of them inflating. Writeup: caio.theodoro.dev/blog/vernier-judge-errors-run-one-way
Beats flag-everything by 274.0, 95% CI [186,326]. Synthetic, not production-validated. Code: https://github.com/caiotheodoro/assay
Self-proposed G702 training tasks lose alone, help as a supplement. Code: https://github.com/caiotheodoro/plumb
Binder vs issued policy, two page images in. 8B QLoRA, program oracle, 0.959 recall. Code: https://github.com/caiotheodoro/suture
schema_only - flag_everything = -114.85, 95% CI [-119.77, -110.26]. Synthetic, not production-validated. https://github.com/caiotheodoro/proctor
Synthetic NFS-e invoices and the pre-registered pinned-DiffusionGemma study. Real invoices stay private.
Dense optical flow asked for a hand's speed reports the background above a displacement knee. Synthetic, no corpus or GPU needed.
28 judge errors on 153 human-labelled frames, 26 of them inflating. Writeup: caio.theodoro.dev/blog/vernier-judge-errors-run-one-way
Five of six were my own harness. Three pre-registered hypotheses since falsified, two of them mine.
Beats flag-everything by 274.0, 95% CI [186,326]. Synthetic, not production-validated. Code: https://github.com/caiotheodoro/assay
qwen3.7-plus beats qwen3.6-plus on accuracy and loses 3.7x more. Code: https://github.com/caiotheodoro/lossbench
Self-proposed G702 training tasks lose alone, help as a supplement. Code: https://github.com/caiotheodoro/plumb
Loses accuracy to DeepSeek v4-flash, wins severity-weighted recall 0.901 vs 0.872 at 1.000 on HIGH. Code: https://github.com/caiotheodoro/reconforge
Binder vs issued policy, two page images in. 8B QLoRA, program oracle, 0.959 recall. Code: https://github.com/caiotheodoro/suture