{"days":30,"runs":[{"day":"2026-11-04","group":"model:pricing","objective":"our USD per 1M input tokens per model class stays undercut_pct below the cheapest comparable hosted API (every competitor row in competitors.yaml, re-read daily","base":15.0,"best":0.0,"candidates":2,"applied":0,"seconds":6.4,"sample":"adoption funnel, 7-day paid calls","error":false},{"day":"2026-11-04","group":"model:tuatara-decide","objective":"accuracy of the real trained classifier (vector + TF-IDF blend + embedding head) on a dev slice of BANKING77 train; embed_head.C under a full training, embed_de","base":0.911,"best":0.914,"candidates":2,"applied":1,"seconds":1801.65,"sample":"BANKING77 train dev slice n=1000 seed=20261104; full + fewshot(5/intent)","error":false},{"day":"2026-11-04","group":"model:tuatara-summarize-1.0","objective":"ROUGE-1 F1 on a seeded CNN/DailyMail VALIDATION sample (3 sentences), each node scored under its own preset, higher is better; a candidate that is more than 50 ","base":0.422218,"best":0.402536,"candidates":11,"applied":1,"seconds":657.1,"sample":"abisee/cnn_dailymail 3.0.0 validation n=150 seed=20261104","error":false},{"day":"2026-11-04","group":"model:tuatara-topics-1.0","objective":"sum of c_v coherence, topic diversity, purity and NMI on a tuning slice of 20NG train (K=50); the published test split is scored once, afterwards","base":2.460307,"best":2.507675,"candidates":5,"applied":1,"seconds":2404.73,"sample":"20NG train tuning slice 20% seed=20261104 K=50 iters=200","error":false},{"day":"2026-11-03","group":"model:pricing","objective":"our USD per 1M input tokens per model class stays undercut_pct below the cheapest comparable hosted API (every competitor row in competitors.yaml, re-read daily","base":6.0,"best":0.0,"candidates":2,"applied":0,"seconds":21.79,"sample":"adoption funnel, 7-day paid calls","error":false},{"day":"2026-11-03","group":"model:tuatara-decide","objective":"accuracy of the real trained classifier (vector + TF-IDF blend + embedding head) on a dev slice of BANKING77 train; embed_head.C under a full training, embed_de","base":0.912,"best":0.836,"candidates":4,"applied":1,"seconds":5593.64,"sample":"BANKING77 train dev slice n=1000 seed=20261103; full + fewshot(5/intent)","error":false},{"day":"2026-11-03","group":"model:tuatara-summarize-1.0","objective":"ROUGE-1 F1 on a seeded CNN/DailyMail VALIDATION sample (3 sentences), each node scored under its own preset, higher is better; a candidate that is more than 50 ","base":0.401688,"best":0.404045,"candidates":10,"applied":1,"seconds":1204.02,"sample":"abisee/cnn_dailymail 3.0.0 validation n=150 seed=20261103","error":false},{"day":"2026-11-03","group":"model:tuatara-topics-1.0","objective":"sum of c_v coherence, topic diversity, purity and NMI on a tuning slice of 20NG train (K=50); the published test split is scored once, afterwards","base":2.366062,"best":2.411867,"candidates":2,"applied":1,"seconds":2470.46,"sample":"20NG train tuning slice 20% seed=20261103 K=50 iters=200","error":false},{"day":"2026-11-02","group":"model:pricing","objective":"our USD per 1M input tokens per model class stays undercut_pct below the cheapest comparable hosted API (every competitor row in competitors.yaml, re-read daily","base":4.0,"best":0.0,"candidates":2,"applied":0,"seconds":7.63,"sample":"adoption funnel, 7-day paid calls","error":false},{"day":"2026-11-02","group":"model:tuatara-decide","objective":"accuracy of the real trained classifier (vector + TF-IDF blend + embedding head) on a dev slice of BANKING77 train; embed_head.C under a full training, embed_de","base":0.899,"best":0.899,"candidates":0,"applied":0,"seconds":2518.66,"sample":"BANKING77 train dev slice n=1000 seed=20261102; full + fewshot(5/intent)","error":false},{"day":"2026-11-02","group":"model:tuatara-summarize-1.0","objective":"ROUGE-1 F1 on a seeded CNN/DailyMail VALIDATION sample (3 sentences), each node scored under its own preset, higher is better; a candidate that is more than 50 ","base":0.406885,"best":0.406885,"candidates":9,"applied":0,"seconds":378.96,"sample":"","error":false},{"day":"2026-11-02","group":"model:tuatara-topics-1.0","objective":"sum of c_v coherence, topic diversity, purity and NMI on a tuning slice of 20NG train (K=50); the published test split is scored once, afterwards","base":2.357655,"best":2.436065,"candidates":7,"applied":1,"seconds":2705.2,"sample":"20NG train tuning slice 20% seed=20261102 K=50 iters=200","error":false}],"changes":[{"day":"2026-11-04","group":"model:tuatara-decide","param":"embed_head.C","old":"10.0","new":"7.5","base_objective":0.911,"candidate_objective":0.914,"guard":null,"status":"applied","measured_delta":null,"measured_at":null,"source":"step"},{"day":"2026-11-04","group":"model:tuatara-topics-1.0","param":"unique","old":"0.2","new":"0.4","base_objective":2.460307,"candidate_objective":2.507675,"guard":null,"status":"applied","measured_delta":null,"measured_at":null,"source":"step"},{"day":"2026-11-04","group":"model:tuatara-summarize-1.0","param":"tuned.length_norm","old":"1.0","new":"1.5","base_objective":0.400944,"candidate_objective":0.402536,"guard":null,"status":"applied","measured_delta":null,"measured_at":null,"source":"step"},{"day":"2026-11-03","group":"model:tuatara-decide","param":"embed_default_weight","old":"1.0","new":"0.9","base_objective":0.833,"candidate_objective":0.836,"guard":null,"status":"reverted","measured_delta":-0.003,"measured_at":"2026-11-04T07:35:37.293308+00:00","source":"step"},{"day":"2026-11-03","group":"model:tuatara-topics-1.0","param":"unique","old":"0.1","new":"0.2","base_objective":2.366062,"candidate_objective":2.411867,"guard":null,"status":"applied","measured_delta":0.050431,"measured_at":"2026-11-04T06:51:40.502212+00:00","source":"advisor"},{"day":"2026-11-03","group":"model:tuatara-summarize-1.0","param":"tuned-best.length_norm","old":"1.0","new":"0.75","base_objective":0.401688,"candidate_objective":0.404045,"guard":null,"status":"applied","measured_delta":0.000791,"measured_at":"2026-11-04T06:33:04.541026+00:00","source":"step"},{"day":"2026-11-02","group":"model:tuatara-topics-1.0","param":"unique","old":"0.0","new":"0.1","base_objective":2.357655,"candidate_objective":2.436065,"guard":null,"status":"applied","measured_delta":0.053647,"measured_at":"2026-11-03T07:15:23.697444+00:00","source":"step"}],"scored":58,"last_run":"2026-11-04"}