=== first first.py: another message: two more towels choice: housekeeping confidence: 0.7779 0.7779 housekeeping 0.0090 maintenance 0.2131 front desk === by_hand_1 by_hand.py: TEMPERATURE = 1 Context: The air conditioner in room 12 has been dripping water onto the bed since last night. Please send someone before 2 pm. I fly home tomorrow. Question: Which team should handle this message? Options: (A) housekeeping (B) maintenance (C) front desk Answer: ( 66 tokens were read; 248320 tokens have a logit the five highest: ['A', 'B', 'C', 'D', 'F'] (A) logit 13.5235 probability 0.5113 housekeeping (B) logit 13.1665 probability 0.3578 maintenance (C) logit 12.1614 probability 0.1309 front desk === by_hand_2 by_hand.py: TEMPERATURE = 0.1 Context: The air conditioner in room 12 has been dripping water onto the bed since last night. Please send someone before 2 pm. I fly home tomorrow. Question: Which team should handle this message? Options: (A) housekeeping (B) maintenance (C) front desk Answer: ( 66 tokens were read; 248320 tokens have a logit the five highest: ['A', 'B', 'C', 'D', 'F'] (A) logit 13.5235 probability 0.9726 housekeeping (B) logit 13.1665 probability 0.0274 maintenance (C) logit 12.1614 probability 0.0000 front desk === choice choice.py: no meanings: every option is None { "model": "decider-2b-v11", "answers": { "team": { "type": "choice", "choice": "housekeeping", "confidence": 0.3328, "x_p_max": 0.4996, "certainty": 0.2386, "probabilities": { "housekeeping": 0.4996, "maintenance": 0.3779, "front_desk": 0.0691, "other": 0.0534 } } }, "usage": { "input_tokens": 71, "output_tokens": 0 } } === noul noul.py: the guest also asks for money back { "refund": { "type": "noul", "noul": 0.8819 }, "today": { "type": "noul", "noul": 0.7527 } } === score score.py: another message: a late check-out next Friday { "urgent": { "type": "score", "score": 0.97, "confidence": 0.0, "x_p_max": 0.3606, "certainty": 0.0016, "legend": { "0": "It can wait a few days", "1": "It should be done today", "2": "It must be done within a few hours" }, "probabilities": { "0": 0.3606, "1": 0.3137, "2": 0.3257 }, "level_fit": { "0": 0.1685, "1": 0.1465, "2": 0.1522 }, "fit_mass": 0.4672 } } === one_call one_call.py: the urgent question is taken out team choice maintenance refund noul 0.0232 {'input_tokens': 169, 'output_tokens': 0} === look_inside look_inside.py: one more line: the signature of the class decide (self, context, questions, **kw) decide_batch (self, requests, max_ctx_tokens=1536) decide_json (self, context, schema, **kw) system_one (self, state, questions, independent=True, max_state_tokens=32768, max_fwd_tokens=65536, layout=None, isolated=None) Jev-shaped requests on top of the decider prompt format (same wire format as TypeSafe's POST /v1/systemone). state str | dict | list JSON state is serialised compactly; questions may name a part by path (`ticket.messages[0].text`) questions {id: {"type": "choice", "instructions": ..., "criteria": {name: description | {...} | [...] | None}} up to 255 options {"type": "score", "instructions": ..., "criteria": [level 0 description, level 1 description, ...]} 2..10 levels {"type": "noul", "instructions": ... (optional), "criteria": {"true": ..., "false": ...} (optional)}} a noul question needs instructions or at least one true/false description. ids are never shown to the model. `instructions` and every description may be a string or any JSON value. (path, device=None, dtype=None, temperature=None, abstain_below=0.0, use_graphs=None, temperature_by_type=None, gguf_file=None, gguf_options=None) === fields fields.py: the legend as a map with the numbers 0, 5 and 10 { "Which team should handle this message?": { "choice": "maintenance: repairs, air conditioner, water, electricity", "confidence": 0.8762, "type": "choice", "probabilities": { "housekeeping: cleaning, towels, bed linen": 0.1014, "maintenance: repairs, air conditioner, water, electricity": 0.8762, "front desk: bookings, check-out, bills, questions": 0.0225 } }, "Does the guest ask for money back?": { "noul": 0.016, "type": "noul" }, "How soon does the guesthouse have to act?": { "score": 9.15, "confidence": 0.8357, "type": "scale", "legend": { "0": "it can wait a few days", "5": "it should be done today", "10": "it must be done within a few hours" }, "probabilities": { "0": 0.0049, "5": 0.1594, "10": 0.8357 } } } === many many.py: abstain_below=0.9 0.934 housekeeping Can we have two more towels, please? 0.888 (not sure) The wifi password on the card does not work. 0.955 front desk What time is breakfast? 0.991 front desk My bill shows three nights but I stayed two. 0.932 maintenance The shower has no hot water. 0.420 (not sure) There is a gecko on my wall and I love it. === compare compare.py: BANK = 5.0 river truth numbers only with the words 3.2 no 0.0937 0.0243 3.9 no 0.0963 0.0244 2.1 no 0.0865 0.0241 3.75 no 0.1079 0.0259 5.3 yes 0.1857 0.7916 3.65 no 0.1073 0.0256 4.0 no 0.1143 0.0251 1.2 no 0.0888 0.0256 3.71 no 0.1090 0.0252 3.0 no 0.1048 0.0254 6.8 yes 0.1996 0.7974 3.69 no 0.1136 0.0254 right with the numbers only: 10 of 12 right with the words: 12 of 12 === too_much too_much.py: a rule about breakfast in place of the refund rule 0.0304 the message alone 0.0196 the booking record 0.0173 the record with the refund rule in it 0.0176 the same, and the question names the message === long_text long_text.py: 50 rules in place of 150 967 tokens in the text 0.4693 maintenance decide(): it reads the first 1,536 tokens 0.4693 maintenance decide() with max_ctx_tokens=8192 0.8478 maintenance decide(), with the message first and the rules after it 0.4693 maintenance system_one(), the same long text