One Pass, No Prose

state     Subject: Checkout broken / Hi, every customer who tries to pay on our store has been getting a 502 at the payment step for the last hour. We are losing orders. Account: acme-retail (Scale plan).
questions
  team      billing · technical · sales · other
  urgent    yes · no
  severity  no impact · minor · major · critical

This sounds like a payment outage, so the technical team should own it. It is urgent because customers cannot check out, and I would call the severity major or critical. I'm about 90% confident on the routing.

…502 at the payment step for the last hour…  Which team should own this?
[slot] billing   [slot] technical   [slot] sales   [slot] other
[slot] urgent?   [slot] severity…

One ticket, three questions, every allowed answer listed. About 45 tokens, written one at a time. Each one is a separate pass through the model. Highlighted: what your code has to dig out, a hedge, and a confidence that is only text. Each [slot] is read by the scoring head. Softmax runs over each question's options, not the whole vocabulary. Jev, one request, 432 ms round trip. Three answers, each with a full distribution. Base Laya, 43 ms on a laptop. Right team, but "not urgent" at 0.80 and "major" rather than "critical".

state     Subject: Checkout broken / Hi, every customer who tries to pay on our store has been getting a 502 at the payment step for the last hour. We are losing orders. Account: acme-retail (Scale plan).
questions
  team      billing · technical · sales · other
  urgent    yes · no
  severity  no impact · minor · major · critical

This sounds like a payment outage, so the technical team should own it. It is urgent because customers cannot check out, and I would call the severity major or critical. I'm about 90% confident on the routing.

…502 at the payment step for the last hour…  Which team should own this?
[slot] billing   [slot] technical   [slot] sales   [slot] other
[slot] urgent?   [slot] severity…

One ticket, three questions, every allowed answer listed. About 45 tokens, written one at a time. Each one is a separate pass through the model. Highlighted: what your code has to dig out, a hedge, and a confidence that is only text. Each [slot] is read by the scoring head. Softmax runs over each question's options, not the whole vocabulary. Jev, one request, 432 ms round trip. Three answers, each with a full distribution. Base Laya, 43 ms on a laptop. Right team, but "not urgent" at 0.80 and "major" rather than "critical".

state     Subject: Checkout broken / Hi, every customer who tries to pay on our store has been getting a 502 at the payment step for the last hour. We are losing orders. Account: acme-retail (Scale plan).
questions
  team      billing · technical · sales · other
  urgent    yes · no
  severity  no impact · minor · major · critical

This sounds like a payment outage, so the technical team should own it. It is urgent because customers cannot check out, and I would call the severity major or critical. I'm about 90% confident on the routing.

…502 at the payment step for the last hour…  Which team should own this?
[slot] billing   [slot] technical   [slot] sales   [slot] other
[slot] urgent?   [slot] severity…

One ticket, three questions, every allowed answer listed. About 45 tokens, written one at a time. Each one is a separate pass through the model. Highlighted: what your code has to dig out, a hedge, and a confidence that is only text. Each [slot] is read by the scoring head. Softmax runs over each question's options, not the whole vocabulary. Jev, one request, 432 ms round trip. Three answers, each with a full distribution. Base Laya, 43 ms on a laptop. Right team, but "not urgent" at 0.80 and "major" rather than "critical".

state     Subject: Checkout broken / Hi, every customer who tries to pay on our store has been getting a 502 at the payment step for the last hour. We are losing orders. Account: acme-retail (Scale plan).
questions
  team      billing · technical · sales · other
  urgent    yes · no
  severity  no impact · minor · major · critical

This sounds like a payment outage, so the technical team should own it. It is urgent because customers cannot check out, and I would call the severity major or critical. I'm about 90% confident on the routing.

…502 at the payment step for the last hour…  Which team should own this?
[slot] billing   [slot] technical   [slot] sales   [slot] other
[slot] urgent?   [slot] severity…

One ticket, three questions, every allowed answer listed. About 45 tokens, written one at a time. Each one is a separate pass through the model. Highlighted: what your code has to dig out, a hedge, and a confidence that is only text. Each [slot] is read by the scoring head. Softmax runs over each question's options, not the whole vocabulary. Jev, one request, 432 ms round trip. Three answers, each with a full distribution. Base Laya, 43 ms on a laptop. Right team, but "not urgent" at 0.80 and "major" rather than "critical".

state     Subject: Checkout broken / Hi, every customer who tries to pay on our store has been getting a 502 at the payment step for the last hour. We are losing orders. Account: acme-retail (Scale plan).
questions
  team      billing · technical · sales · other
  urgent    yes · no
  severity  no impact · minor · major · critical

This sounds like a payment outage, so the technical team should own it. It is urgent because customers cannot check out, and I would call the severity major or critical. I'm about 90% confident on the routing.

…502 at the payment step for the last hour…  Which team should own this?
[slot] billing   [slot] technical   [slot] sales   [slot] other
[slot] urgent?   [slot] severity…

One ticket, three questions, every allowed answer listed. About 45 tokens, written one at a time. Each one is a separate pass through the model. Highlighted: what your code has to dig out, a hedge, and a confidence that is only text. Each [slot] is read by the scoring head. Softmax runs over each question's options, not the whole vocabulary. Jev, one request, 432 ms round trip. Three answers, each with a full distribution. Base Laya, 43 ms on a laptop. Right team, but "not urgent" at 0.80 and "major" rather than "critical".

state     Subject: Checkout broken / Hi, every customer who tries to pay on our store has been getting a 502 at the payment step for the last hour. We are losing orders. Account: acme-retail (Scale plan).
questions
  team      billing · technical · sales · other
  urgent    yes · no
  severity  no impact · minor · major · critical

This sounds like a payment outage, so the technical team should own it. It is urgent because customers cannot check out, and I would call the severity major or critical. I'm about 90% confident on the routing.

…502 at the payment step for the last hour…  Which team should own this?
[slot] billing   [slot] technical   [slot] sales   [slot] other
[slot] urgent?   [slot] severity…

One ticket, three questions, every allowed answer listed. About 45 tokens, written one at a time. Each one is a separate pass through the model. Highlighted: what your code has to dig out, a hedge, and a confidence that is only text. Each [slot] is read by the scoring head. Softmax runs over each question's options, not the whole vocabulary. Jev, one request, 432 ms round trip. Three answers, each with a full distribution. Base Laya, 43 ms on a laptop. Right team, but "not urgent" at 0.80 and "major" rather than "critical".

Four layers shown; real models have 24 to 64. Each layer refines every token's vector using the others. Chatbot ending: the last vector scores every token in the vocabulary, then one is sampled. Repeat per word. Decision-model ending: read the vectors at the option slots and score just those. Who each token can see. This is the main architectural difference between an encoder like Laya and a decoder like Qwen.

Four layers shown; real models have 24 to 64. Each layer refines every token's vector using the others. Chatbot ending: the last vector scores every token in the vocabulary, then one is sampled. Repeat per word. Decision-model ending: read the vectors at the option slots and score just those. Who each token can see. This is the main architectural difference between an encoder like Laya and a decoder like Qwen.

Four layers shown; real models have 24 to 64. Each layer refines every token's vector using the others. Chatbot ending: the last vector scores every token in the vocabulary, then one is sampled. Repeat per word. Decision-model ending: read the vectors at the option slots and score just those. Who each token can see. This is the main architectural difference between an encoder like Laya and a decoder like Qwen.

Four layers shown; real models have 24 to 64. Each layer refines every token's vector using the others. Chatbot ending: the last vector scores every token in the vocabulary, then one is sampled. Repeat per word. Decision-model ending: read the vectors at the option slots and score just those. Who each token can see. This is the main architectural difference between an encoder like Laya and a decoder like Qwen.

Chat model: one prefill pass, then one pass per output token (illustrative 25 ms each). The answer still has to be parsed. Decision model: the same prefill, then the option slots are read. Jev's measured round trip for all three questions was 432 ms; Clef 27B hosted 1,255 ms; Clef Flash on a laptop 5,569 ms; Laya 43 ms. Input tokens are the whole bill. Fewer, better-chosen tokens in the state is the main cost lever. Bars are schematic. A chat request spends most of its wall time in decode; a decision request has no decode. Thirteen questions, one state, one pass. The questions are scored together, so the marginal question is nearly free.

Chat model: one prefill pass, then one pass per output token (illustrative 25 ms each). The answer still has to be parsed. Decision model: the same prefill, then the option slots are read. Jev's measured round trip for all three questions was 432 ms; Clef 27B hosted 1,255 ms; Clef Flash on a laptop 5,569 ms; Laya 43 ms. Input tokens are the whole bill. Fewer, better-chosen tokens in the state is the main cost lever. Bars are schematic. A chat request spends most of its wall time in decode; a decision request has no decode. Thirteen questions, one state, one pass. The questions are scored together, so the marginal question is nearly free.

Chat model: one prefill pass, then one pass per output token (illustrative 25 ms each). The answer still has to be parsed. Decision model: the same prefill, then the option slots are read. Jev's measured round trip for all three questions was 432 ms; Clef 27B hosted 1,255 ms; Clef Flash on a laptop 5,569 ms; Laya 43 ms. Input tokens are the whole bill. Fewer, better-chosen tokens in the state is the main cost lever. Bars are schematic. A chat request spends most of its wall time in decode; a decision request has no decode. Thirteen questions, one state, one pass. The questions are scored together, so the marginal question is nearly free.

Chat model: one prefill pass, then one pass per output token (illustrative 25 ms each). The answer still has to be parsed. Decision model: the same prefill, then the option slots are read. Jev's measured round trip for all three questions was 432 ms; Clef 27B hosted 1,255 ms; Clef Flash on a laptop 5,569 ms; Laya 43 ms. Input tokens are the whole bill. Fewer, better-chosen tokens in the state is the main cost lever. Bars are schematic. A chat request spends most of its wall time in decode; a decision request has no decode. Thirteen questions, one state, one pass. The questions are scored together, so the marginal question is nearly free.

Chat model: one prefill pass, then one pass per output token (illustrative 25 ms each). The answer still has to be parsed. Decision model: the same prefill, then the option slots are read. Jev's measured round trip for all three questions was 432 ms; Clef 27B hosted 1,255 ms; Clef Flash on a laptop 5,569 ms; Laya 43 ms. Input tokens are the whole bill. Fewer, better-chosen tokens in the state is the main cost lever. Bars are schematic. A chat request spends most of its wall time in decode; a decision request has no decode. Thirteen questions, one state, one pass. The questions are scored together, so the marginal question is nearly free.

The ticket, the question and the four options in one row. The four outlined [MASK] chips are the answer slots. Every token's vector is updated using every other token, in both directions, across 28 layers. Base Laya's real output for the ticket. Right answer, but only 62% sure, with 'billing' and 'other' close behind. The same scores at different temperatures. The order never changes; only how spread out the bars are. Base Laya on our jobs: 0.62 on keyword buckets (majority guess 0.54), 0.36 on intent, 0.61 on email category, 0.50 on the seven-way form sort. Fine-tuned on about 1,300 keyword labels: 0.89.

The ticket, the question and the four options in one row. The four outlined [MASK] chips are the answer slots. Every token's vector is updated using every other token, in both directions, across 28 layers. Base Laya's real output for the ticket. Right answer, but only 62% sure, with 'billing' and 'other' close behind. The same scores at different temperatures. The order never changes; only how spread out the bars are. Base Laya on our jobs: 0.62 on keyword buckets (majority guess 0.54), 0.36 on intent, 0.61 on email category, 0.50 on the seven-way form sort. Fine-tuned on about 1,300 keyword labels: 0.89.

The ticket, the question and the four options in one row. The four outlined [MASK] chips are the answer slots. Every token's vector is updated using every other token, in both directions, across 28 layers. Base Laya's real output for the ticket. Right answer, but only 62% sure, with 'billing' and 'other' close behind. The same scores at different temperatures. The order never changes; only how spread out the bars are. Base Laya on our jobs: 0.62 on keyword buckets (majority guess 0.54), 0.36 on intent, 0.61 on email category, 0.50 on the seven-way form sort. Fine-tuned on about 1,300 keyword labels: 0.89.

The ticket, the question and the four options in one row. The four outlined [MASK] chips are the answer slots. Every token's vector is updated using every other token, in both directions, across 28 layers. Base Laya's real output for the ticket. Right answer, but only 62% sure, with 'billing' and 'other' close behind. The same scores at different temperatures. The order never changes; only how spread out the bars are. Base Laya on our jobs: 0.62 on keyword buckets (majority guess 0.54), 0.36 on intent, 0.61 on email category, 0.50 on the seven-way form sort. Fine-tuned on about 1,300 keyword labels: 0.89.

The ticket, the question and the four options in one row. The four outlined [MASK] chips are the answer slots. Every token's vector is updated using every other token, in both directions, across 28 layers. Base Laya's real output for the ticket. Right answer, but only 62% sure, with 'billing' and 'other' close behind. The same scores at different temperatures. The order never changes; only how spread out the bars are. Base Laya on our jobs: 0.62 on keyword buckets (majority guess 0.54), 0.36 on intent, 0.61 on email category, 0.50 on the seven-way form sort. Fine-tuned on about 1,300 keyword labels: 0.89.

# model/head.py, Jeeves (complete)
class PointerHead(nn.Module):
    def __init__(self, d_embed, d_latent=256):
        self.q = nn.Linear(d_embed, d_latent)   # reads <decide>
        self.k = nn.Linear(d_embed, d_latent)   # reads each </opt>
        self.scale = d_latent ** -0.5
        self.temperature = 1.0               # fitted after training

    def forward(self, questions, decide):
        options = self.k(questions)
        query = self.q(decide)
        logits = (options @ query) * self.scale   # one score per option
        return logits / self.temperature

A decoder LLM, unchanged. On its own it can only write. Outlined chips are the structural markers. Optional reasoning (dashed) goes between <think> tags, then the question is repeated before <decide>. The entire decision head. It compares one vector with each option's vector, and that's it. The 9B base stays frozen; only the small LoRA adapters and the head are trained. Bar length shows how far each stage ran: SFT to the end, RL stopped at 64% of its schedule. With reasoning on, the answer still comes from the head, but the request is back to many passes: 3–10 s on an A100 in our runs.

# model/head.py, Jeeves (complete)
class PointerHead(nn.Module):
    def __init__(self, d_embed, d_latent=256):
        self.q = nn.Linear(d_embed, d_latent)   # reads <decide>
        self.k = nn.Linear(d_embed, d_latent)   # reads each </opt>
        self.scale = d_latent ** -0.5
        self.temperature = 1.0               # fitted after training

    def forward(self, questions, decide):
        options = self.k(questions)
        query = self.q(decide)
        logits = (options @ query) * self.scale   # one score per option
        return logits / self.temperature

A decoder LLM, unchanged. On its own it can only write. Outlined chips are the structural markers. Optional reasoning (dashed) goes between <think> tags, then the question is repeated before <decide>. The entire decision head. It compares one vector with each option's vector, and that's it. The 9B base stays frozen; only the small LoRA adapters and the head are trained. Bar length shows how far each stage ran: SFT to the end, RL stopped at 64% of its schedule. With reasoning on, the answer still comes from the head, but the request is back to many passes: 3–10 s on an A100 in our runs.

# model/head.py, Jeeves (complete)
class PointerHead(nn.Module):
    def __init__(self, d_embed, d_latent=256):
        self.q = nn.Linear(d_embed, d_latent)   # reads <decide>
        self.k = nn.Linear(d_embed, d_latent)   # reads each </opt>
        self.scale = d_latent ** -0.5
        self.temperature = 1.0               # fitted after training

    def forward(self, questions, decide):
        options = self.k(questions)
        query = self.q(decide)
        logits = (options @ query) * self.scale   # one score per option
        return logits / self.temperature

A decoder LLM, unchanged. On its own it can only write. Outlined chips are the structural markers. Optional reasoning (dashed) goes between <think> tags, then the question is repeated before <decide>. The entire decision head. It compares one vector with each option's vector, and that's it. The 9B base stays frozen; only the small LoRA adapters and the head are trained. Bar length shows how far each stage ran: SFT to the end, RL stopped at 64% of its schedule. With reasoning on, the answer still comes from the head, but the request is back to many passes: 3–10 s on an A100 in our runs.

# model/head.py, Jeeves (complete)
class PointerHead(nn.Module):
    def __init__(self, d_embed, d_latent=256):
        self.q = nn.Linear(d_embed, d_latent)   # reads <decide>
        self.k = nn.Linear(d_embed, d_latent)   # reads each </opt>
        self.scale = d_latent ** -0.5
        self.temperature = 1.0               # fitted after training

    def forward(self, questions, decide):
        options = self.k(questions)
        query = self.q(decide)
        logits = (options @ query) * self.scale   # one score per option
        return logits / self.temperature

A decoder LLM, unchanged. On its own it can only write. Outlined chips are the structural markers. Optional reasoning (dashed) goes between <think> tags, then the question is repeated before <decide>. The entire decision head. It compares one vector with each option's vector, and that's it. The 9B base stays frozen; only the small LoRA adapters and the head are trained. Bar length shows how far each stage ran: SFT to the end, RL stopped at 64% of its schedule. With reasoning on, the answer still comes from the head, but the request is back to many passes: 3–10 s on an A100 in our runs.

# model/head.py, Jeeves (complete)
class PointerHead(nn.Module):
    def __init__(self, d_embed, d_latent=256):
        self.q = nn.Linear(d_embed, d_latent)   # reads <decide>
        self.k = nn.Linear(d_embed, d_latent)   # reads each </opt>
        self.scale = d_latent ** -0.5
        self.temperature = 1.0               # fitted after training

    def forward(self, questions, decide):
        options = self.k(questions)
        query = self.q(decide)
        logits = (options @ query) * self.scale   # one score per option
        return logits / self.temperature

A decoder LLM, unchanged. On its own it can only write. Outlined chips are the structural markers. Optional reasoning (dashed) goes between <think> tags, then the question is repeated before <decide>. The entire decision head. It compares one vector with each option's vector, and that's it. The 9B base stays frozen; only the small LoRA adapters and the head are trained. Bar length shows how far each stage ran: SFT to the end, RL stopped at 64% of its schedule. With reasoning on, the answer still comes from the head, but the request is back to many passes: 3–10 s on an A100 in our runs.

# model/head.py, Jeeves (complete)
class PointerHead(nn.Module):
    def __init__(self, d_embed, d_latent=256):
        self.q = nn.Linear(d_embed, d_latent)   # reads <decide>
        self.k = nn.Linear(d_embed, d_latent)   # reads each </opt>
        self.scale = d_latent ** -0.5
        self.temperature = 1.0               # fitted after training

    def forward(self, questions, decide):
        options = self.k(questions)
        query = self.q(decide)
        logits = (options @ query) * self.scale   # one score per option
        return logits / self.temperature

A decoder LLM, unchanged. On its own it can only write. Outlined chips are the structural markers. Optional reasoning (dashed) goes between <think> tags, then the question is repeated before <decide>. The entire decision head. It compares one vector with each option's vector, and that's it. The 9B base stays frozen; only the small LoRA adapters and the head are trained. Bar length shows how far each stage ran: SFT to the end, RL stopped at 64% of its schedule. With reasoning on, the answer still comes from the head, but the request is back to many passes: 3–10 s on an A100 in our runs.

Ticket: Subject: Checkout broken … 502 at the payment step for the last hour …
Which team should own this? Answer with one letter.
A. billing   B. technical   C. sales   D. other
max_tokens=1, temperature=0, logprobs=true, top_logprobs=20

An ordinary prompt. The readout is the only unusual part. Top alternatives at the answer position (shape illustrative; the real list is model-specific). Keep the four letters, drop the rest, renormalize. GPT-4.1-mini on the ticket, 2,747 ms for two questions in two calls. 1.00 on technical. Letter bias, missing logprobs in reasoning modes, one generation per question, and one output token billed each time. On the board: Decider chat (Gemma-4-31B, no training) 57.33; simple-jev (Qwen3.8-27B, no training) 55.74; the hosted leader 57.91.

Ticket: Subject: Checkout broken … 502 at the payment step for the last hour …
Which team should own this? Answer with one letter.
A. billing   B. technical   C. sales   D. other
max_tokens=1, temperature=0, logprobs=true, top_logprobs=20

An ordinary prompt. The readout is the only unusual part. Top alternatives at the answer position (shape illustrative; the real list is model-specific). Keep the four letters, drop the rest, renormalize. GPT-4.1-mini on the ticket, 2,747 ms for two questions in two calls. 1.00 on technical. Letter bias, missing logprobs in reasoning modes, one generation per question, and one output token billed each time. On the board: Decider chat (Gemma-4-31B, no training) 57.33; simple-jev (Qwen3.8-27B, no training) 55.74; the hosted leader 57.91.

Ticket: Subject: Checkout broken … 502 at the payment step for the last hour …
Which team should own this? Answer with one letter.
A. billing   B. technical   C. sales   D. other
max_tokens=1, temperature=0, logprobs=true, top_logprobs=20

An ordinary prompt. The readout is the only unusual part. Top alternatives at the answer position (shape illustrative; the real list is model-specific). Keep the four letters, drop the rest, renormalize. GPT-4.1-mini on the ticket, 2,747 ms for two questions in two calls. 1.00 on technical. Letter bias, missing logprobs in reasoning modes, one generation per question, and one output token billed each time. On the board: Decider chat (Gemma-4-31B, no training) 57.33; simple-jev (Qwen3.8-27B, no training) 55.74; the hosted leader 57.91.

Ticket: Subject: Checkout broken … 502 at the payment step for the last hour …
Which team should own this? Answer with one letter.
A. billing   B. technical   C. sales   D. other
max_tokens=1, temperature=0, logprobs=true, top_logprobs=20

An ordinary prompt. The readout is the only unusual part. Top alternatives at the answer position (shape illustrative; the real list is model-specific). Keep the four letters, drop the rest, renormalize. GPT-4.1-mini on the ticket, 2,747 ms for two questions in two calls. 1.00 on technical. Letter bias, missing logprobs in reasoning modes, one generation per question, and one output token billed each time. On the board: Decider chat (Gemma-4-31B, no training) 57.33; simple-jev (Qwen3.8-27B, no training) 55.74; the hosted leader 57.91.

Ticket: Subject: Checkout broken … 502 at the payment step for the last hour …
Which team should own this? Answer with one letter.
A. billing   B. technical   C. sales   D. other
max_tokens=1, temperature=0, logprobs=true, top_logprobs=20

An ordinary prompt. The readout is the only unusual part. Top alternatives at the answer position (shape illustrative; the real list is model-specific). Keep the four letters, drop the rest, renormalize. GPT-4.1-mini on the ticket, 2,747 ms for two questions in two calls. 1.00 on technical. Letter bias, missing logprobs in reasoning modes, one generation per question, and one output token billed each time. On the board: Decider chat (Gemma-4-31B, no training) 57.33; simple-jev (Qwen3.8-27B, no training) 55.74; the hosted leader 57.91.