Weekly Coaching Workbook

The floor scored 224 calls and 5,281 turns this week, carrying 463 detected objections. Sentiment recovered on 56% of them. 19 of 27 reps carry a lag material enough to coach against their own tier, and the most common lag is affordability conversion (7 reps). The hardest objection on the floor is needs to consult a partner, recovering 46% of the time against a floor average of 56%.

The bar behind each value is that rep on their own tier's range; the vertical rule is the tier median. A metric is flagged only when the rep sits at least 0.60 robust z below their tier median, so a rep inside the band gets no manufactured coaching item. Reserved capacity is the point: flagging everybody flags nobody.

One card per rep, the same five metrics on every card, in the same order. A workbook that reorders columns per rep cannot be read across a floor, and a director reading twenty-seven of these has to be able to scan the same position every time.

Spread is the median absolute deviation scaled to a standard-deviation equivalent, which is what the robust z in the rep cards divides by. Mean and standard deviation would let one rep with four calls and a bad Tuesday redefine the tier.

The call, as one request

client = anthropic.Anthropic()      # key from ANTHROPIC_API_KEY, never an argument

message = client.messages.create(
    model="claude-sonnet-5",
    max_tokens=16000,
    system=[{"type": "text", "text": SYSTEM_PROMPT,
             "cache_control": {"type": "ephemeral"}}],
    messages=[{"role": "user", "content": build_user_message(call, window)}],
    output_config={"format": {"type": "json_schema", "schema": TURN_LABEL_SCHEMA}},
)

System prompt, opening block

You are a quality analyst for a consumer finance sales floor. You read one
window of a recorded sales call at a time and return structured labels for
every turn in that window. You are scoring the conversation, not judging the
customer and not writing coaching copy. Another system does the coaching.

TASK
For each turn in the window, return three things:
  1. sentiment: how the speaker of that turn sounds on this turn.
  2. labels: which rebuttal taxonomy labels the turn carries, if any.
  3. confidence: how sure you are about the labels on that turn.
Most turns carry no labels. An empty labels array is the correct and common
answer. Do not reach for a label to fill the field.

The schema the response is constrained to

{
  "type": "object",
  "properties": {
    "turns": {
      "type": "array",
      "description": "One entry per turn in the supplied window, in order.",
      "items": {
        "type": "object",
        "properties": {
          "turn": {
            "type": "integer",
            "description": "Turn index, copied from the input window."
          },
          "sentiment": {
            "type": "number",
            "description": "Speaker sentiment on this turn, -1.0 to 1.0, against the anchor scale in the system prompt."
          },
          "labels": {
            "type": "array",
            "description": "Rebuttal taxonomy labels carried by this turn. Empty array when the turn carries none.",
            "items": {
              "type": "string",
              "enum": [
                "cant_afford",
                "wants_loan",
                "consult_spouse",
                "needs_time",
                "process_explained"
              ]
            }
          },
          "confidence": {
            "type": "number",
            "description": "0.0 to 1.0 confidence in the labels on this turn."
          },
          "evidence": {
            "type": "string",
            "description": "Verbatim span from this turn that carries the label, at most 140 characters. Empty string when labels is empty."
          }
        },
        "required": [
          "turn",
          "sentiment",
          "labels",
          "confidence",
          "evidence"
        ],
        "additionalProperties": false
      }
    }
  },
  "required": [
    "turns"
  ],
  "additionalProperties": false
}

The schema deliberately carries no numeric bounds. The API's JSON schema subset does not enforce minimum and maximum, so writing them in would read as a guarantee the platform never made. Bounds are enforced in validate_turn_record() instead, on every record, from both the live path and the fixture path.

One payload, verbatim from data/agent_features.jsonl

{
  "schema_version": "1.2.0",
  "call_id": "NF-20260720-0052",
  "rep_id": "R-09",
  "tier": "Tier 2",
  "emitted_at_turn": 14,
  "emitted_at_sec": 185,
  "features": {
    "objection": "cant_afford",
    "objections_in_turn": 1,
    "sentiment_now": -0.077,
    "sentiment_mean": 0.2,
    "sentiment_slope_3": -0.177,
    "objections_raised_to_date": 1,
    "objections_handled_to_date": 0,
    "process_explained_count": 2,
    "process_explained_before_first_objection": true,
    "rep_talk_share": 0.723,
    "turns_elapsed": 15,
    "seconds_elapsed": 185
  },
  "recommended_module": "affordability_reframe",
  "label_confidence": 0.942,
  "actionable": true
}

This is the grain everything else is built from. The workbook aggregates these turns; the live agent reads them as they arrive. Nothing in the coaching board exists that cannot be traced back to a row on a page like this one.

One scored record, verbatim from data/scores.jsonl

{
  "call_id": "NF-20260720-0210",
  "engine": "fixture",
  "model": "pre-generated fixture (no API call)",
  "prompt_version": "2026-07-13.3",
  "turns": [
    {
      "turn": 0,
      "sentiment": 0.276,
      "labels": [],
      "confidence": 0.969,
      "evidence": ""
    },
    {
      "turn": 1,
      "sentiment": 0.385,
      "labels": [],
      "confidence": 0.957,
      "evidence": ""
    },
    {
      "turn": 2,
      "sentiment": 0.312,
      "labels": [],
      "confidence": 0.883,
      "evidence": ""
    },
    {
      "turn": 3,
      "sentiment": 0.561,
      "labels": [],
      "confidence": 0.859,
      "evidence": ""
    },
    {
      "turn": 4,
      "sentiment": 0.338,
      "labels": [],
      "confidence": 0.973,
      "evidence": ""
    },
    {
      "turn": 5,
      "sentiment": 0.423,
      "labels": [],
      "confidence": 0.857,
      "evidence": ""
    },
    {
      "...": "15 more turns"
    }
  ]
}