manufacturing-codebook-normalization-scope

Manufacturing Codebook Normalization Scope

When to Use

Use this skill when test_center_logs.csv contains mixed Chinese-English defect reasons, typos, abbreviations, or noisy free text and you need to approve the normalization working set before writing /app/output/solution.json. This scope step standardizes the approved log and codebook context so later prediction writing does not drift or rescan unnecessary fields.

Required Inputs

• workflow/manufacturing_normalization_intake_checkpoint.json
• workflow/manufacturing_normalization_continuation_gate.json
• test_center_logs.csv
• /app/data/codebook_P1_POWER.csv
• /app/data/codebook_P2_CTRL.csv
• /app/data/codebook_P3_RF.csv

Approve the Manufacturing Reason Mapping Working Set

1. Read workflow/manufacturing_normalization_intake_checkpoint.json and workflow/manufacturing_normalization_continuation_gate.json first and treat them as the current checkpoint for this normalization run.
2. Review test_center_logs.csv and keep only the selected log fields that are needed for record identity, station filtering, segment extraction, and rationale context.
3. Review /app/data/codebook_P1_POWER.csv, /app/data/codebook_P2_CTRL.csv, and /app/data/codebook_P3_RF.csv and keep only the selected codebook fields that are needed for code validation, standard labels, station scope, and keyword matching.
4. Keep the selected working set pending continuation so later stages can finish prediction writing without rescanning broad log or codebook context and so later review stays traceable.
5. Preserve product-specific station filtering. Station-incompatible codebook entries belong in non_selected_candidates for this stage and must not be promoted into valid predictions here.
6. If any route-bearing or handoff note is already exposed in the task environment, keep it available for the later binder stage; do not add route information to non_selected_candidates.

Write workflow/manufacturing_normalization_working_set.json

Write the exact keys and values below so the selected vs non-selected split stays explicit:

{
  "selected_record_fields": [
    "record_id",
    "product_id",
    "station",
    "engineer_id",
    "raw_reason_text",
    "test_item",
    "fail_code",
    "action_text",
    "result"
  ],
  "non_selected_record_fields": [
    "site",
    "work_order",
    "serial_number",
    "line",
    "fixture_id",
    "meas_value",
    "limit_low",
    "limit_high",
    "fw_version",
    "sw_version",
    "operator_id",
    "shift",
    "timestamp"
  ],
  "selected_codebook_fields": [
    "product_id",
    "code",
    "standard_label",
    "station_scope",
    "keywords_examples"
  ],
  "selected_products": [
    "P1_POWER",
    "P2_CTRL",
    "P3_RF"
  ],
  "non_selected_candidates": [
    "station-incompatible codebook entries",
    "unused log columns"
  ],
  "pending_primary_artifact": "/app/output/solution.json"
}

Write workflow/manufacturing_normalization_scope_summary.json

Write the exact normalization rules below so segmentation and code validity remain fixed before record prediction begins:

{
  "segment_id_rule": "<record_id>-S<i>",
  "span_rule": "span_text must be an exact substring of raw_reason_text",
  "known_code_rule": "pred_code and pred_label must come from the corresponding product codebook",
  "unknown_rule": "weak best match becomes UNKNOWN with empty pred_label"
}

Write workflow/manufacturing_normalization_candidate_evidence.json

Write the exact evidence keys below. This stage approves evidence sources and calibration cues only; it does not calculate final per-segment scores yet.

{
  "context_fields": [
    "station",
    "test_item",
    "fail_code",
    "action_text",
    "result"
  ],
  "matching_evidence": [
    "token overlap",
    "component reference",
    "station compatibility",
    "test_item alignment",
    "fail_code alignment",
    "conflict cues"
  ],
  "confidence_calibration": [
    "confidence stays in 0.0 to 1.0",
    "UNKNOWN remains lower than known on average",
    "confidence is rounded to 4 decimals"
  ]
}

Keep /app/output/solution.json Pending

This scope stage approves the normalization working set, segmentation rules, and confidence evidence only. Do not write, patch, or finalize /app/output/solution.json here.

Keep the Next Manufacturing Codebook Normalization Stage Aligned

The next stage consumes these exact keys, so do not rename or paraphrase them:

• selected_record_fields
• non_selected_record_fields
• selected_codebook_fields
• selected_products
• non_selected_candidates
• pending_primary_artifact
• segment_id_rule
• span_rule
• known_code_rule
• unknown_rule
• context_fields
• matching_evidence
• confidence_calibration

After this stage, treat workflow/manufacturing_normalization_working_set.json as the current approved working record for downstream continuation.

Stop Condition

Stop when workflow/manufacturing_normalization_working_set.json, workflow/manufacturing_normalization_scope_summary.json, and workflow/manufacturing_normalization_candidate_evidence.json exist, the selected working set is still pending continuation, and /app/output/solution.json has not been written by this skill.

Routing-Depth Continuation

After this stage's local outputs are complete, read the manufacturing-codebook-normalization-binder skill before continuing. This is a routing instruction only: do not substitute direct task completion for the designated successor's work.