D2 — who answered

Traced at commit e53d0dc · libs/contracts/src/lib/{vector-resolution,extractor,canonicalization,contracts,outbox}.ts · libs/temporal/src/{workflows,activities}/call-attempt.ts · apps/worker/src/main.ts · live state read 10 September 2026

Workflow 1 — where the value comes from

Step 1 — the voice provider

1. The provider sends a call analysis after the call ends.
2. The platform reads three fields from it: call_successful, in_voicemail and call_summary.
3. The platform sets WRONG_PERSON only when all three conditions below hold.
4. call_successful is the boolean false.
5. in_voicemail is the boolean false. An absent field does not satisfy this condition.
6. call_summary contains wrong number or wrong person, in any letter case.
7. In a test, an analysis without in_voicemail gave no D2 value. The headline stayed needs_review.
8. The platform stores the value with confidence 0.8 and the source retell_analysis.
9. Step 1 cannot set DECISION_MAKER, GATEKEEPER or REASSIGNED.

Before step 2

1. The workflow asks for extraction only for an answered call.
2. Each answered call makes a model request. The model is gpt-4o-mini unless OPENAI_MODEL names another.
3. Before the request, the extractor removes phone numbers, email addresses and card numbers from the transcript.
4. It also removes social security numbers, street addresses and ZIP codes. It matches fixed patterns only.
5. It does not remove names. It sends the provider’s call summary without any removal.
6. In a test, a phone number with spaces and the whole call summary reached the model unchanged.
7. The extractor sends the model request before it checks the phrase list. A phrase match then discards the paid answer.
8. If the request fails, the extractor uses the phrase list alone.
9. The cost Blueprint does not price this request.
10. A rehearsal runs extraction against an impersonated model and spends no tokens.
11. Extraction needs OPENAI_API_KEY. Without it, steps 2 and 3 do not run.

Step 2 — the phrase list

1. Step 2 outranks the model. A phrase match wins even when the model answered first.
2. It joins every turn of the transcript into one string. It drops the speaker labels.
3. It checks the values in a fixed order: DECISION_MAKER, GATEKEEPER, WRONG_PERSON, REASSIGNED.
4. For each value, it looks for any phrase of that value as whole words.
5. The first value with a match wins. Step 2 then stops.
6. A match replaces the model answer, whatever the model said.
7. The list holds 35 phrases. Their confidence runs from 0.85 to 0.95.
8. Every phrase confidence is above the review threshold of 0.7. A match alone never sends a call to review.

Step 3 — the language model

1. The model receives the transcript with AGENT and USER labels, and the call summary.
2. The platform uses the model answer only when no phrase matched.
3. The platform discards an UNKNOWN answer. That answer never fills D2.
4. An UNKNOWN answer for D2 can still send the call to review. See Outcome resolution.

What the record keeps

1. The platform writes D2 into disposition_vector under the key d2Party.
2. The entry holds the value, the confidence, the evidence and a source label.
3. Step 1 writes the source retell_analysis and no evidence.
4. Step 2 and step 3 both write the source llm.
5. When a phrase replaces the model answer, the record keeps no copy of the model answer.
6. To find a phrase match, compare the single evidence entry with the phrase list below.

Workflow 2 — what each value does

1. D2 is step 4 of the headline waterfall. D5, D6 and D1 come first.
2. WRONG_PERSON and REASSIGNED both give the headline wrong_number.
3. DECISION_MAKER and GATEKEEPER give no headline. The waterfall moves on to D4, then D3.
4. Steps 2 and 3 change the headline only when step 1 left it at needs_review.
5. A later D2 value never replaces a headline that step 1 decided.
6. Step 1 never sets D2 and D4 on the same call. D2 needs call_successful false, and D4 needs it true.
7. A gauge counts D2 values over seven days for calls that reached a person.
8. The gauge shows an empty D2 as UNRECORDED.
9. The gauge also counts phrase matches from the agent’s own words.

Workflow 3 — the agent’s own words

1. The test ran the real extractor on the two turns in the diagram.
2. A model stub answered WRONG_PERSON with confidence 0.9.
3. The extractor returned DECISION_MAKER at 0.9, from the phrase speaking.
4. The phrase came from the agent’s question, not from the caller.
5. The same call without the agent’s turn returned WRONG_PERSON at 0.95.
6. An August 2026 flow backup tells the agent to confirm that it is speaking with the merchant.
7. An agent line with hold on sets GATEKEEPER in the same way.
8. A real wrong-number call gets DECISION_MAKER this way, unless step 1 caught the wrong number.

A caller’s words out of context

1. A second test used one caller line: I'm retired from the navy, I run the shop now. Tell me more.
2. Step 2 matched retired and set REASSIGNED.
3. The merged result had the headline wrong_number, confidence 0.9 and no review.
4. The caller owned the business and asked to hear more.

The phrase list

Step 2 checks the rows from top to bottom. The first row with a match wins. Dictionary version 1.0.

Known defects

1. Step 2 reads the agent. The phrase list receives agent turns and caller turns as one string. Fix: pass only the caller’s turns.
2. Value order beats word order. A transcript with yes this is and wrong number gets DECISION_MAKER. The position of each phrase does not matter.
3. Short phrases match ordinary speech. speaking, hold on, moved on and retired occur in calls with the right person.
4. A match erases the disagreement. The record keeps no copy of the model answer that the phrase replaced.
5. One label for two sources. Step 2 and step 3 both write llm.
6. The model call happens first. Every answered real call pays for a model request, and any phrase match then discards the answer.
7. Names reach the model. Redaction matches fixed patterns only, and the call summary goes to the model unchanged.
8. Step 1 needs an explicit field. An analysis without in_voicemail never gives WRONG_PERSON.

What rehearsal data can show

1. On 10 September 2026, prod held 2,157,695 rehearsal attempts and 4 real attempts.
2. None of the 4 real attempts holds a D2 value.
3. A sample of about 1% of stage attempts held 240 D2 values. The phrase list set none of them.
4. The impersonated model set 213 of them. Step 1 set the other 27.
5. Rehearsal values show that the code path runs. They do not measure how well D2 reads a real caller.
6. The corrections in the review queue are the only real measurement. See Outcome resolution.