Menu-step decomposition (no graph nodes exist): 6 system/code steps (intake trigger + answer-the-call, fetch-record, route-draft-to-approval, log-call, update-record) x $0.025 = $0.15, plus 7 judgment steps (intake extraction + identify, check question, look up, answer, decide next step, draft follow-up). At standard tier: $0.15 + 7 x $0.10 = $0.85 ≈ the $0.855 menu headline (intake $0.13 + module $0.725; no $0.20 parse — a voice call reads no document). Measured: 13.1 credits = $1.31/run across 203 calls on a mid-tier model. All-mid arithmetic: $0.15 + 7 x $0.20 = $1.55. The measured $1.31 sits between all-std $0.85 and all-mid $1.55, so the mid-tier assumption closes the gap: $1.31 − $0.85 = $0.46 ≈ 4.6 of the 7 judgment calls stepping up $0.10 each — mid tier on most but not every judgment turn of an average call, matching the repo's own verdict that 'the entire gap is model tier' (Wünsche shows the same pattern). Voice metering rides on top and is billed separately: $0.1115/min ($391.37 speech / 3,510 min in the order reconciliation) x 3.51 measured min/call (179 min / 51 calls) ≈ $0.39/call, plus toll-free $79.37/mo.

Measured: 13.1 credits = $1.31 per run, measured across 203 calls on a mid-tier model (standard-tier menu price is $0.855 — the entire gap is model tier); voice metering $0.1115/min on top at 3.51 measured min/call (credit-rail measured)

Graph read: none on disk — this is NOT a Beam graph. sol-hudson/hershey/p2-build/build/voice/03-working/agent.json is a 229-byte ElevenLabs voice-agent config (agent_3101kvwph1e3fmesmnhg1zxeyj8z, voice Sarah, tts eleven_turbo_v2, llm claude-sonnet-4-6, public auth); tools run client-side in web/index.html over synthetic scenarios plus a Cloudflare Worker, per lookup-tool-spec.md and the catalog. Breakdown derived instead from the measured 13.1 credits/run and the menu's post-placement-care step list (6 judgment + 5 system) plus HR intake, which the corrections ledger confirms as the deck's declared split.

• The #1 rule from the blueprint: the agent looks up the candidate record only to disambiguate the caller and to build the handoff — it never speaks record contents aloud; search_candidate_record output is marked INTERNAL, and disambiguation questions must reveal nothing from the file (lookup-tool-spec.md).
• C1/C2 (answer + PII check) are real-time; C3/C4 (ATS search + take action) are async and human-validated — every resulting action is drafted as a proposal for approval, never executed live (concept decomposition, canonical scope model).
• Demo stack is deliberately serverless and outside Beam: ElevenLabs voice agent + client-side tools over synthetic candidates.json + Cloudflare Worker to spreadsheets; production contracts map to Airtable/SuccessFactors search and Gmail drafts — a copycat build must not confuse the demo wiring with an integration story.
• Voice-specific guards the docs require: recording-consent/AI disclosure opening (PA is a two-party-consent state), barge-in and sub-1s P95 turn latency on a cascaded Deepgram→Claude→ElevenLabs stack, and an explicit multi-requisition rule (rank by latest activity timestamp, pick the controlling status).

• The live-status contradiction (§4.1): the blueprint's safety model (real-time = standard answers only, ATS lookup async + human-validated) directly conflicts with the demo ask for live status lookup — resolve it before committing an integration story; the recommended line is read-only live lookup with no record changes.
• Telephony lead time (§4.2): real phone numbers take 2-3 weeks to provision — longer than the demo runway; run the demo on web/test voice and name PSTN as roadmap rather than promising a live number.
• Ground the pitch in the real call log, not the blueprint's guesses: 203 logged calls show status+onboarding = ~77% of volume and only ~26% resolved on the spot — the 74% 'email and wait' is what the agent takes over; the blueprint's '53% status + 26% assessment' was corrected (assessment is ~10%, the real #2 is onboarding ~23%).
• Integration honesty (§4.7): Hershey asked about Eightfold + SuccessFactors directly; draw a clean line between what is genuinely connected (sandbox/mock) and roadmap — over-claiming damages the TPRM/legal review that follows.