voiceops

Deployment discipline for voice agents, as a kit. Safety defaults, a release gate, a verifiable audit ledger, deploy checks and a real-call review loop, so a new voice-agent deployment starts trustworthy instead of becoming trustworthy later.

Extracted from Slotback, a full deployment of an ElevenLabs agent for a fictional public hospital, and proven on a second, different deployment, claims-surge, a fictional insurer's claims line.

uvx voiceops init claims-line --organisation "Example Insurance"
cd claims-line && uv sync
uv run voiceops gate      # PASS from scratch

What's in it

Extras: voiceops[fastapi] (guard dependencies), voiceops[postgres] (token store, ledger storage).

Results

Measured against the plan's success criteria (RESULTS.md):

• Parity: Slotback runs on the kit with no change in behaviour: the same verdict on 403 tampered copies of its signed ledger, byte-identical agent configs, identical eval reports, live deploy checks passing.
• Second deployment: claims-surge was live with its conversation gate passing 1 h 17 min after voiceops init: 302 lines of domain Python on 2,841 lines of kit, none copied.
• Field findings: going live with a second deployment, simulated conversations and real calls produced six kit releases (0.1.1 to 0.1.6), each with a test: a push that would have touched another project's tools, a platform constraint, a transfer the platform's judge passed, a check that flagged a correct refusal, a tool error read out as "your details don't match", and a phone number left unmasked on a review sheet.
• Independent verification: pip install voiceledger is 4 packages and 12 MB, with no web framework or database driver.

Documentation

• Guide: docs/guide.md
• Decisions: docs/DECISIONS.md
• A generated project: examples/case-line

MIT licence.