ExactRef

Inspiration

On 2026-09-05 we placed one real CALL-E call (Calls API, Python SDK 0.7.0) to obtain a supplier off-hire reference. The recipient read it out, the agent read it back in full, the recipient said "yes". The structured_result came back schema-valid: 07198SECTIST. The real reference was 07198FECTIST. One character. Confirmed. Wrong.

Every agent integration we looked at would have written that value into a record, because "schema-valid + task_completed + the recipient said yes" is what they treat as truth. ExactRef exists to make that write impossible.

What it does

ExactRef is the write-gate after a CALL-E call. It is not a calling desk.

Character-level check against an intended value. Intended vs extracted are compared after normalising spaces, hyphens and case. The decision names the first differing character (index 5: F → S). F/S, 0/O, 1/I homophones are the same class.
Provenance is a named state, not a score. unknown, spoken_only, conversational_confirmed, mismatch, independently_verified. Only the last is writable. Readback-plus-yes is conversational_confirmed — evidence, not verification.
Second-channel stamp. To open the write a human must type the value from a written channel (email, portal, paper) and claim it. Typing the spoken (wrong) value keeps mismatch.
Task compiler that refuses to lead the call. It compiles an outbound CALL-E task (one full readback, no interruptions, ask for a word for an ambiguous letter, leave contradictory times unresolved) and exits non-zero if the intended identifier leaks into the task text.
Wait honesty. queued + attempt activity is not idle (do not create again); a local wait timeout is not hangup or cancel; completed without structured_result is not an extraction; MCP COMPLETED is not task success — read result{}.
gate for real CALL-E output. Feed it a saved call object from REST, the Python/TypeScript SDK, calle call status --json, or the MCP result{} envelope. It refuses non-terminal calls, then classifies structured_result[field].

How we built it

skills/exact-ref/ — the Agent Skill (Claude Code, Codex, Cursor, OpenClaw, skills.sh): SKILL.md, scripts/exactref.mjs (one zero-dependency Node 18+ file: classify / verify / compile / gate / fixtures; exit 0 only when writable), fixtures including the live F/S call object, safety notes.
Board (Next.js 15, React 19, TypeScript): a docket that groups the off-hire exception OH-01 with its two blocked writes (FS-01 reference, FS-02 collection window) plus three other identifier classes. The identifier pair is the largest object on the page; provenance is text plus shape, never colour alone.
Tests: vitest for the rules and parity between the standalone skill script and the board; Playwright e2e for the write-gate flow.
Two upstream fixes to CALL-E from what we hit while building: call-e-integrations#128 (calle mcp call treats tool isError as failure) and call-e-integrations#129 (Cursor skill reads result{} and marks tool output untrusted).
A feedback lab (docs/feedback/lab/) of verified, reproducible findings across the Python/TypeScript SDKs, CLI, MCP server, and docs, each with one concrete ask.

Challenges

Isolating the F/S stage is not possible from outside CALL-E, so the skill refuses to guess: it compares against the intended value rather than blaming ASR or the LLM.
Top-level queued during an active attempt, a 15-second MCP call timeout, and recover re-sending run_call all make a second dial easy. Wait honesty is a rule set, not a UI hint.
Keeping the package to one product. Off-hire is a proof exception (OH-01), not a second desk.

What we learned

Schema validation checks shape. Readback checks the transcript. Neither checks the fact. The only thing that does is a second channel and a human who types.

What's next

Classify-only MCP tool (accepts an observation, returns a decision, never dials). @exactref/core once the skill is merged. More homophone fixtures only if a real call produces one.

Significant updates during the Submission Period

Everything here was created during the Submission Period (first commit 2026-09-13).

Try it (no credentials, no calls)

git clone https://github.com/Arshgill01/ExactRef && cd ExactRef

# 10-second check: the live miss is blocked, exit 2, mismatch at character 6
node skills/exact-ref/scripts/exactref.mjs classify --intended 07198FECTIST --extracted 07198SECTIST --readback

# Gate a saved CALL-E call object
node skills/exact-ref/scripts/exactref.mjs gate --call skills/exact-ref/references/call-fs01.json --intended 07198FECTIST

# Human types the value from a second channel → independently_verified, exit 0
node skills/exact-ref/scripts/exactref.mjs verify --intended 07198FECTIST --extracted 07198SECTIST --typed 07198FECTIST --second-channel

# Board
npm install && npm test && npm run dev    # http://127.0.0.1:3450 → OH-01 → FS-01

On FS-01: type 07198SECTIST, tick the second-channel box, click "Mark independently verified" — still Mismatch — do not write. Type 07198FECTIST — Independently verified, writable yes. "Place live call" opens a dialog saying live create is disabled.

Awesome PR: https://github.com/CALLE-AI/awesome-phone-call-agents/pull/673