call-sycophancy-guard

call-sycophancy-guard

An agent that agrees with everything confirms nothing.

Language models are demonstrably sycophantic: trained on human feedback, they affirm users far more often than humans do, and that behavior survives into deployed assistants. On a phone call this has a specific cost: the callee pushes back on a fact - "no, it's $50" - and the agent folds, not because it saw evidence, but because agreeing is polite. The closing confirmation then repeats the adopted value, and whatever parses that outcome writes a fact nobody established. This skill catches exactly that sequence.

When To Use

• after any CALL-E call whose result contains amounts, dates, or times the callee disputed, and before the outcome is written anywhere
• with the goal text (or plan JSON) the call was built from, so ground facts can be extracted and checked against adopted values
• before placing a fact-bearing call, to craft a goal that holds facts without being rude

When Not To Use

• to prove the callee was wrong: a capitulation is an unverified stance switch, and the callee holding a paper invoice may well be right; the card routes to verification, never to reversal
• to detect emotional pressure or hostility; use call-fraud-shield for scam patterns and call-emotional-contagion-monitor for affect
• during a call; this is strictly post-call analysis plus pre-call goal crafting, because CALL-E exposes transcripts, not live audio

Workflow

Audit a finished call

python3 scripts/sycophancy_guard.py analyze --transcript path/to/call-result.json \
  [--goal-file path/to/goal.txt]

Reads the real get_call_run result shape ({status, result: {transcript}}) or the flat shape used by sibling skill fixtures. The goal file is plain text or a JSON with a goal field. Emits a card:

• goal_facts: amounts, dates, times extracted from the goal text
• pushback_events[]: turn index, masked span, the responding agent turn, and its stance:
  • HOLDS - restates the record ("our records show $45")
  • CAPITULATES - adopts the contradicting value ("you're right, it is $50") without verifiable evidence
  • VERIFIES - defers to an independent channel (statement number, callback colleague, transfer)
  • UNADDRESSED - pivots away
• outcome_taint: true when a capitulated value (a digit absent from the goal facts) reappears in a closing confirmation turn
• verdict: CLEAN / PRESSURE_TAINTED / UNCERTAIN (UNCERTAIN when pushback was left unaddressed; stance classification runs without a goal file too, but goal facts are then empty and the taint check keys on the transcript's digits alone)
• fields_to_verify_via_second_channel and the matching recommended action