Grounding and Faithfulness
Answer + Evidence + Reasoning: every claim must point at words that were actually said.
Why this matters
“The customer is interested” is an opinion. “The customer is interested — evidence: *‘haan bhej do, main kal tak apply kar dunga’* (turn 10, customer)” is an auditable claim. forces every output to cite its support; asks whether the output is actually entailed by that support.
This single design choice makes errors visible (reviewers check quotes in seconds), enables automatic verification (does the quote exist? was it the customer? does it entail the claim?), and turns “we don't know” into a legitimate answer instead of a guess.
Core concept
Technical explanation + simple intuition + example
Three verdicts: supported, contradicted, insufficient evidence
Given a transcript and a claim, there are three honest answers: - Supported — the transcript implies it. - Contradicted — the transcript implies it's false. - — the transcript doesn't settle it.
Most grounding errors come from merging the last two: treating “not said” as “false”, or treating “plausible” as “said”.
Evidence spans
An is the exact quote that supports the answer. Good evidence is verbatim (can be found by string search), from the right speaker, minimal (just the decisive part), and complete (includes the resolution if there's an objection).
The 10 questions
Every important concept should answer all of these
1What is it?
Requiring outputs to be supported by, and traceable to, specific parts of the source (grounding); and the degree to which outputs are entailed by it (faithfulness).
2Why does it exist?
Fluent outputs can be unsupported; without citations, errors are invisible and unauditable.
3What problem does it solve?
It makes each decision verifiable by humans and machines and gives a principled way to say 'unknown'.
4How does it work internally?
Require {answer, evidence spans, reasoning}; verify quotes exist, check speaker, check entailment; allow insufficient_evidence.
5What does it look like on real transcript data?
Hedges, agent paraphrases, resolved objections, third-party mentions and sarcasm all create claims that look supported by a partial quote.
6What can go wrong?
Cherry-picked quotes (the objection, not the resolution), paraphrased 'quotes' that don't exist, agent speech cited as customer evidence, 'not mentioned' treated as 'false'.
7How do I evaluate it?
Verdict accuracy, evidence overlap (precision/recall of cited turns vs gold), quote-existence rate, speaker-correctness rate.
8When should I use it?
Any output a human might dispute or act on: intent, consent, compliance, complaints.
9When should I NOT use it?
Pure aggregate topic stats where individual audit isn't needed — still useful for sampling audits.
10How does it affect downstream systems?
Evidence enables verification stages, faster human review, dispute resolution and error analysis.
Interactive lab: Evidence-based verification
For each claim, select the evidence turns in the transcript, choose Supported / Contradicted / Insufficient evidence, and write your reasoning. Get the reference verdict and see your answer as an auditable record.
Try it yourself
Conceptual (A), applied (B) and engineering (C) exercises
Graded by a transparent keyword rubric (not an LLM) — it checks for the key ideas, then shows a model answer for self-assessment.
Real-world application
A dispute team received a complaint: “I never agreed to this loan.” The system's record showed intent=interested with evidence “haan… theek hai” (turn 14). Replaying turn 13 revealed the agent's question: “sir aapko documents WhatsApp kar du?” — the customer agreed to receive documents, not to the loan. The evidence field made the error findable in two minutes and led to a new verifier rule: evidence must entail the specific claim, not merely be positive.
Common mistakes
It converts uncertainty into confident negatives.
Fix: Use a three-way verdict with insufficient evidence.
'haan' to 'shall I send documents?' ≠ agreeing to the loan.
Fix: Include the adjacent question; check entailment of the specific claim.
They can't be verified and often distort.
Fix: Require verbatim quotes; string-check them.
Engineering decision
How strict should evidence requirements be?
- Decisions are contestable or regulated
- Humans review outputs
- You need a verification stage
- Never skip entirely for customer-facing decisions — at least sample-audit
Quiz
Pass mark 60%. Missed questions go to your mistake log and spaced review.