> ## Documentation Index > Fetch the complete documentation index at: https://docs.heymilo.ai/llms.txt > Use this file to discover all available pages before exploring further. # Adjusting Scoring (Interview Evaluation) > Use Knowledge Base prompts to adjust how the AI scores candidate answers across your entire workspace — without editing each question individually. Knowledge Base prompts with Scope set to **Interview Evaluation** can be used to adjust overall candidate scoring on questions answered in their interview. This scoring baseline applies to all interview agents in the workspace. For finer, question-specific scoring tuning, update the answer criteria (what's a 10 and what's a 1) inside the interview agent itself. To access Knowledge Base prompts, go to **Interviewers → Tools → Knowledge Base** in the sidebar, or use search and type "knowledge base." See [Knowledge Base](knowledge-base) for how to navigate and manage entries. Knowledge Base list showing Overall scoring calibration entry with Interview Evaluation scope tag To add a new entry, click **Add Knowledge** and fill in the Title, set Scope to **Interview Evaluation**, paste your prompt into Content, and make sure Active is toggled on. Add knowledge form with Overall scoring calibration filled in The Knowledge Base prompts referenced in this guide use **Scope = "Interview Evaluation."** These are separate from FAQ-style entries you use to answer candidate questions during interviews. *** ## How to make interview scoring stricter or less strict ### 1. The one thing to know first All scoring lives in Knowledge Base entries with **Scope = "Interview Evaluation."** These apply to every agent in the workspace. After any edit, click **Reanalyze** on a candidate to see the new score on their existing interview — you do not need a new interview to test scoring changes. Think of the scale in two parts: * **The floor (1–3):** genuinely weak answers — buzzwords, evasive, off-topic, refused. * **The middle and top (4–10):** relevant answers of varying quality. **Golden rule:** To make scoring less strict, raise the middle/top — never the floor. To make it stricter, tighten the middle/top. Leave the floor alone unless you specifically want weak answers to move. *** ### 2. Your main dial: "Overall scoring calibration" 90% of adjustments should happen in this one entry. It defines what each 1–10 band means. To shift strictness, you widen or narrow the upper bands. **To make scoring LESS strict** — broaden the high bands so more answers qualify: > 7–8: A solid, relevant, on-topic answer with real substance, even if brief, unpolished, or without a specific example or numbers. Most genuine, on-point answers belong here. *(Pulling "relevant but brief" answers up into 7–8 is what lifts scores.)* **To make scoring MORE strict** — raise the bar for high scores: > 7–8: A strong, relevant answer that includes a specific example or concrete supporting detail. General or brief answers belong in 5–6. > > 9–10: Reserved for answers with a specific example AND a clear outcome or result. **Rule of thumb:** Move the description of a typical "decent" answer between the 5–6 band (stricter) and the 7–8 band (more lenient). That single change moves overall scores the most. *** ### 3. Supporting levers (use only if the main dial isn't enough) | Entry | Direction it pushes | When to use it | | ------------------------------------------------- | ----------------------------- | ---------------------------------------------------------------------------- | | Overall scoring calibration | Both (main dial) | First choice for any change | | Buzzword-only and evasive answers | Down (sets the floor) | Only to change how weak answers score | | Scripted or over-polished answers | Down | If good, articulate candidates score too low — soften it | | What earns a strong score | Up | To credit more kinds of "grounding" (details, reasoning, not just examples) | | Calibrate to the role's level | Context | If entry-level candidates are judged too harshly | | When question type conflicts with its rubric | Up (non-behavioral questions) | If motivation/values/"what you'd bring" questions score low for "no example" | | Don't penalize depth interviewer didn't probe for | Up | If brief answers are punished when the bot never followed up | *** ### 4. Copy-paste examples **Make it less strict (lenient calibration):** ``` Apply a generous calibration. A solid, relevant, on-topic answer with real substance belongs in 7–8 even if brief, unpolished, or without a specific example or numbers. Reserve 9–10 for clearly grounded answers (example, concrete details, or strong reasoning) — perfection and metrics are not required. Only genuinely weak answers (buzzword-only, evasive, off-topic, refused) score 1–3. Do not default overall fit downward. ``` **Make it stricter (demanding calibration):** ``` Apply a rigorous calibration. Reserve 9–10 for answers with a specific example AND a clear outcome or result. 7–8 requires a concrete supporting detail or example. General, brief, or unsupported answers belong in 5–6. Weak, evasive, or empty answers score 1–3. Do not inflate scores for fluency or confidence alone. ``` **Soften the floor slightly (if weak-but-relevant answers score too low):** ``` A relevant, on-topic answer with real substance is not a bottom-scoring answer even if brief — score it mid-range. Reserve 1–3 only for buzzword-only, evasive, off-topic, or refused answers. ``` *** ### 5. Do's and don'ts * **Do** change one entry at a time, then reanalyze so you know what caused the shift. * **Do** keep a known weak candidate and a known strong candidate as reference points. After each change, reanalyze both. * **Don't** raise the floor when trying to be more lenient. Keep the 1–3 band wording (buzzword/evasive/refused) intact — otherwise weak candidates climb too. * **Don't** stack many new entries. Edit the calibration dial rather than adding more; overlapping entries dilute each other. * **Remember:** a question's own Evaluation Criteria / Score 1 / Score 5 override the Knowledge Base. If one specific question scores wrong while others are fine, fix it on that question — not in the Knowledge Base. See [Configuring Your AI Interviewer](../../getting-started/log-in/configuring-your-ai-interviewer) for how to update per-question criteria. *** ### 6. Quick verification checklist After any change: 1. Reanalyze a strong candidate. Did they move in the intended direction? 2. Reanalyze a weak candidate. Did they stay low? If they jumped, the floor was affected — undo the floor change. 3. If both look right, the change is good. If not, adjust the calibration band wording and re-test. *** ## Knowledge Base (Interview Evaluation) — Reference prompts These are the full prompt texts for each entry. Copy them into your Knowledge Base entries with **Scope = "Interview Evaluation"** and edit the wording as needed using the guide above. ### Overall scoring calibration ``` Apply a moderately generous calibration. The interviewer's scoring has been running too strict, so anchor the scale so that a solid, relevant answer lands comfortably in the upper-middle, and reserve low scores for genuinely weak answers. Per-question band guide (1–10), unless a question's own evaluation criteria explicitly define a different bar: 9–10: Directly answers the question with clear, grounded substance — a real example, concrete accurate details, or clear relevant reasoning. Does NOT require perfection, metrics, or a polished STAR structure. 7–8: A solid, relevant, on-topic answer with real substance, even if brief, unpolished, or without a specific example or numbers. Most genuine, on-point answers belong here. 5–6: On-topic but thin or generic — some relevance but limited substance or specificity. 3–4: Largely off-target, vague with little relevance, or unsupported claims only. 1–2: Buzzword-only, evasive, off-topic, refused, or shows no real understanding. This calibration raises the middle and upper of the scale; it does NOT raise the floor — genuinely weak, evasive, or empty answers stay at 1–3. When judging overall fit (match_score), let it reflect these calibrated question scores rather than defaulting downward. ``` ### Scripted or over-polished answers ``` Being articulate, fluent, well-prepared, or polished is NOT a negative — do not lower a score simply because an answer sounds smooth or well-structured. Judge the substance behind it, not the delivery. Only be cautious when a polished answer is also hollow: generic phrasing that could apply to any candidate or any company, with no specific details, genuine reasoning, or personal grounding. In that case, don't award a high score for delivery alone — score it on the limited substance actually present. A polished answer that includes concrete details, accurate specifics, genuine reasoning, or a real example should be scored on its merits like any other strong answer. Where a question's own evaluation criteria or grading instructions say otherwise, follow those. ``` ### Calibrate to the role's level ``` Judge each answer against what is reasonable for the level and requirements of the role being interviewed for, as reflected in the question's own evaluation criteria. Do not hold a junior or entry-level candidate to a senior standard — for example, don't require deep strategic leadership, large-scale ownership, or quantified business impact where the role does not call for it. Likewise, don't over-credit a senior candidate merely for clearing a junior bar. When unsure of the level, judge whether the answer would be considered solid by a reasonable hiring manager for that specific role. ``` ### Don't penalize depth the interviewer didn't probe for ``` Score the substance the candidate actually provided, not depth the interviewer never asked for. If a question was not followed up and the answer is on-topic and relevant but brief, evaluate what was said on its merits rather than penalizing the candidate for not elaborating on something they were never prompted to expand on. A lack of follow-up by the interviewer is not evidence against the candidate. ``` ### When question type conflicts with its rubric ``` A question's generated evaluation criteria or Score 1/5 anchors sometimes ask for things that do not fit the question type — for example, demanding a "specific example", "STAR structure", or "measurable outcome" on a question that is really about motivation, company fit, role excitement, self-awareness, or the strengths a candidate would bring. When this happens, prioritise the question type over the rubric wording: For non-behavioral questions (motivation, culture/values fit, why-leaving, role excitement, development areas, experience overview, and "what value/strengths would you bring"), do NOT lower the score solely because the answer lacks a past example, a STAR story, or numbers. Treat any such requirement in the rubric as not applicable, and score on relevance, specificity, genuineness, and role-relevant substance instead. Only behavioral/situational questions ("Describe a time when...", "Tell me about a time...", "Give an example of...") genuinely require a real past example, and only those should be penalized for hypothetical-only answers. This does not lower the floor: buzzword-only, evasive, off-topic, or refused answers still score low on every question type. The goal is to stop type-mismatched rubric wording from capping otherwise strong, relevant answers. ``` ### Buzzword-only and evasive answers ``` Reserve the bottom of the scale for answers that are genuinely weak: buzzword-only answers (e.g. "ownership," "accountability," "growth mindset") with no substance behind them, evasive or off-topic answers, refusals, or answers showing no real understanding. For these, score low and note what was missing (e.g. "answer lacked any specific substance"). For behavioral/experience questions, if a candidate's example describes a related but different situation than what was asked, lower the score and note the mismatch. A relevant, on-topic answer that addresses the question with real substance is NOT a bottom-scoring answer even if it is brief or lacks a polished example — score it in the mid-range, and reserve low scores for the cases above. ``` ### What earns a strong score ``` Use the full scale and score confidently when an answer is grounded in the substance the question calls for. Grounding can take several forms — a specific real example, concrete and accurate details (named brands, products, values, events, actions taken), a described outcome or result, or clear and relevant reasoning — plus a direct connection to what was asked. A quantified outcome (metrics, percentages, volumes, timeframes) strengthens an answer where the question is about measurable results, but is not required for a strong score, and its absence should not cap an otherwise specific, relevant answer. Do not hold back high scores for answers that earn them. Defer to a question's own evaluation criteria when it defines what a strong answer looks like. ``` *** ## Next steps * [Knowledge Base](knowledge-base) — How to create and manage Knowledge Base entries * [Configuring Your AI Interviewer](../../getting-started/log-in/configuring-your-ai-interviewer) — Per-question criteria and scoring weights * [How Scoring Works](../../getting-started/log-in/how-scoring-works) — How overall scores are built and what they mean * [Interview Evaluation](interview-evaluation) — Language proficiency rubrics