Key takeaways
- Planned for the first release, alongside the live AI interview and the text interview.
- The AI reads the transcript only. It never sees the camera or the screen recording.
- Nine default criteria cover what a founder knows and can show, never how they sound or look.
- Every suggested score cites the exact transcript passage, and criteria without enough evidence say so.
- Analysts own the result: each criterion can be overridden, with a reason that stays on the record.
Planned feature: not in Delsivo yet, and details may change before it ships.
What the founder interview scoring rubric measures
The default rubric starts from what a first call is meant to find out, across both parts of the interview: the pitch and the answers to the AI interviewer's questions. Your fund can edit the criteria, their definitions and their weights, and each edit is saved as a new rubric version so older interviews keep the version they were assessed against.
| Criterion | What a strong answer shows |
|---|---|
| Problem insight | Specific knowledge of who has the problem and what it costs them |
| Solution and differentiation | Why this approach, and what others miss |
| Traction and metrics evidence | Numbers with definitions and periods, and how they were measured |
| Market understanding | Who buys, how many of them, and the assumptions behind the size |
| Go-to-market and sales motion | How customers are found and closed today |
| Moat and defensibility | What gets harder to copy over time |
| Financials and unit economics | Costs, margins, burn and runway the founder can explain |
| Team and execution | What the team has shipped and why it can do this |
| Ask and use of funds | What the round pays for and what it should prove |
What is deliberately left out
Delivery is not substance. The coaching engine Delsivo's scoring is adapted from also measured clarity, composure and speech pace. Those dimensions were removed for investor use. The assessment never looks at:
- The camera or screen recording: the AI is given the transcript only
- Accent or fluency in English
- Presentation polish, confidence or concision
- Appearance, the room or the background
- Words per minute, filler words or talk ratio, which are not computed for investors
Leaving these out does not make any assessment neutral by itself. It narrows what the AI is asked to judge, and people still make the call. People at your fund can watch the recording, and founders are told what is assessed before they start.
How a score is put together
- 1After the call, the founder's audio is transcribed again with segment timestamps and merged with the interviewer's questions, labelled by speaker. Low-confidence segments are flagged.
- 2The assessment reads your rubric version, the questions, the transcript of the pitch and the answers, the deck facts and open claims. Contact details are redacted from the copy it reads.
- 3For each criterion it returns a 0 to 10 score or no score, a rationale and citations to transcript segments, with a separate summary of the pitch and of the questions.
- 4Any quote that can't be found in the stored transcript is dropped, along with its citation.
- 5Code computes the weighted total from the scored criteria. Unscored criteria show as not enough evidence.
- 6New claims and claim updates it notices, such as 12 of 20 customers being unpaid pilots, go to the claim tracker as suggestions.
Reviewing and overriding
Suggested scores appear as AI findings, styled differently from anything an analyst has verified. Each one has the rationale and a link that plays the cited moment in the recording or highlights the text answer. The player has a bookmark for the pitch and for each question, and a view switch between screen, camera and both.
An analyst can accept, reject or override any criterion. An override needs a reason, which is kept with who made it and when. Team members can also add their own rating per criterion, and when the fund turns on independent review, each member's rating stays hidden from the others until they have rated.
How format fairness is checked
A founder who takes the text interview should not be marked down against one who pitches on camera with the same content. Before release, a paired set of video, voice-only and text interviews with equivalent content must receive assessments within a set tolerance. The same check runs as part of Delsivo's AI evaluations whenever the prompt or model changes.