# Grounded interview notes

> Emoti's interview notes follow a template the team chooses, and every section, highlight and answer points back to the moment in the transcript it came from. A language model writes the note when one is configured; otherwise a rules engine assembles it from transcript quotes, and the app labels which one answered.

Published by Emoti Research on 2026-11-01, updated 2026-11-01. Research paper, version 1.0. For hiring managers. Canonical: https://emoti.ai/research/grounded-interview-notes

## Abstract

Emoti writes interview notes in the shape of a template the team chooses, and ties each note back to the transcript: sections list the moments they draw on, answers to questions quote the transcript with times, and every time opens the transcript at that moment. When a language model is configured, it writes notes and answers under instructions to use only what the transcript says, and Emoti discards any reference that doesn't match a real turn. When no model is available, a deterministic rules engine assembles notes and answers from existing sections and transcript quotes, and the interface says which one answered. This paper describes templates, both writing paths, the Ask feature, how quotes link to moments, how people edit and restore notes, and the failure modes we know about, including references that point to the wrong turn, paraphrase that word matching misses, and transcription errors that no later step can fix.

## What does grounded mean in an interview note?

A note is grounded when every statement in it can be traced to a moment in the conversation, and a reader can reach that moment in one click.

Interview notes are read by people who weren't on the call: hiring managers, panel members, sometimes a reviewer months later. A fluent note that can't be checked is worse than a short one that can, because it invites a debrief to argue from memory. Research points the same way. In one study, taking notes improved how accurately interviewers recalled what candidates said but not the accuracy of their judgments, while reviewing conventional notes did improve judgment (Middendorf & Macan, 2002). US federal guidance on structured interviews asks for notes good enough to document the reasoning behind each rating, with actual examples of answers (US Office of Personnel Management, 2008). Notes are the record that judgment is checked against.

Emoti's notes follow five rules:

1. Write only from the transcript.
2. Point to the moments each section draws on.
3. Say plainly when a topic wasn't discussed.
4. Label text that Emoti wrote.
5. Keep a person's edits visible and reversible.

## How do templates shape a note?

A template is an ordered list of sections, each a heading plus a one-line instruction, and every note has exactly those sections in that order.

Emoti includes templates for a recruiter screen, a technical interview, a portfolio review, a behavioral interview, a hiring manager intake, a debrief, a reference check, a client call, a one-page candidate brief, and a plain list of questions and answers. Teams can write their own. The recruiter screen template looks like this:

| Section | Instruction Emoti follows |
| --- | --- |
| Why they're looking | What is prompting a move now, in their words. Note anything that would make them stay. |
| Current role | Scope, team, and the most relevant work for this role. Include numbers they gave. |
| Fit with the brief | Compare against the role's must-haves. Mark each as clear, partial or not discussed. |
| Compensation and logistics | Salary expectations, notice period, location, visa, other processes and timing. |
| Next steps | What was agreed and who owns it. |

The template decides what the note pays attention to. That is useful, because it puts every candidate's evidence in the same places. It is also a responsibility: a section that asks for concerns will tend to find some. Choosing a different template for a recorded call rewrites the note under the new sections.

## How does Emoti write a note?

With a model configured, a language model writes the note from the full transcript and Emoti checks its references. Without one, a rules engine rebuilds the note from quotes. Either way, the note has the template's sections in the template's order.

### With a language model

Emoti sends the model the template and the transcript, with a start time on every turn, and asks for one section per template section, each listing the start times of the turns it drew on. The instructions are short: use only what the transcript says; prefer one fact per line, with specific numbers and short quotes; and if the call didn't cover a section, write one line saying it wasn't discussed. The reply must fit a fixed structure of heading, body and references.

Emoti then checks the references. Any time that doesn't match the start of a real turn is dropped before the note is saved. If the model declines, returns nothing usable, or the request fails, Emoti uses the rules engine instead. Requests go through Emoti's model gateway, which runs them on the models a workspace admin assigns to each task and tries the next one in order if a model fails; the method doesn't depend on which model answers.

### Without a model

The rules engine works on words. It lowercases the text, drops short and common words, and compares what is left. For each template section:

1. It counts how many words from the section's heading and instruction appear in each existing note section. If the best match shares at least two words, that text and its references are reused under the new heading.
2. If no section is close enough, it picks up to three transcript turns that share the most words with the section, puts them in time order, and quotes each with the speaker's name and time, cut at 180 characters.
3. If no turn shares any words, the section says "Not discussed in this call."

When the rules engine rebuilds a note, the app says so. Some edits also need a model: a section can be shortened or turned into bullets without one, but adding detail and translating can't be done by rules, and the app says that too.

### What a grounded section looks like

Here is how a "Fit with the brief" section might read for a backend role whose must-haves are six years of backend work, production Go and on-call experience. The example is invented.

- Six years building backend services: clear. "Eight years, the last four on the payments platform." (3:12)
- Production Go: partial. Uses Go for internal tools; the services they own are in Java. (7:45)
- On call for a service they helped build: not discussed.

The section ends with a "From the call" row showing 3:12 and 7:45, and each time opens the transcript at that moment. The line that says "not discussed" matters as much as the others. It tells the panel what the next interviewer should ask, instead of leaving a gap that reads as a pass.

## How does Ask answer questions about a call?

Ask answers one question about one call by quoting the transcript, with a time after each point, and labels every answer with what produced it.

With a model, the question and the timestamped transcript go to the model with instructions to use only the transcript, quote short phrases, put the time after each point, stay under 150 words, and, if the call didn't cover the question, say so in one sentence and suggest what to ask next time. The answer appears as it is written. Each question is answered on its own; earlier answers stay on the page but aren't sent as context.

Without a model, the rules engine answers in one of four ways:

- **Questions about concerns, strengths or motivation** are answered from the call's highlights of that type, when it has any, each with its title, time and quote.
- **Questions about what the candidate asked** return the turns from people outside the hiring team that contain a question mark.
- **Other questions** become a word search. Common words and participants' names are ignored, and a few groups of related words are expanded, so "salary" also looks for "compensation", "base" and "equity". Words match on their first five letters, so "relocate" finds "relocation". Turns are scored by shared words with a small preference for the candidate's own answers, and the best match is kept along with anything at least 60 percent as good, up to two. When a match is the interviewer's question, the reply that follows it is added. At most four turns are quoted.
- **If nothing matches**, the answer says so and points to the call's first highlights.

Each answer is labeled "Answered by Emoti" when a model wrote it, or "Matched from the transcript by Emoti's built-in rules" when the rules engine did.

## How does a reader check and change a note?

Every time in a note, a highlight or an answer is a link to the transcript, and every section can be edited and restored.

### Links to moments

A time written as minutes and seconds in square brackets becomes a small button. Pressing it switches to the transcript, scrolls to the turn that was under way at that second, and highlights it. Note sections end with a "From the call" row of times taken from their references. Highlights carry a type (strength, concern, question, logistics or motivation), a short title, a quote and a time.

The transcript can be searched: matching turns stay bright, the rest are dimmed, and the number of matches is shown. Beside it, a map of who spoke when and the share of talk time on each side show the shape of the call.

### Edits and sharing

A section can be edited by hand, shortened, rewritten as bullets, given more detail or translated. An edited section is marked "Edited", and "Restore what Emoti wrote" brings back the original. Summaries and answers that Emoti wrote carry a label that says so. A note can be shared by link: the shared page shows the summary, notes and highlights, while the transcript and recording stay private to the workspace. A note can also be copied in one step for pasting into an applicant tracking system.

### Recording disclosure

The notetaker is set to tell people they are being recorded. With disclosure on, which is the default, the calendar invite says the call is recorded and transcribed and what the notes are used for, the notetaker joins under a visible name, and it posts a message in the call chat when recording starts. If anyone asks it to stop, it is designed to leave and delete the recording. A team can turn disclosure off, and the settings page then warns that they are responsible for telling people and, in many places, for getting their consent first.

## Limitations

Grounding narrows what can go wrong. It doesn't remove it.

- **Transcription errors flow into everything.** If a word is misheard or a turn is given to the wrong speaker, the note, the quotes and the answers inherit the error. Speech recognition is not equally accurate for everyone: in a study of five commercial systems, the average word error rate was 0.35 for Black speakers and 0.19 for white speakers (Koenecke et al., 2020).
- **A reference can be real and still wrong.** Emoti checks that each reference points to a turn that exists. It doesn't check that the turn supports the line it is attached to.
- **Quotes in model-written text are not checked word for word.** A quote can be shortened or paraphrased, and a model can in principle invent one. A wrong time opens whatever turn was under way at that second.
- **Summaries can lose hedges.** "I think we cut it by about a third" can become "cut it by a third". The instruction to quote short phrases reduces this; reading the moment removes it.
- **Word matching misses meaning.** The rules engine finds turns that share words with the question. It misses paraphrase, can pick a turn that mentions a word in passing, and can reuse an existing section under a heading it only loosely fits, because two shared words are enough.
- **"Not discussed" means "not found".** Both paths can miss a topic that was covered in other words.
- **Notes record what was said, not whether it is true.** A grounded note can faithfully quote an inaccurate claim.

## What we have not measured yet

We have not yet measured:

- How often each line of a model-written note is fully supported by the turns it references, by call type and by language.
- How often the rules engine picks the turns a person would pick for the same question.
- How transcription accuracy varies by accent, language and audio quality in real calls, and how those errors change the notes.
- Whether debriefs that use grounded notes reach different decisions, or reach them with more agreement.
- How well translated sections keep the meaning of the original.

These are the measurements the method needs before anyone should rely on a note more than on their own reading of the transcript.

## Key terms

- **Notetaker**: Software that joins a call, records and transcribes it, and drafts notes. Emoti's notetaker writes notes in the template chosen for each kind of call and, by default, tells everyone the call is recorded.
- **Transcript highlight**: In Emoti, a marked moment in a call's transcript with a type (strength, concern, question, logistics or motivation), a short title, a quote and a time that opens the transcript at that moment.
- **Evidence**: In Emoti, the line of profile or transcript text behind each judgment, shown next to the judgment so a person can check it.
- **Debrief**: The meeting after interviews where interviewers share their ratings and evidence and the hiring team decides. It works best when every interviewer has recorded ratings independently beforehand.

## Questions

### Does Emoti make hiring decisions from interview notes?

No. Notes, highlights and answers are records for people to read. Every hiring decision is made by a person.

### Can a team write its own note template?

Yes. A template is a list of section headings with one instruction each. Personal templates appear beside the built-in ones, and choosing one rewrites the note for that call.

### Which languages can Emoti's notes be written in?

The notetaker can write notes in the language spoken most on the call or in a chosen language, including English, Mandarin Chinese, Malay, Hindi, Japanese, German, French, Spanish, Portuguese and Polish. Translating a single section needs a language model.

### Why keep a rules engine when a model writes better notes?

So every feature still works without a model, and so a team can see exactly how an answer was found. Rules answers only quote the transcript, and they are always labeled as coming from rules.

## Sources

- Middendorf, C. H., & Macan, T. H. (2002). Note-taking in the employment interview: Effects on recall and judgments. Journal of Applied Psychology, 87(2), 293–303. https://doi.org/10.1037/0021-9010.87.2.293
- US Office of Personnel Management (2008). Structured Interviews: A Practical Guide. US Office of Personnel Management. https://www.opm.gov/policy-data-oversight/assessment-and-selection/structured-interviews/guide.pdf
- Koenecke, A., Nam, A., Lake, E., Nudell, J., Quartey, M., Mengesha, Z., Toups, C., Rickford, J. R., Jurafsky, D., & Goel, S. (2020). Racial disparities in automated speech recognition. Proceedings of the National Academy of Sciences, 117(14), 7684–7689. https://doi.org/10.1073/pnas.1915768117