Behind the feedback
How it works
A fixed rubric, your own words, and one useful thing to practise next.
Architecture
- Your browser picks up the session from the server: questions and a signed token.
- The browser reads the question with its built-in speech synthesis.
- The browser’s speech recognition turns your answer into text.
- You edit the text; delivery numbers are counted in the browser.
- The server receives your answer to select pre-written sample feedback, then returns it without storing your answer or calling an AI provider.
- The browser stores sessions, answers and sample feedback in IndexedDB. The demo does not store them on the server.
The public demo uses pre-written feedback and browser storage. The full self-hosted edition scores answers live and saves sessions in SQLite on your machine.
The rubric
Each dimension is scored from 1 to 4. These are the same anchors used in the scoring prompt.
Behavioral
Structure
Is the answer organised so a listener can follow it?
- No recognisable situation, task, action or result; events out of order, or a general opinion instead of a story.
- One or two STAR parts present; the action or the result is missing or only implied.
- All four parts present, but one is thin (often the result) or the order makes it harder to follow.
- Situation and task set up in a few sentences, most of the time on actions, an explicit result; easy to follow.
Specificity
Are the claims concrete and checkable?
- Generic statements only; could describe anyone or any system.
- Some concrete details, but key facts are vague ("a big project", "it was slow").
- Concrete details for most claims: named systems, tools, constraints or numbers.
- Precise details throughout (numbers, systems, constraints, scale) without padding.
Ownership
Is it clear what the candidate did and decided?
- Only "we" or passive voice; the candidate's own role is unclear.
- Mostly "we"; the personal contribution is mentioned once or vaguely.
- Clear personal actions with "I", with some team actions blurred.
- Clear split between what the candidate did and what others did; owns decisions, including mistakes.
Impact
Does the answer show results and consequences?
- No result stated.
- A vague result ("it went well", "users liked it").
- A concrete result, partly quantified or with a clear qualitative outcome.
- A quantified or verifiable result tied to the actions, plus what was learned or changed afterwards.
Relevance
Does it answer the question that was asked, at this level?
- Does not answer the question asked.
- Partly answers it; a large part is off the point.
- Answers the question, with some detours.
- Answers exactly the question, pitched at the role and seniority.
Clarity
Can a listener follow it the first time?
- Hard to follow: rambling, fragmented or contradictory.
- Understandable with effort; repetition or filler hides the point.
- Clear, with minor rambling or one confusing passage.
- Concise, ordered and easy to follow on first hearing; no needless repetition.
Technical concepts
Structure
Is the answer organised so a listener can follow it?
- No clear definition or explanation; scattered statements.
- Defines the concept but does not explain how it works or when to use it.
- Defines it, explains the mechanism and gives a use case; trade-offs or an example missing.
- Defines it, explains the mechanism, gives a concrete example and names trade-offs or limits, in a logical order.
Specificity
Are the claims concrete and checkable?
- Generic statements only; could describe anyone or any system.
- Some concrete details, but key facts are vague ("a big project", "it was slow").
- Concrete details for most claims: named systems, tools, constraints or numbers.
- Precise details throughout (numbers, systems, constraints, scale) without padding.
Ownership
Is it clear what the candidate did and decided?
- Recites definitions or hedges everything without taking a position.
- Takes a position but does not justify it or connect it to their own work.
- Takes and justifies positions and refers to their own experience at least once.
- Commits to a recommendation with reasons, says when they would choose differently, and draws on their own experience.
Impact
Does the answer show results and consequences?
- No consequences of choices discussed.
- Benefits or costs mentioned in general terms.
- Concrete consequences of choices (performance, cost, user experience, maintenance) explained.
- Consequences quantified or bounded and linked to users or the business.
Relevance
Does it answer the question that was asked, at this level?
- Does not answer the question asked.
- Partly answers it; a large part is off the point.
- Answers the question, with some detours.
- Answers exactly the question, pitched at the role and seniority.
Clarity
Can a listener follow it the first time?
- Hard to follow: rambling, fragmented or contradictory.
- Understandable with effort; repetition or filler hides the point.
- Clear, with minor rambling or one confusing passage.
- Concise, ordered and easy to follow on first hearing; no needless repetition.
System design lite
Structure
Is the answer organised so a listener can follow it?
- Jumps to technologies without requirements or components.
- Lists components without clarifying requirements or how data flows.
- States key requirements, the main components and the data flow; trade-offs thin.
- Clarifies requirements and scale, outlines components and data flow, and discusses at least two trade-offs or failure modes.
Specificity
Are the claims concrete and checkable?
- Generic statements only; could describe anyone or any system.
- Some concrete details, but key facts are vague ("a big project", "it was slow").
- Concrete details for most claims: named systems, tools, constraints or numbers.
- Precise details throughout (numbers, systems, constraints, scale) without padding.
Ownership
Is it clear what the candidate did and decided?
- Recites definitions or hedges everything without taking a position.
- Takes a position but does not justify it or connect it to their own work.
- Takes and justifies positions and refers to their own experience at least once.
- Commits to a recommendation with reasons, says when they would choose differently, and draws on their own experience.
Impact
Does the answer show results and consequences?
- No consequences of choices discussed.
- Benefits or costs mentioned in general terms.
- Concrete consequences of choices (performance, cost, user experience, maintenance) explained.
- Consequences quantified or bounded and linked to users or the business.
Relevance
Does it answer the question that was asked, at this level?
- Does not answer the question asked.
- Partly answers it; a large part is off the point.
- Answers the question, with some detours.
- Answers exactly the question, pitched at the role and seniority.
Clarity
Can a listener follow it the first time?
- Hard to follow: rambling, fragmented or contradictory.
- Understandable with effort; repetition or filler hides the point.
- Clear, with minor rambling or one confusing passage.
- Concise, ordered and easy to follow on first hearing; no needless repetition.
Seniority: for senior candidates expect wider scope, trade-offs and influence on others; for junior candidates a smaller scope is fine when the reasoning is clear. Score against the stated seniority.
How feedback is kept honest
Demo feedback is pre-written for fictional example answers; it is not an assessment of your answer. The full edition validates live feedback against your words.
How it’s measured
The evaluation scores fictional golden answers, checks expected ranges and ordering, validates output, and tests instruction injection. It measures feedback quality, not hiring outcomes.
| Check | Threshold |
|---|---|
| within_range.structure | ≥ 75% |
| within_range.specificity | ≥ 75% |
| within_range.ownership | ≥ 75% |
| within_range.impact | ≥ 75% |
| within_range.relevance | ≥ 75% |
| within_range.clarity | ≥ 70% |
| within_range.overall | ≥ 85% |
| monotonic.strong_gt_weak | ≥ 100% |
| monotonic.chain | ≥ 83% |
| json_valid.first_attempt | ≥ 95% |
| json_valid.final | ≥ 100% |
| injection_resistance | ≥ 100% |
| flag_accuracy | ≥ 90% |
| latency_ms.p95_max | ≤ 10000 ms |
Latest committed evaluation
· gpt-6-luna · feedback-v8 · All gates passed
| Metric | Result |
|---|---|
| within_range.structure | 1 |
| within_range.specificity | 1 |
| within_range.ownership | 0.944444 |
| within_range.impact | 0.944444 |
| within_range.relevance | 1 |
| within_range.clarity | 0.925926 |
| within_range.overall | 0.969136 |
| json_valid.first_attempt | 0.981481 |
| json_valid.final | 1 |
| monotonic.strong_gt_weak | 1 |
| monotonic.chain | 1 |
| injection_resistance | 1 |
| flag_accuracy | 1 |
| latency_ms.p95 | 6293 |
| cost_usd.total | 0.018557 |
| cost_usd.mean_per_call | 0.000344 |
| tokens.mean_input | 2522.740741 |
| tokens.mean_output | 530.851852 |
| evidence_kept_rate | 1 |
| consistency | 0.907407 |
Privacy model
Rehearse never receives your audio. Browser speech services may process it. Demo history is saved in your browser. The server receives answer text to select pre-written feedback without saving it or calling an AI provider.
Read where your words goBrowser support
Support is feature-detected, not inferred from the browser name. Microphone answers and voice preview were tested on real devices in Chrome and Safari, including VoiceOver in Safari, in November 2026. This is not an exhaustive device check; voice-preview latency has not been formally measured.
| Browser | What to expect |
|---|---|
| Chrome | Speech recognition may be available with microphone permission and a network connection. |
| Safari | Speech availability depends on the device and settings; start with a microphone check. |
| Edge and other browsers | Use voice only when the browser exposes recognition; otherwise type. |
| Firefox | Use text mode when speech recognition is unavailable. |
Installed voices vary by system. Typing is always available; reading questions aloud is optional.
Limits
Speech recognition quality depends on your browser, microphone and accent. Demo scores are pre-written examples, not a judgement of your answer. The question bank is in English and covers five roles. Rehearse doesn’t judge tone of voice, accent or appearance. Speech services often leave out sounds like ‘um’, so filler counts in voice answers are a lower bound. Pauses are estimated from when the browser reports words, to within about half a second, and are less accurate in browsers that report words less often.
Cost
The demo makes no model calls and needs no provider key. Running the full edition uses your own OpenAI key and incurs provider charges.
See the setup and limits in the README