TAC×Learning Commons

We scored a teacher's voice the way you score text.

Seven Learning Commons Feedback Evaluators, run against a transcribed 65-second voice note. Same teacher, same student, two takes.

Play the voice note ↓ What's real here
The instrument

One student response. Two takes.

Press play. The transcript is what Whisper heard; the profile below is what Learning Commons' rubric made of it.

0:00 / 0:00
The student wrote

Productive Coaching profile
    The finding

    The three that failed are not random.

    They are the failure modes speech has and writing doesn't.

    01 · MANAGEABILITY

    Talking is cheap

    Five separate issues in sixty-five seconds. Typing that costs ten minutes, so nobody types it. Voice makes over-feedback nearly free — and the student still can't act on all five.

    02 · WITHHOLDING ANSWERS

    Thinking aloud leaks answers

    Mid-sentence, the teacher recites the exact statistic the student was supposed to go find. You don't do that when you type. You do it constantly when you talk.

    03 · TONE

    Prosody doesn't survive

    Flagged on four words: "you're doing that thing again." In the room, warmth in the voice carries it. In a transcript, warmth is gone and the rubric sees a comment aimed at the student.

    The transcript is not the feedback.

    Provenance

    Exactly what is real.

    This is a prototype built for one conversation. It is worth being precise about which parts are Learning Commons' and which parts we invented.

    Evaluators
    realLearning Commons' published prompts, used verbatim. All 14 system.txt/user.txt files were sha256-verified byte-identical against each evaluator's own config.json manifest in learning-commons-org/evaluators (MIT).
    Output contract
    realEach evaluator's published output_schema.json, enforced as a strict structured-output schema. The scores, reasoning, key features and proposed adjustments below are returned verbatim by the model — not written by us.
    Model
    realgpt-5.4-2026-03-05 at temperature 1 — the configuration Learning Commons validated against, read from their config.json.
    How it was called
    realDirect OpenAI chat/completions calls using LC's published artifacts. This did not route through the LC-hosted SDK or playground — the SDK runs on the developer's own OpenAI key, so this is the same computation by a different path.
    Student text
    syntheticInvented for this demo. No TAC production data. No student PII. Learning Commons requires de-identified inputs.
    Teacher takes
    syntheticBoth notes were written for this demo to be representative of how teachers actually talk, then spoken by a text-to-speech voice — no real teacher was recorded.
    Waveform
    realRMS peaks computed from the actual audio files on this page.
    Cost & latency
    real$0.042 per voice note for all seven evaluators; ~5.4s wall clock running them in parallel. Measured across 14 calls.
    Run variance
    These evaluators run at temperature 1. Re-running the same input can flip a borderline dimension — we saw Take B score 7/7 and 6/7 on identical input. Learning Commons documents this: treat scores as directional signals, not verdicts. The results shown here are one captured run.
    Known limits
    Learning Commons documents these as Early Access, validated for grades 8–9 on short claim-with-evidence completions — though every config.json in the repo declares grades 6–12. Binary output. Not for grading or placement without human review.