Can a model of one
beat a model of everyone?

Does an emotion model adapted on a person’s own labelled diary entries predict that person’s later self-reported feelings better than a generic model trained on everyone — and how many entries does the adaptation take?

The question matters beyond one app. Consumer AI in 2026 personalises by centralisation: a large model trained on everyone, tuned to an individual by sending that individual’s data to it. The alternative is a model of one, built and kept on the person’s own device. That alternative is defensible today on ownership and privacy grounds alone, and I have argued it in a position paper. What has not been shown — by me or, as far as I can find, by anyone in this configuration — is the predictive claim: that a model fit to one person beats a model fit to a population, on that person, at a sample size a real user could plausibly produce.

That is an empirical question with a number attached, and it can come back negative. If personalisation does not help at 25 entries, that is a publishable result and it changes what the product should honestly claim.

The longer arc this belongs to

The question above is one falsifiable step, not the destination. What this platform is ultimately for is a model of a person that lives on that person’s own device — built from their own longitudinal record, owned by them, and useful to them as a mirror of what they have actually written rather than as an oracle that claims to be them. Every research decision here is subordinate to that: the on-device constraint, the falsifiable evaluation, the refusal to collect anything.

The reason to work on it now is that the substrate has stopped being hypothetical. A phone in 2026 ships a multi-billion-parameter language model, small trainable adapters, on-device speech recognition and sentence embeddings, and a neural accelerator measured in tens of trillions of operations per second. Assembling a genuinely complete personal model on the device itself looks less like a research fantasy each year and more like an engineering schedule. The open questions are moving from “can it run?” to “how would you know it is any good?” — and the second question is the one with no established answer, because a model built to be private cannot be watched in production.

That is the through-line for a collaboration worth more than a single pilot: how a person is represented on their own hardware, how that representation is measured without surveilling it, how it should change as the person changes, and what it must refuse to do. The cohort study is simply the first piece of it that is small enough to run this academic year.

A collaborator is not being asked to build infrastructure. Four things are already running, and the instrument has already been used to fail my own model in public.

• A shipping testbed with the measurement built in. DailyVox is a live iOS app that transcribes speech on-device with no network permission at all. Since v1.9 it ships a seven-class self-labelling control, so participants generate labelled ground truth as a normal part of journaling — not in a separate survey instrument that decays after week two.
• A falsifiable evaluation harness (TwinEval). Every metric is lift over an explicit baseline; label-permutation negative controls collapse to chance when the signal is destroyed (p = 0.000 against the null). The harness is the contribution, not just plumbing for the claim.
• A position paper with a discrimination study. A shared on-device model prompted to imitate the author is reliably separable from the author’s real writing (AUC 0.81); a model actually fit to the author is not (AUC 0.43) — evidence that prompt conditioning is not personalisation and that the path runs through on-device adaptation.
• Population-scale baselines on licence-clean public corpora. 1,473 writer-self-labelled entries, 15,809 held-out utterances, 1,307 writers’ openness self-reports. Reproducible without anyone’s diary.

Results so far, including the ones that went against me

Three of those four are the product losing. An instrument that cannot catch its own maker’s model is not an instrument, so the negatives are the evidence that the method has teeth.

The study exists as a complete pre-registration draft: hypotheses, split rule, power tables, analysis code, and a pre-written result sentence for every branch of the outcome tree, so no wording is chosen after the data is seen.

Power, stated before recruitment

At N = 15 the per-person win probability is 0.828 under a true +10-point effect, with a 2.9-point accuracy quantum and a 6.9% tie rate. Below eight participants the single most likely outcome is “we could not tell” — and that sentence is pre-written like every other, because a pilot that can only report success is not a pilot.

The anti-fabrication gate

A leave-one-participant-out donor arm is pre-registered as an interpretive gate: adapt on other participants, test on this one. If donor adaptation matches personalised adaptation, the personalisation claim is withdrawn and the finding is published instead as spoken-register domain adaptation. This exists because the generic head is trained on written-register corpora while diary entries are transcribed speech, so an apparent personalisation lift could be shared register adaptation with no person-specific signal in it at all. The study is publishable on either branch.

Prior art the design must answer

MoodScope (LiKamWa et al., MobiSys 2013) is the closest ancestor and the effect-size prior: N = 32 over two months, with mood accuracy rising from 66% to 93% after personalised training. It ran on phone-usage sensors rather than speech, but it is the same generic-versus-personal contrast, measured thirteen years ago. Han et al. (IMWUT 8(4), 2024) is the closest methods precedent: a systematic evaluation of personalised affect models on WESAD at N = 15, using leave-one-participant-out cross-validation explicitly to correct earlier within-participant evaluation — which is the same reasoning behind our donor arm.

Two lineages are easy to overclaim and we do not. LEIA (Aroyehun et al., EPJ Data Science 2023) is a five-class model, so our seven-class label set does not map onto it without a documented collapse; its real value here is its held-out-user and future-posts splits, which are direct precedent for our design. Meta-PerSER (Interspeech 2025) personalises to the annotator rather than the speaker, so it is a weaker ancestor than its name suggests.

The novelty claim is therefore a conjunction rather than a single new idea: on-device, diary-register transcribed speech, self-labelled ground truth generated in normal use, consent-first with no egress, and falsifiable by construction. One narrower claim does look genuinely unoccupied: we could find no prior work applying an exact sign-flip permutation test to per-participant Δaccuracy in an ML evaluation, nor one using leave-one-participant-out as an anti-fabrication gate. At N ≤ 15 the 32,768 sign assignments are exhaustively enumerable, so the test is exact rather than Monte-Carlo. A reviewer who thinks the conjunction is thin is exactly the reader this document wants.

Three blockers. None is technical, and that is the point — they are the specific things one unaffiliated person cannot supply, stated plainly so nobody discovers them late.

• No ethics process. A human-subjects study needs a real one. An independent researcher can write a self-assessment, and reviewers at IMWUT or CHI will notice it is not an IRB. This is the single largest gap and the one a university closes by existing.
• The consent document needs its second version before anyone enrols. The current form was written for a different, retired study: it promises data will “never be used to train anything,” which this study’s per-person head would violate, and it recruits for the wrong artifact. The in-app recruitment card is held closed until that is fixed. Zero participants have been enrolled and zero data collected, so the fix is free today and unfixable retroactively the moment it is not.
• Recruitment and review capacity. Eight to fifteen participants, each journaling to roughly 60 self-labelled entries over about two months. One person working evenings cannot recruit, consent, support and review a cohort; a supervised group can do it as a term project.

One further item is deliberately open rather than blocked: the pre-registration is still a draft. It becomes unamendable the moment it is frozen, so it stays unfrozen until someone qualified has attacked it. Reviewing it before the freeze is the highest-leverage thing a collaborator can do, and it has to happen first.

This is a list of probable targets, not a plan. Nothing here is committed to and the right subset depends on which framing the collaboration chooses (see the note on framing at the end of this section). Everything falls after December 2026, so all of it is still reachable. Dates marked verified were read off the venue’s own site in September 2026; where a 2027 call is not yet posted the prior year’s date is given and labelled as such. Nothing is interpolated silently, and anything load-bearing should be re-checked before it is relied on.

First, a format that fits this study better than any conference

Our own power analysis says the single most likely outcome at 8–14 participants is inconclusive. Every conference on the list below rewards results, which is precisely the wrong incentive for a study of this size. A Registered Report inverts that: the design is peer-reviewed and the paper is accepted in principle before data is collected, so an inconclusive result is publishable by construction rather than by luck.

We checked, and there is no registered-report track at any speech, NLP, affective-computing or ACM HCI venue — not CHI, not IMWUT, not IEEE TAC, not Interspeech. The live routes are outside those communities:

• PCI Registered Reports — the recommended route. Free, rolling, no word limit, and its policy states in terms that it “sets no minimum requirement for statistical power, Bayes factors, or the precision of parameter estimates” and welcomes non-NHST evidence thresholds. That sentence is the answer to the single biggest objection this study faces. Stage 1 acceptance is then portable to 41 friendly journals that commit to accepting a positively recommended manuscript without further peer review — so one Stage 1 review can be honoured by more than one journal. Review is open rather than blinded, and no parallel journal submission is permitted while it runs.
• Royal Society Open Science and PeerJ Computer Science are the two nominated outlets that fit best — the second being CS-native. RSOS carries the most small-N-tolerant language we found anywhere: its Stage 1 method must address statistical power, “although we recognise that classic power analysis may not be appropriate or possible for all research fields or designs.” PeerJ CS applies the same criteria and adds “power analysis where applicable”. Note RSOS excludes clinical research, so this is framed as measurement and ML evaluation rather than as a health study.
• The JMIR route, if MEDLINE indexing and a health audience matter more. Stage 1 goes to JMIR Research Protocols, which issues a registered-report identifier and in principle guarantees acceptance of the results paper “independently of whether the research turns out to be negative”; results then land in JMIR Formative Research, a journal created precisely because pilot and feasibility work is not published in outcomes-focused venues, and which sets no minimum sample size. It has already published a person-specific smartphone-and-mental-health study at N = 5 and an n-of-1 study fitting participant-specific models against a person-specific baseline — which is this study’s design, already in print. Roughly $4,400 in fees before waivers, and the one route here with a real cost attached; note also that JMIR’s fee-waiver form requires a department chair’s signature, which an unaffiliated researcher does not have.
• Avoid Collabra: Psychology as the nominated outlet. It requires power ≥ 0.90 for every pre-registered hypothesis test unless the submission reaches its Level-6 bias control, which is hostile to a cohort this size. Listed here so nobody discovers it late.
• One thing to settle before any Stage 1 freeze. PCI RR, RSOS and Frontiers all require a public data or code archive at Stage 2, and the introduction and methods freeze at in-principle acceptance. So the protocol has to state exactly what leaves the device — aggregate labels, per-person model weights, or nothing but summary statistics. The no-egress design is an asset at Stage 1 only if that archiving plan is written down before the freeze.

This is the single most useful thing a collaborator could push for, because the pre-registration already exists — the expensive part of a Registered Report is written. It also converts the honest weakness of the study into a structural strength. Worth noting what we did not find: no ACM or IEEE venue appears on the Center for Open Science’s participating-journals list at all, the sole exception being ACM Games: Research and Practice.

Probable venues, January 2027 onward

Four things worth knowing before choosing

• Interspeech has a real fit problem, not just a deadline. Our classifier reads transcripts, not the acoustic signal, and Interspeech’s paralinguistics community is built on the premise that affect lives in the signal — prosody, voice quality, timing. A reviewer will ask why the audio was discarded, and as currently framed the study has no answer. Adding even a modest prosodic-feature arm converts that objection into a modality comparison. Without one, this is the audience most likely to rule it out of scope.
• ACII 2027 cannot be planned around. It would be the best topical fit, but as of September 2026 it has no host city, no dates and no call — ACII 2026 closed on 10 September and the next host is customarily named at the closing session, so an announcement may be days away. Worth watching; not worth scheduling against. Note also that a predatory publisher advertises a similarly-named “ICACII” event; it is not ACII.
• The big ML conferences are the wrong room for the cohort study. An 8–15 participant study with no new architecture and no benchmark number reads as thin to NeurIPS and ICML reviewer pools. The companion methodology paper is a different matter: NeurIPS’s renamed Evaluations & Datasets track explicitly invites evaluation methodology and negative results — but it requires public data release by camera-ready, which diary audio cannot satisfy. That track is reachable only with a synthetic benchmark or the harness alone.
• The same fact reads as a contribution or as a defect, depending on the room. Zero network egress is the point at PoPETs, ACM HEALTH and PervasiveHealth, where architectural privacy is a result. At CHIL, ML4H and ACM MM the identical property reads as “no public dataset” and counts against the paper. This is the one thing on this page that should be decided before a word is written, because it changes the framing rather than the findings.

The timing still works. A cohort recruited in October, journaling for about two months, freezes data around late December — which puts analysis in January, in range for CHI Posters on 21 January, UMAP on 29 January and IMWUT on 1 February. Pursuing a Registered Report in parallel costs nothing, because it uses the pre-registration that already exists.

There are two separable papers here, which matters for a student’s workload: the methodology and harness paper (already largely written) and the cohort study paper (the one that needs people). They can go to different venues, and the second has the cleaner student-sized scope. For the methodology paper, ACM TiiS looks like the closest scope match found — rolling, no page limit, fully open access, and its call names personalisation, affect-aware interfaces and evaluation frameworks together; it desk-rejects work that addresses only intelligence or only interaction, which this does not. ACM TOCHI supports a more ambitious option worth considering: publish both as a single long paper — the method, with the cohort study as its demonstration — which suits TOCHI’s depth norm and carries no conference deadline at all.

Shapes this can take

• The study may return a null, or an “inconclusive”. At 8–14 participants, inconclusive is the single most likely outcome, and a +6-point true effect is essentially undetectable at this scale. Every branch has a pre-written sentence and every branch is publishable, but nobody should enter this expecting a confident positive. The mitigation is structural rather than statistical: a Registered Report (§05) is accepted on the design before the result exists, which is the only route that makes an inconclusive outcome publishable by construction. If we pursue one route beyond the cohort itself, it should be that.
• Novelty is a conjunction, not a first. MoodScope showed a personalisation curve in 2013. If a reviewer decides the conjunction is insufficiently new, that is a real risk and it is better argued now than discovered at submission.
• Attrition is the operational risk. Sixty labelled entries over two months is a genuine ask of a volunteer. The mitigation is that labelling happens inside an app people already use daily rather than in a separate instrument, but the cohort should be over-recruited.
• One author, so far. Everything above is a single person’s work with a single diary, which is the honest limitation of all of it — and the reason this document exists.

On ethics, settled before anyone asks: informed consent, no compensation, data kept local to the participant’s own machine, deletion honoured immediately, withdrawal honoured any time before publication, and aggregate numerics only in anything published. Those terms predate anyone asking and are not negotiable. Where a collaborating institution’s process is stricter, its process governs.

The full pre-registration draft — power tables, split rule, outcome sentences, consent draft and the decision log — is available on request and is the right thing to read next. The most useful first response to it is suspicion.

A thirty-minute conversation about whether the design holds is enough to decide whether any of this is worth an ethics application.

hello@getdailyvox.com →

In a hurry? The two-page brief is this document compressed to something printable. Background reading: the methodology paper (PDF), the research programme and open problems, and the system architecture.