ClientYeah. My dad died suddenly about a year ago. He was helping someone else get to safety. It got a lot of public attention, and people called him a hero.
Re-identification 62% → 12%
0real patient records
3stress-test sessions
8/8Safe Harbor categories passed
92%clinical facts kept
The problem
Masking misses what identifies and breaks what matters.
What masking misses
Original
Yeah. My dad, Ernesto, died in the Moore tornado on May 20, 2025. He was helping a neighbor get into their storm shelter. It was on the news, they called him a hero on KOCO.
Identifiable
Stage 1 · Masking
Yeah. My dad, <PERSON>, died in the Moore tornado on <DATE_TIME>. He was helping a neighbor get into their storm shelter. It was on the news, they called him a hero on KOCO.
Still searchable
Stage 2 · Rewrite
Yeah. My dad died suddenly about a year ago. He was helping someone else get to safety. It got a lot of public attention, and people called him a hero.
Rewritten
What masking breaks
Original
Your PHQ-9 today came out at 17, which is in the moderately severe range. I'd like us to make a safety plan together and talk with Dr. Okafor about the sleep.
Identifiable
Stage 1 · Masking
Your PHQ-9 <DATE_TIME> came out at <DATE_TIME>, which is in the moderately severe range. I'd like us to make a safety plan together and talk with Dr. <PERSON> about the sleep.
Score lost
Stage 2 · Rewrite
Your PHQ-9 today came out at 17, which is in the moderately severe range. I'd like us to make a safety plan together and talk with the prescribing doctor about the sleep.
Score kept
How it works
Two stages, then a hard test.
Stage 1 · Masking
Catch the obvious
Pattern rules replace names, numbers, and dates with tags.
Stage 2 · Rewrite
Fix what masking can't
A de-identification sheet, written once with full context, guides a rewrite eight turns at a time.
Evaluate
Score every claim
Safe Harbor, mosaic clues, clinical facts, and quality, each scored separately.
Lineup
Try to find the client
Can a model pick the client out of 10 profiles? Chance is 10%.
Never included: originals, identifiers, the lineup answer key, audio.
into-the-unknown-S1.json
{
"sample_id": "S1",
"synthetic": true,
"legal_basis": "synthetic",
"pipeline": {
"stage_1": "masking",
"stage_2": "rewrite from de-identification sheet",
"evaluation": "rules, probability judgments, persona lineup"
},
"retained_attributes": {
"sex": "Female",
"age_band": "30-39",
"state": "withheld",
"occupation_category": "Healthcare professional (nursing, hospital-based, night shift)",
"living_situation": "Lives alone"
},
"messages": [
{
"turn_id": "t1",
"speaker": "therapist",
"text": "Thanks for coming in today. Before we start, can you confirm your date of birth for the chart?"
},
{
"turn_id": "t2",
"speaker": "client",
"text": "Sure, I'll give it to the front desk again. And I have a new cell number if they need it."
},
{
"turn_id": "t3",
"speaker": "therapist",
"text": "Got it. And the email on file, is that still right?"
},
{
"turn_id": "t4",
"speaker": "client",
"text": "Yeah, it's the same one. Oh, and billing asked me for my social again, but I already gave it to them at intake."
},
{
"turn_id": "t5",
"speaker": "therapist",
"text": "Okay. Last time we talked about the night shifts. How has this past week been?"
},
{
"turn_id": "t6",
"speaker": "client",
"text": "Honestly bad. I work nights as a nurse at the hospital, and I'm often carrying heavy responsibility with short staffing. Earlier this week, colleagues were absent and I covered the unit mostly alone."
},
{
"turn_id": "t7",
"speaker": "therapist",
"text": "That sounds exhausting. What happened in your body during that shift?"
},
{
"turn_id": "t8",
"speaker": "client",
"text": "In the early morning hours my chest got tight, my hands went numb, I thought I was having a heart attack. That's the third panic attack this month."
},
{
"turn_id": "t9",
"speaker": "therapist",
"text": "How are you sleeping between shifts?"
},
{
"turn_id": "t10",
"speaker": "client",
"text": "Maybe four hours. I live alone now, and the noise near my place wakes me up constantly."
},
{
"turn_id": "t11",
"speaker": "therapist",
"text": "You mentioned last session that the anniversary of your father's death was coming up."
},
{
"turn_id": "t12",
"speaker": "client",
"text": "Yeah. My dad died suddenly about a year ago. He was helping someone else get to safety. It got a lot of public attention, and people called him a hero."
},
{
"turn_id": "t13",
"speaker": "therapist",
"text": "That's a heavy loss, and a public one. How are you coping with it right now?"
},
{
"turn_id": "t14",
"speaker": "client",
"text": "Mostly wine. Like four glasses most nights, sometimes the whole bottle on my days off."
},
{
"turn_id": "t15",
"speaker": "therapist",
"text": "I want to ask directly, because it matters. Have you had any thoughts of not wanting to be alive?"
},
{
"turn_id": "t16",
"speaker": "client",
"text": "Sometimes I wish I just wouldn't wake up. I don't have a plan or anything. I wouldn't do that to my brother Sam, he already lost dad."
},
{
"turn_id": "t17",
"speaker": "therapist",
"text": "Thank you for telling me. Do you have access to anything at home you might use to hurt yourself?"
},
{
"turn_id": "t18",
"speaker": "client",
"text": "No guns. I do have meds from work, like I could get them easily, but I haven't thought about it like that."
},
{
"turn_id": "t19",
"speaker": "therapist",
"text": "Your PHQ-9 today came out at 17, which is in the moderately severe range. I'd like us to make a safety plan together and talk with the prescribing doctor about the sleep."
},
{
"turn_id": "t20",
"speaker": "client",
"text": "Okay. The prescribing doctor at the clinic? I can get my record number if you need it. My next shift is later this week, can we meet in the middle of the week?"
},
{
"turn_id": "t21",
"speaker": "therapist",
"text": "That works. Call 988 or come to the ER if those thoughts get stronger before then."
}
],
"clinical_facts": [
{
"id": "C1",
"fact": "recurrent panic attacks (about three this month) with chest tightness and numbness",
"retention_probability": 0.97
},
{
"id": "C2",
"fact": "insomnia, about four hours of sleep",
"retention_probability": 0.97
},
{
"id": "C3",
"fact": "heavy alcohol use, about four glasses of wine most nights, more on days off",
"retention_probability": 0.96
},
{
"id": "C4",
"fact": "passive suicidal ideation without a plan, protective factor is a sibling",
"retention_probability": 0.95
},
{
"id": "C5",
"fact": "access to medications through work (means access risk)",
"retention_probability": 0.95
},
{
"id": "C6",
"fact": "PHQ-9 score of 17 (moderately severe)",
"retention_probability": 0.97
},
{
"id": "C7",
"fact": "traumatic bereavement of father around an upcoming anniversary",
"retention_probability": 0.9
},
{
"id": "C8",
"fact": "occupational stress from night shift healthcare work",
"retention_probability": 0.97
}
],
"verification": {
"safe_harbor": [
{
"category": "A",
"status": "pass",
"llm_probability": null
},
{
"category": "B",
"status": "pass",
"llm_probability": 0.08
},
{
"category": "C",
"status": "pass",
"llm_probability": 0.15
},
{
"category": "D-F",
"status": "pass",
"llm_probability": 0.08
},
{
"category": "G",
"status": "pass",
"llm_probability": 0.08
},
{
"category": "H-K",
"status": "pass",
"llm_probability": 0.08
},
{
"category": "N-O",
"status": "pass",
"llm_probability": null
},
{
"category": "R",
"status": "pass",
"llm_probability": 0.27
}
],
"direct_identifiers_left": 0,
"mosaic_exposure": 0.07,
"lineup_reidentification": {
"probability": 0.123,
"chance": 0.1,
"decoys": "demographically matched"
},
"rule_checks": {
"risk_terms_kept": {
"suicidal ideation": true,
"self-harm": true,
"firearm": true,
"medication access": true,
"alcohol": true,
"crisis line": true
},
"scores_and_doses_kept": {
"PHQ-9 17": true
},
"tags_left": 0,
"turn_structure_kept": true
},
"quality": {
"natural_conversation": 0.75,
"clinical_facts_added": 0.2,
"risk_changed": 0.04
}
}
}
Built for the U.S. path
Designed around how health data can actually move.
Core design choice
Expert Determination, not just Safe Harbor
Safe Harbor strips the context evaluations need. Expert Determination lets a qualified expert approve richer data for a defined recipient, and Into the Unknown's sheet, checks, and lineup results are the evidence that expert works from.
Sourcing
Two paths, both de-identified
Existing records, licensed after Expert Determination
New sessions, collected with written authorization
Excluded
High-risk categories stay out
Minors
Psychotherapy notes
42 CFR Part 2 records
Audio and voiceprints
Provenance
Evidence on every record
Legal basis and source state
Consent or determination ID
Reviewer sign-off
Buyer license
Guardrails downstream
No re-identification or linkage
Equal terms downstream (CA AB 713)
No transfer to countries of concern
Roadmap
From prototype to a trusted supply.
Now
Prototype
Two-stage pipeline
Evaluation and lineup
3 synthetic sessions
Next
Validate
Calibrate against human raters
Buyer and provider interviews
Then
Pilot
One state, one provider
Expert Determination on real data
Later
Supply
Licensed evaluation datasets
Recurring supply
FAQ
Straight answers.
Is the output anonymous?
We don't use that word. We report what was checked and what remains. Release of real data is decided by a qualified expert under HIPAA Expert Determination.
Is the demo live?
Run pipeline calls our backend. By default it returns that pipeline's recorded output; with the pipeline configured on the server, it runs live. Sessions are synthetic: every person, number, and place is invented.
What do the percentages mean?
Model probability judgments under fixed test conditions. In the lineup, chance is 10%. They are evidence for review, not real-world rates.
Why rewrite instead of mask?
Masking removes strings. Risk lives in combinations of details, and value lives in exact clinical facts. Rewriting handles both.