# Clinical Trial OS — Product Requirements Document

Oct 1, 2026 · @Tuan · revised the same day after the CEO review ([ceo-review.md](ceo-review.md)) and the first thin build ([build-plan.md](build-plan.md))

Trial OS is an AI platform that designs, simulates and helps run clinical trials and studies. It learns how trials really behave in a country by ingesting that country's finished studies, then uses what it learned to design the next one. It starts in Vietnam and Southeast Asia, where Life AI has direct access to hospitals, doctors and government, and scales globally one country pack at a time.

## 1. Overview

Trial OS's edge is access. It can obtain finished studies, clinicians' judgement and regulators' feedback in Vietnam and Southeast Asia, and turn them into a model of local behaviour that global vendors do not have.

**Vision.** Every study a country runs makes its next study faster, smaller and more likely to succeed.

**Problem**

- **Designs assume the wrong patients and practice.** Protocols written from Western data misjudge local biology, usual care, adherence and site behaviour.
- **What a study learns is lost when it ends.** The lessons sit in reports, logs, review letters and people's heads, and the next team starts again.
- **Trials stall on operations.** Enrolment, amendments, deviations and approvals cause most delay.

Evidence for each is in Appendix A.

**Product.** Trial OS runs one loop: ingest a finished study, learn local behaviour from it, design and simulate the next study with that knowledge, monitor it as it runs, and feed the results back.

**Why Life AI**

| Access | What it gives the product |
| --- | --- |
| Hospitals | Finished study files and data; records for recruitment counts |
| Doctors and investigators | Review of what the system learned; practice knowledge no dataset holds; first users |
| Government | Review feedback, pilot dossiers and early regulatory precedent |
| Life AI's own studies in diabetes, autism and stroke | The first study packages to ingest |

**Staging**

1. **Vietnam.** Build the loop on Life AI's and partner hospitals' studies.
2. **Southeast Asia.** Repeat it in Thailand and Indonesia, one country pack each.
3. **Global.** Partners build country packs with the same playbook, and the engines are licensed.

**What is different.** Other vendors sell simulators, patient twins or operations agents built on US and EU data. None found in public sources learns site and investigator behaviour from a country's own past studies. Appendix D has the comparison.

## 2. Goals, non-goals and success metrics

The product succeeds if a finished local study becomes a reusable behaviour profile within days, and a new design built on that profile predicts its own trial well enough to plan on.

**Goals**

1. Turn any finished local study into a local behaviour profile within days.
2. Produce a locally calibrated, regulator-ready study design from a brief within a week.
3. Predict enrolment, retention and data quality with stated intervals that hold up.
4. Cut sample size and duration with validated statistical methods, without weakening the evidence.
5. Build a new country pack in three months wherever access exists.

**Non-goals**

- Data-capture, trial-management and trial-master-file systems. Trial OS integrates with them.
- Replacing randomised control arms with synthetic ones.
- Autonomous regulatory submission or automated unblinded decisions.
- Predicting whether a drug will work.
- Drug discovery.

**Success metrics**

| Metric | Target | By month |
| --- | --- | --- |
| Finished studies ingested | 1, then 10 | 3, then 12 |
| Time to ingest one study package | ≤ 10 working days, then ≤ 2 | 3, then 12 |
| Extracted behaviour parameters reviewed by a named investigator | ≥ 80% | 3 |
| Transport gap estimated against a baseline-adjusted analysis | Reported with patient count and interval | 9 |
| Replay on truncated data: coverage and width of 80% intervals for pre-listed quantities | Reported | 3 |
| Letters of intent from demos | 2 from 10 | 4 |
| Paid design engagements | 3 | 12 |
| Protocol and dossier drafting time | ≤ 5 working days | 12 |
| First dossier passes the ministry validity check at first filing | 1 | 12 |
| Live trial using a pre-specified Trial OS model | 1 started | 24 |
| Enrolment forecast error on live studies | Within ± 50%, then ± 25% | 18, then 24 |
| Second country pack build time | ≤ 3 months | 24 |
| Revenue from customers based outside Vietnam | ≥ 30% | 24 |

## 3. Users and use cases

Trial OS has seven kinds of user. The first to serve are hospital research units, which supply finished studies, and the two buyers: sponsors and government.

| User | Job to be done | What Trial OS gives them |
| --- | --- | --- |
| Hospital research unit or principal investigator | Learn from finished studies; write a fundable, approvable protocol | A retrospective of each study; a structured protocol and dossier draft |
| Sponsor clinical lead | Decide whether and how to run a study locally | A costed design with size, duration and risk, each with a range |
| Biostatistician | Choose and defend a design | Simulated operating characteristics and a draft analysis plan |
| Research-organisation project lead | Quote, plan sites, forecast enrolment | Site ranking and an enrolment forecast with intervals |
| Regulatory affairs | Clear site ethics, national ethics and the ministry | A dossier checked against local rules and past review comments |
| Ethics committee member or regulator | Judge the design | A plain-language rationale and the model's evidence file |
| Ministry science administration or research funder | Raise national trial capacity; see what publicly funded studies taught | An annual lessons report and a view of timelines across studies |

**Use cases**

| ID | As a | I want to | So that | Features |
| --- | --- | --- | --- | --- |
| UC-1 | Hospital research unit | Hand over a finished study's files and data | I get a retrospective and its lessons are kept | F1 |
| UC-2 | Investigator | Confirm or correct what the system learned about local practice | Designs reflect how we really work | F1.6 |
| UC-3 | Sponsor clinical lead | Turn a study brief into a design with realistic size, time and cost | I can decide and budget | F2, F3 |
| UC-4 | Biostatistician | Compare designs under local behaviour | I can defend the choice and the analysis plan | F3, F4 |
| UC-5 | Research-organisation project lead | Count eligible patients and forecast enrolment by site | My quote and timeline hold | F5 |
| UC-6 | Regulatory affairs | Draft a dossier that answers the comments reviewers usually raise | Approval takes fewer rounds | F6 |
| UC-7 | Study manager | See patient and site compliance problems early | I act before data is lost | F7 |
| UC-8 | Country lead | Build a pack for a new country from its finished studies | Trial OS works there in three months | F8 |
| UC-9 | Ethics committee member | See why a design was chosen and how far its models can be trusted | I can approve or challenge it | F6, F9 |
| UC-10 | Ministry or research funder | See what publicly funded studies taught, across studies | Policy and funding go where trials stall | F1, F1.7 |

**Study types.** Drug, biologic and vaccine trials come first. Device and new-technique trials follow, since Vietnamese law makes them mandatory for higher-risk devices. Observational and registry studies use the same ingest and design features, and need their own confounding-control methods later.

## 4. How it works

Trial OS is a learning loop around three engines. Each finished study is ingested, turned into a local behaviour profile, and used to design, simulate and monitor the next one.

```mermaid
flowchart LR
  A["1 Ingest<br/>a finished study package"] --> B["2 Understand<br/>design retrospective and feedback digest"]
  B --> C["3 Learn<br/>local behaviour profile, reviewed by clinicians"]
  C --> D["4 Design<br/>synopsis, simulation, design report"]
  D --> E["5 Run<br/>monitor the live trial"]
  E -. "the finished trial becomes the next study to ingest" .-> A
```

*The learning loop: five steps, one return path.*

The loop starts with studies that already exist, so the product is useful before any new trial runs.

**What "local behaviour" means.** It is five layers of parameters, each stored with an interval and the evidence behind it.

| Layer | Examples | Learned from |
| --- | --- | --- |
| Patient | Adherence, visit attendance, dropout timing, use of traditional medicine, reasons for refusing consent | Visit and dosing records, case report forms, patient feedback |
| Investigator and site | Enrolment pace, causes of screen failure, deviation types, data-entry lag, query turnaround | Enrolment, deviation and query logs; monitoring reports |
| Clinical practice | Usual care actually given, dose titration habits, co-prescriptions, referral paths | Concomitant-medication records, hospital records, clinician review |
| System | Ethics and ministry timelines, recurring review comments, import and contract delays | Review letters, submission records |
| Design lessons | Assumed against actual effect size, variance and event rates; amendments and their causes | Protocol, analysis plan, amendments, final report |

**How each value is labelled.** Every parameter carries a source label (learned, elicited or assumed), a prediction interval for a new study, and tags for where it applies: indication, phase, site type, date and rule set. A value counts as learned only above a minimum of evidence, proposed as 30 patients or 3 sites. One study cannot show how much studies differ from each other. That spread comes from published ranges and clinician elicitation until several studies are pooled, and the screen shows the within-study interval beside the wider prediction interval so the reader sees how much of the width is elicited. Global defaults, the baseline every local value is compared with, carry their citation; a default without one is labelled assumed and flagged.

**What each study type can teach**

| Study type | Patient | Investigator and site | Clinical practice | System | Design lessons |
| --- | --- | --- | --- | --- | --- |
| Randomised trial | Full | Full | Partial: care is set by protocol | Full | Full |
| Cohort or registry | Attendance and dropout; no dosing | Recruitment pace and data lag | Full | Ethics review only | Partial |
| Cross-sectional or genomic | Consent and baseline only | Recruitment pace only | Partial | Ethics review only | Little |

The demo shows only the layers its study can supply.

**General framework, local learning**

1. **General engines.** Design, simulation and inference work the same everywhere and are pre-trained on global public data.
2. **Local learning.** Ingested studies and clinician review build the behaviour profile and fine-tune the patient models, inside the country.
3. **Honest transfer.** Every output reports how much local evidence supports it. Thin evidence widens the interval.

```mermaid
flowchart TB
  subgraph ENG["General engines (the same in every country)"]
    direction LR
    D["Design engine"]
    S["Simulation engine"]
    I["Inference engine"]
  end
  L["Localization layer<br/>behaviour profile and country pack: population, standard of care, rules"]
  subgraph DATA["Data base"]
    direction LR
    G["Global public data"]
    N["Ingested local studies<br/>patient-level data stays in the hospital"]
    C["Clinician review and elicitation"]
  end
  DATA --> L
  L -- "calibrates" --> D
  L -- "calibrates" --> S
  L -- "calibrates" --> I
  I -. "finished trials update the profile for later studies" .-> L
```

*Trial OS architecture: three engines, one localization layer, one data base.*

The localization layer holds the behaviour profile and calibrates each engine. Each model is frozen before the trial that uses it, and finished trials update it for later ones.

**Design principles**

1. Randomisation stays. Models adjust and plan; they do not replace controls.
2. Every number carries an interval and its provenance.
3. A named human signs every regulated artefact.
4. Patient-level data stays in the hospital and country it came from. Only derived parameters move.
5. Protocols are structured data in USDM and ICH M11 from the first draft.
6. Clinicians can overrule what the data suggests, and the override is recorded.

## 5. Features

Trial OS has 67 features in nine areas. 16 are in the demo, 22 more complete the first release, 22 follow by month 24 and 7 come later.

| Priority | Meaning |
| --- | --- |
| Demo | Working in the 12-week demo |
| P0 | First release, by month 6 |
| P1 | Stages 1 and 2, by month 24 |
| P2 | Stage 3 |

Acceptance thresholds are proposals to confirm with the first users.

### 5.1 Study Ingest and local behaviour (F1)

Turns a finished study into a structured record, a retrospective and a local behaviour profile. It takes the protocol, amendments, analysis plan, case report forms, datasets, enrolment, deviation and query logs, monitoring reports, review letters, final report and interview notes.

| ID | Feature | What the user gets | Acceptance test | Priority |
| --- | --- | --- | --- | --- |
| F1.1 | Study package intake | Upload of study files in Vietnamese or English: born-digital Word, PDF, spreadsheets and text. Scanned pages are listed and labelled "not read" until F1.10. A checklist shows what is missing | File type correct for ≥ 90% of a hand-labelled set of 100 files | Demo |
| F1.2 | Structured study record | Protocol converted to USDM and ICH M11 data; datasets mapped to a common schema; de-identified inside the hospital | ≥ 90% correct on a hand-labelled set of 200 protocol fields and 50 dataset variables | Demo |
| F1.3 | Design retrospective | Planned against actual: enrolment, screen failure by criterion, dropout, variance, effect, timeline, amendments | Matches a statistician's hand extraction on ≥ 90% of 30 items | Demo |
| F1.4 | Feedback digest | Comments from ethics committees, the ministry, monitors, investigators and patients grouped into themes, each linked to its source text | A reviewer agrees with ≥ 85% of theme assignments | Demo |
| F1.5 | Local behaviour extraction | Patient, site, practice and system parameters, each with an interval and evidence count | 30 parameters re-computed by hand agree; each carries its source label, interval type and tags | Demo |
| F1.6 | Clinician confirmation | Investigators confirm, correct or reject each parameter; a structured debrief and case vignettes fill gaps where data is missing | ≥ 80% of parameters reviewed by a named investigator | Demo |
| F1.7 | Behaviour profile store | Versioned profiles by country, indication and site type, pooled across studies | A second study updates the pooled values and their intervals | P1 |
| F1.8 | Retrospective report for the data owner | A plain-language report on the hospital's or sponsor's own study | Delivered within 10 working days of intake | P0 |
| F1.9 | Continuous capture | Feedback and operations data from running trials flow into the profile | First live study connected | P1 |
| F1.10 | Scanned-document reading | Text recognition for scanned Vietnamese and English pages, feeding F1.2 with the same source links | Character and field accuracy measured on surrogate Vietnamese medical scans and reported | P0 |

### 5.2 Design studio (F2)

Turns a study brief into a structured protocol tested against local lessons.

| ID | Feature | What the user gets | Acceptance test | Priority |
| --- | --- | --- | --- | --- |
| F2.1 | Brief to structured protocol | A synopsis, then a full protocol, as USDM and ICH M11 data | Exports every M11 section; each element traces to a field | Demo |
| F2.2 | Eligibility optimiser | Per criterion, the share of local patients excluded | Eligible share within 10 points of actual on a held-out cohort | P0 |
| F2.3 | Schedule and burden check | Visits and procedures scored against local attendance behaviour | Flags visits whose predicted attendance falls below a set level | P1 |
| F2.4 | Standard-of-care checker | Comparators and concomitant drugs checked against local practice and the insurance list | Flags any drug absent from the insurance list | P0 |
| F2.5 | Design critique from local lessons | Warnings where a draft repeats a choice that failed in an ingested study | Each warning cites the study and the evidence | P0 |
| F2.6 | Device, new-technique and observational templates | Study designs beyond drug trials | One study of each type designed end to end | P1 |
| F2.7 | Master protocols | Platform designs with shared controls and arm entry and exit | Operating characteristics simulated under all planned adaptations | P2 |
| F2.8 | Feasibility and design report | The first sellable output: recommended design, sample size, eligibility, enrolment forecast, retention, timeline, risks and an evidence table | A sponsor reviewer can trace every number to its source label | Demo |

### 5.3 Trial simulator (F3)

Runs the planned study thousands of times under local behaviour, covering patient trajectories, intervention effects, and patient and investigator compliance.

| ID | Feature | What the user gets | Acceptance test | Priority |
| --- | --- | --- | --- | --- |
| F3.1 | Patient trajectory model | The predicted outcome course under usual care, per indication, with a model card | Held-out correlation reported with interval, and its gain over baseline covariates | P0 |
| F3.2 | Expected intervention effect | A published effect entered with its source and a sensitivity range | Every assumed effect cites a trial | P0 |
| F3.3 | Effect transport | Published effects re-weighted to the local population | Assumptions listed; range shown | P1 |
| F3.4 | Patient compliance model | Adherence, missed visits and dropout drawn from the behaviour profile | Re-running the ingested study's own design reproduces its observed dropout curve. This checks the code, not the model. | Demo |
| F3.5 | Investigator and site compliance model | Enrolment pace, deviations and data lag by site type | Predicted site deviation rates beat a flat-rate baseline on held-out sites | P1 |
| F3.6 | Virtual trial runs | At least 1,000 runs per design and 10,000 for null scenarios; power, type I error, size, duration and cost with intervals | Reproducible from seed and version | Demo |
| F3.7 | Scenario comparison | Designs and assumptions side by side, including global defaults against the local profile. Every global default carries its published source, or is labelled assumed and flagged as unsourced | Re-run of 1,000 trials in ≤ 1 hour; no global default shown without its citation or its flag | Demo |
| F3.8 | Replay validation | A finished study's later data hidden, predicted, then compared; and a calibration check on synthetic studies with known truth | Profile re-fitted on data dated before a calendar cut, with a test that nothing later reaches the fit; coverage and width of 80% intervals reported for pre-listed quantities; coverage of every stated interval measured on at least 100 synthetic studies | Demo |

### 5.4 Statistical design and analysis (F4)

Produces the analysis plan and runs the pre-specified analysis.

| ID | Feature | What the user gets | Acceptance test | Priority |
| --- | --- | --- | --- | --- |
| F4.1 | Estimand and analysis plan | A draft following ICH E9(R1) | A biostatistician signs off with edits logged | P0 |
| F4.2 | Pre-specified covariate adjustment | An analysis using a prognostic model frozen before the trial | Same data and plan give the same result on re-run | P0 |
| F4.3 | Sample-size justification | A document an ethics committee or regulator can read | Figures match the simulation output | P0 |
| F4.4 | Group-sequential monitoring | Interim looks with alpha spending | Boundaries match a reference package | P1 |
| F4.5 | Monitoring-committee reports | Reports with unblinded data firewalled | No unblinded field reachable by a blinded role | P1 |
| F4.6 | Anytime-valid monitoring | Continuous looks beside the sequential design | Simulated type I error at or below nominal under all planned adaptations | P2 |
| F4.7 | Bayesian borrowing | A sensitivity analysis with robust priors | Prior weight falls when external and trial data conflict | P2 |
| F4.8 | Subgroup estimates | Effects by genotype and by site, pre-specified only | No subgroup can be added after the plan is locked | P2 |

### 5.5 Recruitment and feasibility (F5)

Counts who can be recruited, where, and how fast.

| ID | Feature | What the user gets | Acceptance test | Priority |
| --- | --- | --- | --- | --- |
| F5.1 | Recruitment data register | The sources that may lawfully be used: hospital records, registries, insurance claims, national health books | Each source has a recorded legal basis | P1 |
| F5.2 | Eligible-patient counts | Counts per site from hospital records, computed inside the hospital | Within 20% of coordinator chart review | P1 |
| F5.3 | Site ranking | Certified sites ranked by expected enrolment and past behaviour | Ranking correlates with actual enrolment on past studies | P1 |
| F5.4 | Enrolment forecast | A forecast with intervals, built first from ingested enrolment logs | Back-tested on ingested studies for the demo; on live studies within ± 50% at month 18 and ± 25% at month 24 | Demo |
| F5.5 | Pre-screening | Candidate lists for site coordinators | Criterion-level accuracy ≥ 85% against coordinator review | P1 |

### 5.6 Regulatory and ethics (F6)

Drafts what reviewers need and applies what reviewers said last time.

| ID | Feature | What the user gets | Acceptance test | Priority |
| --- | --- | --- | --- | --- |
| F6.1 | Dossier drafting | A Vietnamese dossier for site ethics, national ethics and ministry review | No missing mandatory item in 9 of 10 test dossiers | P0 |
| F6.2 | Country rules engine | Required documents, steps and time limits by study type | Drug trials under Circular 50/2025 covered; devices added in P1 | P0 |
| F6.3 | Review-comment pre-emption | The draft checked against comments reviewers raised on ingested studies, used with the applicant's permission | Each flag cites the past comment | P0 |
| F6.4 | Consent form | A bilingual form at a stated reading level | Passes a readability check and investigator review | P0 |
| F6.5 | Submission tracker | Actual review times recorded and fed into the profile | First three submissions tracked | P1 |

### 5.7 Live trial monitor (F7)

Compares the running study with its simulation and flags compliance problems early.

| ID | Feature | What the user gets | Acceptance test | Priority |
| --- | --- | --- | --- | --- |
| F7.1 | Forecast against actual | Enrolment, retention and data flow against the simulation | Updated weekly | P1 |
| F7.2 | Patient compliance signals | Missed visits and dosing gaps flagged by patient and site | Flags precede dropout in at least half of cases | P1 |
| F7.3 | Investigator and site compliance signals | Deviations, data-entry lag and query backlog flagged by site | Flags precede monitor findings in at least half of cases | P1 |
| F7.4 | Data-quality anomalies | Unusual patterns in entered data | Checked against monitor reports | P1 |
| F7.5 | Data-capture integration | A connection to the sponsor's data-capture system | Two systems connected | P1 |
| F7.6 | Adherence support | Reminders through local messaging channels | Measured against a no-reminder arm | P2 |
| F7.7 | Dataset build | Cleaned CDISC datasets | Built in ≤ 1 week | P2 |

### 5.8 Localization framework (F8)

Packages what was learned so the general engines work in a new country.

| ID | Feature | What the user gets | Acceptance test | Priority |
| --- | --- | --- | --- | --- |
| F8.1 | Country pack | Behaviour profile, population, standard of care and rules in one versioned package; Vietnam first | Schema validated now; tested when a second pack loads with no engine change | P0 |
| F8.2 | Evidence label | On every output: learned, elicited or assumed, with the local evidence behind it. A prediction for a new study also shows where its between-study spread came from | No output shown without its label and interval; every evidence table has an interval column and a spread-source column | Demo |
| F8.3 | Country-pack playbook | The steps to build a pack: access agreement, three to five studies ingested, clinician review | Followed for the second country | P1 |
| F8.4 | Pharmacogenomic options | Genotype frequencies and stratification choices | Frequencies sourced and cited | P1 |
| F8.5 | Second country pack | A pack for one more Southeast Asian country | Built in ≤ 3 months | P1 |
| F8.6 | Partner-built packs | Tools for a partner to build a pack without Life AI staff on site | One pack built that way | P2 |

### 5.9 Platform and governance (F9)

Keeps data where it belongs and makes every output checkable.

| ID | Feature | What the user gets | Acceptance test | Priority |
| --- | --- | --- | --- | --- |
| F9.1 | In-country processing | Patient-level data stored and processed where it originated | No patient-level record leaves the in-country environment | Demo |
| F9.2 | Consent registry | Each dataset tagged with its consent scope | A record without scope cannot enter training | Demo |
| F9.3 | Evidence file | A credibility file per model, following FDA's seven-step framework | Complete for the first model used in a design | P0 |
| F9.4 | Human sign-off | A named signature on every regulated output | No export without it | P0 |
| F9.5 | Source linking | Every generated statement linked to its source | Reviewer audit finds < 2% unsupported statements | P0 |
| F9.6 | Reproducibility | Simulations reproducible from seed and version | Identical output on re-run | P0 |
| F9.7 | Vietnamese and English | Interface, inputs and outputs in both | All first-release outputs in both languages | P0 |
| F9.8 | Security | Role-based access, encryption, breach notice within 72 hours | Passes an external review | P0 |
| F9.9 | Audit trail and electronic signatures | Records consistent with ICH E6(R3) | Validated before the first live trial | P1 |
| F9.10 | Legal conformity | A conformity plan under AI Law 134/2025; impact assessments under Law 91/2025 | Filed before first commercial use | P0 |

**Data-use terms and conflict rules**

- **Which studies qualify.** Investigator-initiated or publicly funded studies, and sponsored studies only with the sponsor's written permission.
- **Ownership.** Owners keep their data. Trial OS receives a licence to derive parameters, and patient-level data stays in the hospital.
- **Pooling is opt-in.** A pooled value is shown only when at least five sites contribute to it.
- **Sites are not exposed.** No site-identifiable output goes to a third party without that site's written consent.
- **Studies are not exposed.** Outside its owner, a study is cited without its name.
- **One side per dossier.** Trial OS never works for both the applicant and the reviewer on the same dossier.
- **Children's and genomic data.** Excluded from pooling by default and reviewed separately.
- **What the hospital gets.** Its retrospective report within 10 working days, paid reviewer time, and authorship on methods papers that use its study.

## 6. Demo

The demo takes one finished study, turns it into a local behaviour profile on screen, and uses that profile to produce a feasibility and design report for the next study. It runs about 30 minutes on real files, and anything mocked is labelled.

**Storyline.** A hospital hands over a finished study. Half an hour later, a sponsor holds a design report for the next study, with every number labelled by where it came from.

| Act | Step | Minutes | What the audience sees | Features | Tier |
| --- | --- | --- | --- | --- | --- |
| 1. Ingest | 1. Intake | 3 | The study package is dropped in. A checklist shows what was found and what is missing. Scanned pages are listed as not read. | F1.1 | Must |
|  | 2. Study record | 2 | The protocol as structured data; the dataset mapped and de-identified; a quality report | F1.2, F9.1, F9.2 | Must |
| 2. Understand | 3. Design retrospective | 3 | Planned against actual, line by line: enrolment, screen failures by criterion, dropout, variance, timeline, amendments | F1.3 | Must |
|  | 4. Feedback digest | 3 | Themes from ethics, ministry, monitor and investigator comments. Clicking a theme opens the source text. | F1.4 | Must |
| 3. Learn | 5. Behaviour profile | 3 | Parameters by layer, each labelled learned, elicited or assumed, with its interval and tags, beside the global default | F1.5, F8.2 | Must |
|  | 6. Clinician review | 3 | An investigator confirms two parameters, corrects one, and a debrief answer fills a gap | F1.6 | Must |
| 4. Apply | 7. New brief | 2 | A new study question becomes a protocol synopsis | F2.1 | Must |
|  | 8. Simulate twice | 4 | The same design under cited global defaults and under the local profile: time to enrol, retention, power, duration. Inputs that fell back to a global default are listed with their source. | F3.4, F3.6, F3.7 | Must |
|  | 9. Enrolment forecast | 2 | A forecast by site type with its interval, back-tested on the ingested study | F5.4 | Must |
|  | 10. Design report | 2 | The report a sponsor would buy, ending in an evidence table | F2.8 | Must |
| 5. Trust | 11. Replay and calibration | 3 | The profile re-fitted on the study's first two-thirds by calendar date; the last third predicted, then revealed. Then: how often the stated intervals held the truth on synthetic studies | F3.8 | Must |
|  | 12. Limits | 1 | What one study can and cannot show |  | Must |
|  | 13. Sign-off | 1 | A named biostatistician signs the design report; export unlocks only then | F2.8 | Must |

**The demo ends on the signed design report**, the first thing a buyer pays for. **Should-show if time allows:** the design critique (F2.5) and the evidence file (F9.3). **Stretch:** the dossier draft with comment pre-emption (F6.1, F6.3), and the trajectory model with covariate adjustment (F3.1, F4.2).

**Replay is protected.** It is the only step that answers the biostatistician's question. If time runs short, cut in this order: the design critique, model-based theming in the feedback digest (keep the keyword baseline with human review), then the Vietnamese interface beyond the report itself.

**Which study qualifies**

- It is interventional or longitudinal, with dated visits.
- It has at least two sites, or one site with a dated enrolment log.
- The data owner gives written permission, and consent covers reuse.
- It is cleared for display to people outside the study.

If no Life AI study qualifies by week 2, use a partner's investigator-initiated trial. The full request list for the three studies is in [data-request.md](data-request.md).

**Who sees it, and what counts as success**

| Audience | The question they are answering | Success signal |
| --- | --- | --- |
| Hospital research unit | Is handing over our studies worth it? | A second study offered for ingest |
| Sponsor clinical lead | Would I pay for this report? | A signed letter of intent, then one paid pilot |
| Regional research organisation | Would I use this in feasibility and bids? | A pilot on a live bid |
| Biostatistician | Do I believe the numbers? | No unexplained result under questioning |
| Ministry science administration and research funders | Would a national programme be worth funding? | A named counterpart and a budget route |

The target is two letters of intent from ten demos.

**Build plan: 12 weeks, about 48 engineer-weeks**

| Weeks | Work | Output |
| --- | --- | --- |
| 1–2 | Study inventory and permission check. Three spikes on surrogate Vietnamese documents: extraction accuracy on born-digital files (scan accuracy measured for F1.10, not built), dataset mapping, simulator skeleton. The synthetic golden path and calibration check already run. | Go or no-go on real data; spike results |
| 3–5 | Intake, study record, design retrospective, feedback digest | Acts 1 and 2 |
| 6–8 | Behaviour extraction with labels, clinician review, enrolment forecast | Act 3 and step 9 |
| 9–10 | Synopsis, simulator, design report | Act 4 |
| 11–12 | Replay and calibration on the real study; dry runs with a biostatistician and an investigator | Act 5 and fixes |

**Team.** One product and clinical lead, one biostatistician, two machine-learning engineers, one data engineer and one full-stack engineer, with a part-time regulatory specialist and a clinician from the study's field. The system design is in [demo-architecture.md](demo-architecture.md), and the screens are in [../design](../design/README.md).

**What the demo does not prove.** One study cannot show that local knowledge makes a design better. Step 8 shows what changes when local inputs replace global ones. Calibration on synthetic studies shows the stated intervals mean what they say; it does not show that one study predicts the next. Proof needs replay across several studies, then a prospective trial.

**If the package is thin.** With only a protocol and dataset, acts 1 to 3 still run. The site and system layers then come from the debrief and are labelled as elicited. With no usable patient-level data, the demo falls back to published figures and loses its point.

## 7. Release plan

Trial OS ships in four stages over 48 months: a proof on one study, then Vietnam, Southeast Asia and global. Spending stays small until Gate 0 shows that a finished study can be ingested and its profile reviewed.

```mermaid
flowchart LR
  S0["Stage 0 · Proof<br/>one Life AI study"] --> G0{{"Gate 0<br/>study ingested, profile reviewed, replay reported"}}
  G0 --> S1["Stage 1 · Vietnam<br/>partner hospitals, first release"]
  S1 --> G1{{"Gate 1<br/>ten studies ingested, three paid design studies, one dossier submitted"}}
  G1 --> S2["Stage 2 · Southeast Asia<br/>Thailand and Indonesia"]
  S2 --> G2{{"Gate 2<br/>pack built in ≤ 3 months, one live trial, one customer outside Vietnam"}}
  G2 --> S3["Stage 3 · Global<br/>partner-built packs, engine licences"]
```

*Release plan: four stages, three gates, 48 months.*

| Stage | Where | Access used | What ships | Gate to pass |
| --- | --- | --- | --- | --- |
| 0. Proof | One Life AI study | Life AI's own data and investigators | Study Ingest, behaviour profile, demo | An eligible, consented study ingested; investigators review the profile; replay on truncated data reported |
| 1. Vietnam | Partner hospitals | Hospitals, doctors and the ministry | First release: design studio, simulator, dossier drafting, Vietnam pack | Ten studies ingested; three paid design studies; one dossier submitted |
| 2. Southeast Asia | Thailand and Indonesia | Local hospital and regulator relationships | Second country pack, live trial monitor, recruitment features | Pack built in ≤ 3 months; one live trial; one customer outside Vietnam |
| 3. Global | Countries reached through partners | Research organisations and academic networks | Partner-built packs, platform trials, engine licences | A pack built without Life AI staff on site |

**Why it scales.** The engines are general and only the country pack is local. Expansion is a repeatable playbook: secure access, ingest three to five finished studies, run clinician review, release the pack.

**What does not scale by itself.** Access. Outside Southeast Asia, Trial OS depends on partners who hold the hospital and regulator relationships. Stage 3 tests whether a partner can run the playbook alone.

**Thailand and Indonesia need different entries**

- **Indonesia.** Health data and specimens must [stay in the country](https://resourcehub.bakermckenzie.com/en/resources/global-data-and-cyber-handbook/asia-pacific/indonesia/topics/data-localization-and-regulation-of-non-personal-data). The ministry already runs a clinical research centre, [INA-CRC](https://ina-crc.kemkes.go.id/en), with feasibility support and a site-capacity database. Enter as its supplier, not its rival.
- **Thailand.** It is the most mature market in the region, with [central ethics review](https://clinregs.niaid.nih.gov/country/thailand), a [national registry](https://www.thaiclinicaltrials.org/) and established research organisations. The gap is smaller, so enter through a research organisation or a university trial unit.
- **Access is not yet shown.** Life AI's announced partners are a [Kalbe Farma laboratory subsidiary](https://www.globenewswire.com/news-release/2025/12/11/3203822/0/en/Kalbe-Farma-and-LIFE-AI-Unveil-StrokeGENME-to-Personalize-Stroke-Prevention-at-National-Scale.html) in Indonesia and a [volunteer medical association](https://www.globenewswire.com/news-release/2026/01/12/3216875/0/en/LIFE-AI-Builds-Upon-Global-Biohub-by-Initiating-Deployment-of-PREVANA-Subnets-in-Thailand.html) in Thailand. Neither runs trials.
- **Test.** Name one research director and one regulator contact in each country, and make the two calls this month.

**Gate 0 is the cheapest and the most likely to fail.** It needs one study whose consent covers reuse, whose files are complete enough to ingest, and whose investigators will review the result.

## 8. Commercial plan

Revenue comes first from design studies sold to sponsors and research organisations working in Vietnam and Southeast Asia, then from country-pack and engine licences worldwide. Vietnam is where the product is proven, and it is too small to be the whole market.

**Market reality**

- Vietnam had 318 studies recruiting or about to start on 1 Oct 2026. The six largest Southeast Asian markets together had 2,390.
- Published estimates of Vietnam's trial market for 2025 run from $29.8M to $120M. Treat both as low confidence.
- An illustration: 10% of 71 new trials a year at an assumed $40,000 per design study is about $280,000. Even at double that count it stays under $600,000. Vietnam alone does not carry a company.
- Domestic sponsors led Vietnamese registrations in 2020–2024, 198 to 154, but most are universities and hospitals. Commercial local buyers are few. The most-studied conditions are lung cancer and infertility (40 studies each), then tuberculosis, infections and type 2 diabetes (35 each).

**Customers**

| Segment | Why they would buy | What they buy | Priority |
| --- | --- | --- | --- |
| Domestic pharma, vaccine and biotech makers | Few in-house statisticians; new rules to meet | Design study and dossier | Second. Few such buyers, and vaccine endpoints sit outside the EMA-qualified scope. |
| Multinational sponsors and regional research organisations | Need feasibility and ethnic-factor bridging for Vietnam | Localization report; platform licence | First |
| Device, diagnostic and digital-health makers | Trials are mandatory for higher-risk devices | Lower-priced design templates | Second |
| Hospital research units and universities | Many protocols, little support | Free or subsidised use in exchange for data partnership | First, as the source of studies to ingest rather than revenue |
| Government and research funders | Wants more and better local trials, and to know what publicly funded studies taught | A national study-learning programme: publicly funded studies ingested, an annual lessons report, timelines across studies | First |
| Global sponsors | Evidence that models and effects transfer to Asian patients | Country-pack licence | Later |

**Government as a buyer is unproven.** No ministry budget line, tender or programme for a trial platform was found. The evidence of demand is the [May 2025 industry roadmap](https://assets.kpmg.com/content/dam/kpmg/vn/pdf/2025/05/roadmap-clinical-trials-in-vietnam-en.pdf), which recommends an online portal and a trial database, and Circular 50/2025, which allows online filing. Two meetings should establish the budget line, the procurement route and the counterpart.

**Offers and price hypotheses.** None of these prices has been tested with a buyer.

| Offer | Unit | Price hypothesis | Value anchor |
| --- | --- | --- | --- |
| Feasibility and design report | Per protocol | $30,000–80,000 | A Phase III amendment cost a median $535,000 in a 2016 global study; local costs are lower |
| Localization and bridging report | Per product and country | $20,000–50,000 | Shows whether existing data covers local patients, or what a local study must add |
| Dossier preparation | Per submission | $10,000–20,000 | Approval plus ethics took nearly 160 days in the last published data |
| Platform licence | Per organisation per year | $100,000–250,000 | Used across every bid and study |
| Live-trial monitoring | Per trial per month | To be set after the first live trial | Enrolment delay and deviation cost |
| National study-learning programme | Per year | To be scoped with the ministry | The 2025 roadmap calls for a trial database by 2026–27 |

**Ingest as the way in.** A free study retrospective for hospitals, and a paid one for sponsors, opens each account and adds a study to the behaviour profile.

**The first sellable deliverable** is the feasibility and design report, with a fixed table of contents:

1. Recommended design and the alternatives considered.
2. Sample size and power under local behaviour.
3. Who the eligibility criteria exclude locally.
4. Enrolment and screen-failure forecast by site type, with intervals.
5. Retention and adherence to expect.
6. Timeline to first patient and to last visit.
7. Risks seen in past local studies, and how to reduce them.
8. An evidence table giving the source label of every number.

Costs come from a calculator that uses the sponsor's own unit costs.

**Revenue sketch.** At these prices, three design studies in year 1 bring $90,000–240,000. A year 2 of eight reports across Vietnam, Thailand and Indonesia, four bridging reports and one licence brings about $0.4–1.1 million. Stage 0 needs about 48 engineer-weeks, plus a lead and a biostatistician for 12 weeks. Costs are not yet priced, so there is no break-even estimate.

**Go-to-market**

1. **Months 0–4.** Interview ten buyers on price before the demo. Sign two design partners: one sponsor and one research organisation. Agree study hand-over with two hospitals.
2. **Months 4–12.** Founder-led sales of design studies. Publish the transport result. Run one pilot dossier with the ministry's science administration.
3. **Months 12–24.** Sell through regional research organisations. Enter Thailand and Indonesia with partners that already have sites there.
4. **Months 24 on.** License country packs to global platforms and sponsors.

**Partners**

| Role | Candidates | What they provide |
| --- | --- | --- |
| Data and clinical | Hanoi Medical University, K Hospital's research centre, Bach Mai, stroke registry investigators, OUCRU | Cohorts, clinician review, site behaviour data |
| Channel | Big Leap, SMART Research, Novotech | Sponsors and live studies |
| Regulatory | Ministry of Health science administration; national ethics committee | A pilot review and early precedent |
| Infrastructure | An in-country cloud provider | Data residency |
| US | A US-led SURPASS team | Method credibility and a route to FDA-facing work |

**Funding**

- **Stage 0** is small enough to fund internally or from a research grant.
- **Non-dilutive routes** are Vietnamese science and technology programmes and international research funders. A SURPASS sub-award is possible but unlikely.
- **Equity** makes sense after Gate 1. Recent rounds show appetite and well-funded rivals: Biorce $52M, QuantHealth $45M, Faro $37.3M.

**Moat, ranked honestly**

1. **Behaviour profiles from ingested and live studies.** Strongest. It rests on access that global vendors lack, and it grows with every study ingested.
2. **Consented local cohorts and clinician-validated practice maps.** Moderate. IQVIA could copy it with one hospital deal.
3. **Regulatory precedent in Vietnam.** Moderate, and available to whoever gets there first.
4. **Software.** Weak.

**The service trap.** A design studio can become a consultancy that never turns into a product. The rule is to take no engagement that fails to add data or reusable country-pack content. Track hours per design study, which should fall with each one.

## 9. Risks and critical review

The two risks most likely to end the project are the data and the buyers, and both can be tested within four months for little money.

| Risk | Why it is real | Early test | Stop or change if |
| --- | --- | --- | --- |
| Local data is too thin or not consented for model training | The one public study is cross-sectional genotyping of 250 children | Data audit in weeks 1–2 | No cohort has follow-up outcomes and consent for secondary use |
| Localization adds little to outcome models | A global model may already fit local patients well | The transport experiment | The locally fitted model's gain is negligible. Value then lies in practice and regulatory localization only. |
| Buyers will not pay | Vietnam logs roughly 70 or more new registered trials a year, mostly academic, with small budgets | Ten price interviews; two letters of intent | Fewer than two letters from ten demos |
| The "operating system" framing invites head-on competition | Medidata, Veeva, IQVIA and Pi Health own operations software | Buyer reaction in interviews | Buyers compare it to their EDC. Reposition as a design and localization layer that plugs in. |
| Regulators do not accept model-based designs | No ASEAN precedent was found | A pilot dossier reviewed informally | The ministry asks for the model to be removed. Covariate adjustment alone keeps randomisation, so this risk is lower than for synthetic controls. |
| Predictions are wrong and trusted anyway | Published outcome-prediction accuracy is modest | Calibration on synthetic studies with known truth, then replay on completed studies | Intervals fail to cover the truth at their stated rate |
| Investigator modelling meets resistance | Doctors may object to being scored, and research organisations hold the data | Conversations with two partner sites | No partner will share deviation data. Model at site level and share results with sites. |
| Data law tightens | A draft law would ban export of core data; AI Law requires conformity assessment | Counsel opinion before any model leaves Vietnam | Model export is treated as data export. Run inference in-country for foreign clients. |
| Consent and conflict of interest | Research data reused for a commercial product; children's data in autism | Ethics review of the reuse plan | An ethics committee objects. Re-consent or exclude the cohort. |
| A funded rival arrives | Pi Health has Asian sites and the same positioning | Watch for Southeast Asian site announcements | A rival signs a national hospital data deal first. Partner or sell rather than race. |
| A SURPASS bid consumes the team | A summary is four pages, but teaming takes weeks | Time-box to the two events and one summary | No US prime shows interest by mid-November |
| Finished studies lack operations records | Deviation logs, queries and review letters are often not kept, or not shared | Inventory one study package in week 1 | Only a protocol and dataset exist. Build the site and system layers from clinician interviews, and label them as elicited. |
| Access does not travel | Hospital and government relationships are personal and national | The second country pack | The second pack takes far longer than three months. Move to partner-led packs earlier. |
| The second markets are harder than assumed | Indonesia keeps health data in-country and its ministry already runs a trial coordination centre. Thailand has incumbents, a registry and central ethics review. | Two calls in each country this month | No named hospital and regulator contact by Gate 1. Supply the existing national body instead of competing with it. |

**Claims to state carefully**

- **"Simulate expected intervention effects."** These are published effects transported to a local population with wide intervals. They are not forecasts of a new drug's efficacy.
- **"Doctor compliance."** It is learned from deviation, query and monitoring records in ingested studies, and from clinician review. It is only as good as those records.
- **"Trained on local studies."** Fine-tuned on local studies. Training from scratch needs far more data than a few disease cohorts.
- **"A platform globally."** The global asset is a fast, repeatable way to build a country pack. The second pack is the proof.

**What would raise confidence**

- A consented diabetes cohort with at least 12 months of follow-up.
- One research organisation willing to share site-level enrolment and deviation history.
- One sponsor paying for a design study before the software is finished.
- A full-time lead and a full-time biostatistician. Part-time ownership will not clear Gate 1.

## 10. Decisions and open questions

Three decisions are needed now, and the first questions below decide whether Gate 0 can be passed at all.

**Decided on 1 Oct 2026**

- Life AI will gather the data from its diabetes, stroke and autism studies. The request list is in [data-request.md](data-request.md).
- Thailand and Indonesia are the second markets.
- Sponsors and government are the first buyers.

**Decided on 1 Oct 2026, after the CEO review**

- Replay (F3.8) is in the demo and protected, with a calibration check on synthetic studies. Scanned-document reading moves to the first release (F1.10).
- Every global default cites its source or is flagged as unsourced.
- The demo ends on the signed design report. The dossier draft stays a stretch item.
- Until real data arrives, the two zones run as two processes on one machine with separate databases. The hospital server request still goes in week 1.
- Every evidence table shows an interval and where the between-study spread came from.

**Three decisions needed**

| Decision | Recommendation | Reason |
| --- | --- | --- |
| Lead indication | Type 2 diabetes first, stroke second, autism deferred | HbA1c is a continuous endpoint, the only kind covered by the EMA-qualified twin method. Stroke has a 121,715-patient Vietnamese quality registry, with access not yet secured. Autism endpoints are behavioural and heterogeneous. |
| Business form | Tech-enabled design service first, software licence later | Local sponsors run too few trials to buy software. A service also generates the operations data the models lack. |
| SURPASS posture | Use it as the design blueprint, and seek a sub-performer role on a US-led team | A prime team must bring three drug arms and six sites, and win FDA clearance of a master-protocol IND within 24 months. Solution Summaries are due 30 Nov 2026. |

**Open questions**

- [ ] Which finished study has the fullest package for the demo ingest?
- [ ] Do its files include deviation logs, data queries, monitoring reports and review letters?
- [ ] Which investigators will review the behaviour profile, and which hospitals will hand over studies next?

* [ ] For each diabetes, stroke and autism study: design, patient count, sites, follow-up length, endpoints, and whether treatment and adherence were recorded?
* [ ] Does consent cover secondary use, model training and commercial use?
* [ ] Who owns each dataset: Life AI, the hospital, or both? Where is it held?
* [ ] Which legal entity builds and owns Trial OS? This affects SURPASS eligibility and data-law duties.
* [ ] Is there a US entity that could join a SURPASS team by 30 Nov 2026?
* [ ] Which hospital or research organisation would share site-level enrolment and deviation history?
* [ ] Who are the full-time lead and biostatistician?
* [ ] What is the budget ceiling for Stage 0?
* [ ] Is type 2 diabetes agreed as the lead indication?
* [ ] Should the product drop "OS" from its name, given Pi Health's positioning?

## Appendix A. Problem evidence and Vietnam context

Trials fail on design and operations as often as on biology, and designs imported from Western data fit Southeast Asian patients poorly. Trial OS addresses both, starting where the second problem is sharpest.

**The general problem.** ARPA-H puts drug development at more than a decade, $1–2 billion and a 90% failure rate. The operational record (Tufts CSDD, via secondary sources) explains part of it:

- 53% of studies extended enrolment timelines, and 11% of sites enrolled no one (2012 study).
- 76% of protocols need at least one substantial amendment, up from 57% in 2016. Median direct cost in 2016 was $141k in Phase II and $535k in Phase III.
- Site activation takes 31.4 weeks. Dropout runs at 19.1% and screen failure at 36.3%.
- A Phase III protocol averages 119 deviations, touching about a third of patients.

**The local problem.** Patients, practice and systems differ in ways a protocol must reflect:

- **Biology.** About 60% of East Asians carry a CYP2C19 no-function allele, against 30% of Europeans. WHO sets Asian BMI action points at 23.0 and 27.5. Intracerebral haemorrhage is about 19% of strokes in one Vietnamese multi-centre registry, against 8–15% in the US, UK and Australia.
- **Practice.** A pooled estimate puts thrombolysis at about 5.5% of stroke patients, with a wide interval. One study found 14% arrive by ambulance. In 2021 ministry figures, about 35% of people with diabetes were diagnosed and 23.3% managed at a facility. Traditional medicine is about a quarter of commune-level visits, 16% at district and 11% at provincial outpatient level.
- **System.** Out-of-pocket spending is 39.4% of health expenditure. The insurance drug list limits which comparators are realistic.

**Vietnam is under-trialled.** It has about 13 registered studies per million people, a fifth of Thailand's rate.

| Country | Studies with a local site | Recruiting or about to start | Per million people (approx.) |
| --- | --- | --- | --- |
| Singapore | 4,058 | 669 | 690 |
| Malaysia | 2,587 | 443 | 72 |
| Thailand | 4,269 | 590 | 60 |
| Vietnam | 1,317 | 318 | 13 |
| Philippines | 1,341 | 110 | 11 |
| Indonesia | 1,711 | 260 | 6 |

Counts are from a ClinicalTrials.gov mirror dated 1 Oct 2026. Per-million rates are derived from approximate populations.

**Why now**

- **Regulation moved.** Circular 50/2025/TT-BYT replaced the 2018 trial rules on 27 Feb 2026. Before it, approval plus ethics took nearly 160 days, and the full path could exceed a year. No timings under the new rules are published yet.
- **Bridging evidence matters more.** Under the 2025 registration rules, a new drug licensed elsewhere with complete clinical data can skip local trials if that data supports analysis of Asian ethnic factors under ICH E5. Vaccines are excluded. The rules were reissued on 1 Oct 2026 and need re-checking.
- **Data is arriving.** 1,268 of 1,650 hospitals had electronic records by 19 Sep 2026.
- **The method has a sponsor.** ARPA-H now funds simulation-first, continuously analysed trials, and FDA is piloting real-time review.

## Appendix B. ARPA-H SURPASS: review and what to adopt

SURPASS is the best public specification of a simulation-first trial platform, and Trial OS should copy its architecture and metrics. It is a poor funding target for a Vietnam-led team.

**What it is.** SURPASS (Simulation-augmented, Real-time Platform Adaptive Seamless Trials) was announced on 30 Sep 2026 under solicitation ARPA-H-SOL-26-164. It aims to cut clinical development from over a decade to under four years. It funds perpetual adaptive platform trials that share infrastructure and control groups and analyse data as it accrues. The budget is not disclosed.

**Structure.** Five years: a 24-month build stage, then 36 months of prospective validation. The gate between them is FDA permission to proceed on a master-protocol IND and at least one arm.

**The three technical areas and their targets**

| Area | Metric | SURPASS final target | Trial OS target by month 24 (proposed) |
| --- | --- | --- | --- |
| 1. Phaseless design engine | Simulator build by a non-programmer | ≤ 1 month | ≤ 1 week for a supported indication |
|  | Re-simulation after a design change | ≤ 5 hours | ≤ 1 hour |
|  | Control allocation cut by in silico methods | ≥ 80% by year 5 | No fixed promise. Covariate adjustment alone gave 10–30% smaller total samples in the method paper. |
| 2. Continuous inference engine | Interim analysis frequency | Monthly or more | Monthly |
|  | Sample size against a non-adaptive reference | ≥ 50% lower | Reported per design, no fixed promise |
| 3. Agentic operations layer | Enrolment prediction accuracy | Within ± 25% | Within ± 50% by month 18, ± 25% by month 24 |
|  | Data cleaning and dataset build | ≤ 1 week | Not targeted before month 24 |
|  | Human hours against historical | ≤ one quarter | Measured from the first engagement |
|  | Ethics approval per new arm | < 1 week | Dossier ready in ≤ 5 working days |

SURPASS also requires that methods "demonstrate those properties under the complete set of planned adaptations" wherever type I error control is needed. At least one Model Master File must be made public once FDA accepts it. Other final targets include go/no-go decisions 2.5 times faster, time to a first effective drug halved, first dose within three months of IND submission, contracts within a week, and cost per arm below the lowest research-organisation quote.

**What a prime bidder must bring**

- One core team covering all three areas.
- At least three drug or biologic arms. At least two should be unapproved for any indication, including one with no prior human data.
- Preferably at least two for-profit sponsors, one a startup or small company.
- At least six clinical sites by programme end.

**Eligibility.** Non-US entities may participate and are encouraged to team with US entities. No award goes to an entity organised under the laws of a "covered foreign country". The statute it cites names Russia, Iran, North Korea and China, plus any country later designated. Neither Vietnam nor Singapore is named.

**Companion awards**

| Project | Purpose | Performer | Award |
| --- | --- | --- | --- |
| [STACK](https://arpa-h.gov/explore-funding/awards/4742) | AI-assisted site activation, converting research-naïve sites | Evidence Health | Up to $41.18M |
| [COMMONS](https://arpa-h.gov/explore-funding/awards/4743) | Consent-governed national real-world data, over 40 million people | Evidence Health | Up to $49.9M |
| [CINCH](https://arpa-h.gov/explore-funding/awards/4740) | Patient-initiated cancer record review and navigation | Courage Health | Up to $8.95M |

**Critical reading**

- **It is a US industrial-policy programme.** The stated aim is US leadership in clinical development, and STACK funds domestic sites. A Vietnam-sited platform runs against that.
- **Several targets are stretch goals.** An 80% control reduction needs external borrowing that regulators have not broadly accepted. Trial OS should not promise it.
- **It leaves gaps Trial OS can own.** The solicitation covers drugs and biologics only and is silent on ethnic factors and on moving models across populations. It asks for deviation detection and retention models, but not for simulating investigator behaviour at design time.
- **Each companion has a Vietnamese analogue.** STACK maps to readying hospitals beyond the 46 facilities KPMG counted as GCP-certified in 2024. COMMONS maps to a consent-governed national data layer. CINCH maps to patient navigation through the 30 million VNeID health books.

**Recommended posture**

1. Adopt the three-engine architecture and report against the SURPASS metrics from day one.
2. Join the 15 Oct webinar and the 6 Nov Proposers' Day, and list in the teaming directory.
3. Offer a US prime one specific thing: evidence on how prognostic models transfer to an Asian population.
4. Pitch the same blueprint to Vietnam's Ministry of Health as a national programme under Resolutions 57 and 72.

Key dates: Solution Summary due 30 Nov 2026; pitch package due about 22 Jan 2027, to be confirmed in feedback letters. Metrics were read from a mirror of the solicitation PDF, because SAM.gov was unreachable.

## Appendix C. State of the art

Three of the capabilities in the brief are validated, three are partly supported, and four are weak or unproven. The product should be built on the validated ones and treat the rest as labelled research.

| Capability | Best evidence | Maturity | What Trial OS does with it |
| --- | --- | --- | --- |
| Prognostic covariate adjustment (digital-twin-as-covariate) | [EMA qualified PROCOVA](https://www.ema.europa.eu/en/documents/regulatory-procedural-guideline/qualification-opinion-prognostic-covariate-adjustment-procovatm_en.pdf) in 2022 for Phase 2/3 continuous endpoints. The [method paper](https://www.degruyterbrill.com/document/doi/10.1515/ijb-2021-0072/html) claims 10–30% smaller samples. | Validated | Core of v1. A weak model costs power, not type I error, if it is frozen before the trial and trained on independent data. |
| Eligibility optimisation and patient matching | [Trial Pathfinder](https://www.nature.com/articles/s41586-021-03430-5) more than doubled the eligible pool. [TrialGPT](https://www.nature.com/articles/s41467-024-53081-z) cut screening time 42.6%. | Validated retrospectively | In v1. |
| Platform and adaptive trials | [RECOVERY](https://www.ukri.org/who-we-are/how-we-are-doing/research-outcomes-and-impact/mrc/recovery-trial-identifies-covid-19-treatments/) enrolled over 40,000 and started within six weeks. STAMPEDE changed standard of care four times. | Proven in practice | Target design pattern for stage 3 of the release plan. |
| Anytime-valid inference | [Safe logrank test](https://arxiv.org/abs/2011.06931) behaves like O'Brien-Fleming, with a larger maximum sample. | Sound theory, little regulatory precedent | Offered beside group-sequential designs, never alone. The cited method suits time-to-event endpoints. |
| External controls and Bayesian borrowing | FDA guidance is still draft. Population mismatch causes bias and type I error inflation. | Conditional | Sensitivity analysis only in v1. |
| Adherence and dropout modelling | Pharmacometric adherence models are established. No validated general dropout model was found. | Partial | Built per indication from local data. |
| Trial-outcome prediction | [HINT](https://arxiv.org/abs/2102.04252) reports PR-AUC of 0.61–0.62 for Phase II and III. A [2025 review](https://www.sciencedirect.com/science/article/pii/S1359644625000455) finds models struggle on external validation. | Weak | Not sold as probability of success. |
| Generative patient twins and trial agents | A [Sept 2026 review](https://www.nature.com/articles/s44222-026-00487-7) finds that agentic AI has "no demonstrated randomized-trial benefit to date". | Unproven | Scenario exploration, labelled as such. Twin evidence in the papers read is retrospective. |
| Investigator and site behaviour models | One [2026 paper](https://link.springer.com/article/10.1186/s13104-026-07823-5) flags outlier sites across 39,936 sites. No generative model exists. | Research gap | The product's main research bet. |
| Transport across populations | Methods exist. No one has measured how much a US-trained prognostic model loses on Asian patients. | Research gap | The first experiment (Appendix E). |

**Regulatory position**

- **FDA.** The AI credibility guidance of Jan 2025 and the Bayesian guidance of Jan 2026 are both still draft. Adaptive-design guidance is final (2019). A real-time trial pilot with AstraZeneca and Amgen began in April 2026.
- **ICH.** E6(R3) good clinical practice was adopted in Jan 2025. The M11 structured protocol became final FDA guidance on 22 May 2026. E20 on adaptive designs is still draft.
- **EMA.** PROCOVA is the only qualified twin-based procedure. EMA did not qualify how the prognostic model is built, and asked that differences between training and trial populations be investigated. Guidance on AI in clinical development is planned through 2028.
- **Vietnam.** AI Law 134/2025 took effect on 1 Mar 2026 and requires conformity assessment for high-risk systems. No ASEAN regulator precedent for digital twins was found.
- **Standards.** CDISC USDM v4.0 (June 2025) is aligned with M11, so a protocol can be held as structured data.

The practical conclusion is that the regulator-safe path is narrow. Use models to adjust and to plan, keep randomisation, and pre-specify everything.

## Appendix D. Competitors

The market is split three ways: statistical simulators, patient twins and operations agents. No company found joins all three, and none is built on Southeast Asian data.

| Company | What it does | Latest verified signal | Gap against Trial OS |
| --- | --- | --- | --- |
| [Pi Health](https://www.pihealth.ai/our-approach) | "AI-native operating system for clinical trials": EDC, CTMS, eTMF, start-up. Built its own cancer hospital in India. | $30–40M raised (reports differ); [GSK collaboration](https://www.pihealth.ai/press.html), Sep 2025 | Oncology only. No Southeast Asian sites verified. No trajectory simulation. |
| [Unlearn](https://www.unlearn.ai/press) | Prognostic twins with PROCOVA to shrink control arms | $50M Series C, Feb 2024; TrialPioneer launched Jan 2026 | Neurology-centred. Patient outcomes only. No Asian data verified. |
| [QuantHealth](https://www.fiercehealthcare.com/finance/quanthealth-raises-45m-series-b-accelerate-ai-driven-clinical-trial-simulations) | Patient-level trial simulation; 600+ trials simulated | $45M Series B, 4 Aug 2026 | Data geography undisclosed. No operations layer. |
| [Phesi](https://www.phesi.com/) | Protocol optimisation, site selection, standard-of-care twins | 325M+ patients, 485,000 trials | Analytics and consulting. Asian depth not stated. |
| [Faro AI](https://fortune.com/press-releases/faro-series-b-scale-agentic-ai-clinical-development-2026-08-26/) | Structured study design, document generation, agents | $37.3M Series B, 26 Aug 2026; 6 of top 10 pharma. Rival Biorce raised $52M in Feb 2026. | No outcome or site simulation. |
| [Medidata](https://finance.yahoo.com/healthcare/articles/medidata-launches-medidata-plus-ai-050000927.html) | Protocol optimisation, synthetic control arm, AI layer | Medidata Plus launched 23 Jul 2026; 38,000+ trials | Enterprise pricing. No Southeast Asian localisation shown. |
| [IQVIA](https://www.iqvia.com/newsroom/2026/03/iqvia-unveils-iqvia-ai-a-unified-agentic-ai-platform) | 150+ agents; predictive site and enrolment models | IQVIA.ai launched 16 Mar 2026; offices in Hanoi and Ho Chi Minh City | Vietnam data is market audit, not patient-level. |
| [Citeline](https://www.globenewswire.com/news-release/2026/09/10/3359588/0/en/citeline-launches-patient-intelligence-to-inform-clinical-trial-feasibility-and-protocol-planning.html) | Feasibility and standard-of-care intelligence | Patient Intelligence launched 10 Sep 2026: US plus 19 countries | Starts in oncology. Southeast Asia not named. |
| [Cytel](https://cytel.com/news-and-events/cytel-launches-east-horizon/) and [Berry](https://www.berryconsultants.com/resource/berry-consultants-releases-facts-8) | Adaptive design simulators (East Horizon, FACTS 8) | FACTS 8 released Oct 2025 | Statistician tools with no data or operations. |
| [Lindus Health](https://www.bioxconomy.com/clinical-and-research/lindus-health-raises-55m-to-advance-ai-tech-expand-team) | AI-enabled contract research organisation | $55M Series B, Jan 2025 | US and UK focus. |
| [Novotech](https://www.businesswire.com/news/home/20250330371581/en/Novotech-Welcomes-New-Investment-From-GIC-Temasek-and-Existing-Investor-TPG-to-Accelerate-Global-Growth) | Asia-Pacific contract research with a Vietnam office | New investment from GIC and Temasek, Mar 2025 | No proprietary AI platform verified. A channel as much as a rival. |
| [Big Leap](https://bigleapresearch.co/), [SMART Research](https://www.smartresearch.com.vn/About-us.html), [VietStar](https://vietstar-research.com/about-us-vietstar-research.htm) | Vietnamese contract research services | Big Leap: 40+ studies across 36 hospitals | Services only. Partners, not competitors. |
| [Mesh Bio](https://www.mobihealthnews.com/news/asia/mesh-bio-bags-35m-bring-digital-twin-tech-hong-kong-indonesia-philippines) and [Oncoshot](https://technode.global/2026/09/11/japans-mitsui-taps-singapores-oncoshot-for-ai-oncology-data-platform-across-ihh-hospital-network/) | Southeast Asian patient-level data: metabolic twin; federated oncology records | Oncoshot–Mitsui deal across 89 IHH hospitals, 11 Sep 2026 | Not trial design platforms. Mesh Bio overlaps in diabetes. |
| [Claude for Life Sciences](https://hitconsultant.net/2026/01/12/anthropic-debuts-claude-for-healthcare-and-life-sciences/), [Veeva AI](https://www.veeva.com/resources/veeva-ai-agents-to-be-released-across-all-veeva-applications/) | Protocol drafting and operations agents from model and suite vendors | Claude protocol drafting with a Medidata connector, Jan 2026 | Generic. Makes drafting a commodity. |

**White space**

- Simulating investigator and site behaviour inside trial design. Current tools predict site performance or detect anomalies after the fact.
- Vietnamese patient-level data, and local start-up work: Vietnamese-language dossiers and ethics and ministry flows.
- Mid-size Asian sponsors. Enterprise suites target top-20 pharma, and local research organisations have no AI product.

**Strongest threats**

- Pi Health extending from India into Southeast Asia with the same positioning.
- IQVIA or Medidata adding a Vietnam data partnership to agents already in production.
- Protocol authoring becoming a free feature of foundation models.

The consequence for scope: do not build EDC, CTMS or eTMF. Integrate with them, and put the effort into simulation and localisation.

## Appendix E. Local data fit and data law

Existing local studies can supply population priors now. Trajectory and behaviour models need follow-up and operations data that an audit must confirm.

**What Life AI's studies can and cannot support.** The brief says Life AI has run studies in diabetes, autism and stroke. Only the autism genotyping paper was found in public, so this table states needs and leaves supply to an audit.

| Use | Data it needs | Likely from existing studies? |
| --- | --- | --- |
| Population priors | Cross-sectional clinical and genomic data | Yes |
| Genotype-based enrichment | Genotypes linked to phenotype | Likely |
| Prognostic model for covariate adjustment | Outcomes over time under usual care, hundreds to thousands of patients | Unknown; depends on follow-up |
| Adherence and dropout model | Dosing and visit-level records | Unlikely unless interventional |
| Investigator and site model | Deviation logs, query logs, enrolment timelines across sites | No; needs partners |
| Treatment-effect transport | Local trial results | No; relies on published trials and bridging |

**The first experiment.** Measure the transport gap. Under covariate adjustment, the sample-size saving is about the squared correlation between the model's prediction and the observed outcome. A correlation of 0.3 saves 9%; 0.5 saves 25%. That is against an unadjusted analysis. What matters is the gain over adjusting for baseline HbA1c alone, measured in patients who meet trial criteria.

1. Take a prognostic model trained on non-Vietnamese data.
2. Score its correlation with observed outcomes in a Life AI cohort.
3. Re-fit the model on a local training split with local covariates, then re-score on a hold-out from another site. A simple rescaling would change nothing.
4. Report the difference as patients saved per trial.

This result is publishable and is what to offer a SURPASS team. It is planned for month 9, outside the demo.

**Indication fit**

| Indication | Local facts | Fit |
| --- | --- | --- |
| Type 2 diabetes | About 5 million adults; 35 registered Vietnamese trials; HbA1c is continuous; disease appears at lower BMI | Lead |
| Stroke | About 200,000 cases a year; 121,715 patients in the RES-Q registry; CYP2C19 affects antiplatelet response | Second. The 90-day disability scale is ordinal, outside the qualified method's scope. |
| Autism | 0.758% prevalence at 18–30 months; interventions mainly behavioural | Deferred. Use the genomic data for stratification research. |

**Legal limits on the data**

- Health records are sensitive personal data under Law 91/2025, in force since 1 Jan 2026. Any transfer abroad needs an impact assessment, and consent must be specific to the purpose.
- Sensitive data on 10,000 or more citizens is "important data". On 100,000 or more it is "core data", which needs a prior ministry assessment before leaving Vietnam. Important data needs an impact assessment filed 15 days before transfer.
- A draft Data Security Law would ban export of core data outright.

The design response is to train inside Vietnam and move only models and aggregates. Life AI's Life Cloud proposal uses the same pattern, with data held at hospital nodes.

## Appendix F. Sources

All pages below were opened during research on 1 Oct 2026. Competitor sources are linked in Appendix D.

**Limits on verification**

- SAM.gov was unreachable. SURPASS metrics come from a mirror of the solicitation PDF.
- ClinicalTrials.gov blocked access. Country counts come from a third-party mirror.
- Tufts CSDD figures come through trade-press summaries, not the original papers.
- Vietnamese legal texts were read on legal databases through an automated reader. Check article numbers and day counts before quoting.
- Life AI's data holdings rest on self-reported claims. No audited cohort size was found.
- Market-size figures are from market-research firms and differ fourfold.

**SURPASS and US policy**

- [ARPA-H press release, 30 Sep 2026](https://arpa-h.gov/news-and-events/hhs-launches-surpass-and-new-efforts-accelerate-faster-smarter-clinical-trials)
- [SURPASS programme page](https://arpa-h.gov/explore-funding/programs/surpass)
- [Solicitation ARPA-H-SOL-26-164, mirror copy](https://everglade.com/wp-content/uploads/SOL_26_164_SURPASSISO.pdf)
- [STAT on the undisclosed budget](https://www.statnews.com/2026/09/30/hhs-arpa-h-clinical-trials-artificial-intelligence-surpass-program/)
- [FDA real-time trial pilot](https://www.clinicalresearchnewsonline.com/news/2026/04/29/fda-piloting-real-time-review-of-clinical-trial-data-from-astrazeneca--amgen)
- [FDA draft guidance on AI for regulatory decisions](https://www.fda.gov/regulatory-information/search-fda-guidance-documents/considerations-use-artificial-intelligence-support-regulatory-decision-making-drug-and-biological)
- [FDA draft guidance on Bayesian methods](https://www.federalregister.gov/documents/2026/01/12/2026-00325/)
- [ICH M11 as final FDA guidance](https://www.federalregister.gov/documents/2026/05/22/2026-10295/)
- [ICH E6(R3)](https://www.fda.gov/regulatory-information/search-fda-guidance-documents/e6r3-good-clinical-practice-gcp)
- [CDISC USDM and Digital Data Flow](https://www.cdisc.org/ddf)

* [50 U.S.C. § 3059, definition of covered foreign country](https://www.law.cornell.edu/uscode/text/50/3059)

**Methods and benchmarks**

- [EMA qualification opinion on PROCOVA](https://www.ema.europa.eu/en/documents/regulatory-procedural-guideline/qualification-opinion-prognostic-covariate-adjustment-procovatm_en.pdf)
- [Digital twins and transportability, npj Digital Medicine 2026](https://www.nature.com/articles/s41746-026-02871-4)
- [Tufts: protocol deviations by phase](https://www.appliedclinicaltrialsonline.com/view/quantifying-protocol-deviation-experience-by-clinical-phase)
- [Tufts: enrolment performance](https://www.appliedclinicaltrialsonline.com/view/enrollment-performance-weighing-facts)
- [Tufts: amendment costs](https://intuitionlabs.ai/articles/clinical-trial-protocol-amendments-cost-data)
- [Tufts: site activation](https://www.globenewswire.com/news-release/2018/03/07/1417625/0/en/First-Comprehensive-Clinical-Site-Initiation-Benchmark-Developed-by-Tufts-Center-for-the-Study-of-Drug-Development-Finds-Study-Startup-Is-Lengthy-and-Inefficient.html)
- [Tufts: dropout and screen failure](https://www.appliedclinicaltrialsonline.com/view/can-recruitment-and-retention-get-any-worse)
- [CYP2C19 allele frequencies](https://www.frontiersin.org/journals/cardiovascular-medicine/articles/10.3389/fcvm.2022.991646/full)
- [WHO Asian BMI action points](https://research.vu.nl/en/publications/appropriate-body-mass-index-for-asian-populations-and-its-implica-2/)
- [Stroke subtypes in Asia](https://www.j-stroke.org/journal/view.php?number=163) and [Vietnam registry share](https://karger.com/cee/article/15/1/81/918886/Stroke-Epidemiology-in-Asia)
- [STAMPEDE at 25](https://www.innovative-ctu.ucl.ac.uk/news/news-stories/2024/september/25-at-25-transforming-prostate-cancer-treatment-through-the-stampede-clinical-trial)

**Vietnam: regulation and data law**

- [Circular 50/2025/TT-BYT on drug trials](https://thuvienphapluat.vn/van-ban/The-thao-Y-te/Thong-tu-50-2025-TT-BYT-quy-dinh-ve-thu-thuoc-tren-lam-sang-648617.aspx)
- [Circular 43/2024/TT-BYT on ethics committees](https://thuvienphapluat.vn/van-ban/Bo-may-hanh-chinh/Thong-tu-43-2024-TT-BYT-thanh-lap-Hoi-dong-dao-duc-trong-nghien-cuu-y-sinh-hoc-635867.aspx)
- [Pharmacy Law amendment 44/2024](https://www.tilleke.com/insights/vietnam-amends-law-on-pharmacy/)
- [Criteria for exemption from local trials](https://thuvienphapluat.vn/chinh-sach-phap-luat-moi/vn/ho-tro-phap-luat/chinh-sach-moi/85541/tieu-chi-de-xac-dinh-truong-hop-mien-thu-lam-sang-truoc-khi-cap-phep-luu-hanh-thuoc-tai-viet-nam)
- [Device and new-technique trial rules](https://thuvienphapluat.vn/chinh-sach-phap-luat-moi/vn/ho-tro-phap-luat/chinh-sach-moi/55109/quy-dinh-moi-ve-thu-nghiem-lam-sang-trong-kham-chua-benh-tu-01-01-2024)
- [Personal Data Protection Law 91/2025](https://www.dfdl.com/insights/legal-and-tax-updates/vietnam-personal-data-protection-2026-what-foreign-organizations-need-to-know/)
- [Law on Data: core and important data thresholds](https://www.morihamada.com/en/insights/newsletters/121451)
- [Draft Data Security Law](https://www.techtimes.com/articles/320437/20260714/vietnam-adds-fourth-data-law-banning-export-core-data-before-october-vote.htm)
- [AI Law 134/2025](https://www.allenandgledhill.com/vn/vn/perspectives/articles/32668/vnkh-vietnam-s-new-law-on-artificial-intelligence-risk-based-regulatory-framework-in-force-1-march-2026)
- [Electronic medical record uptake, Sep 2026](https://www.vietnam.vn/en/so-co-so-y-te-tham-gia-benh-an-dien-tu-tiep-tuc-tang)
- [VNeID health books](https://mst.gov.vn/30-trieu-so-suc-khoe-tren-ung-dung-vneid-dam-bao-phai-thuc-chat-thiet-thuc-197260322003007912.htm)
- [Ministry list of GCP-certified sites](https://asttmoh.vn/thuc-hanh-tot-thu-thuoc-tren-lam-sang-gcp/)

**Vietnam: market and disease context**

- [KPMG, Pharma Group and OUCRU roadmap](https://assets.kpmg.com/content/dam/kpmgsites/vn/pdf/2025/05/roadmap-clinical-trials-in-vietnam-en.pdf)
- [Trial counts by country](https://clinicaltrials.gg/locations/vietnam)
- [Vietnamese trial registrations 2010–2024](https://journal-format2.inforang.com/journal/view.html?doi=10.59931%2Frcp.25.0002)
- Market estimates: [The Report Cubes](https://www.thereportcubes.com/report-store/clinical-trials-cro-market-vietnam) and [Expert Market Research](https://www.expertmarketresearch.com/reports/vietnam-clinical-trials-market)
- [Diabetes: Ministry of Health figures](https://cand.vn/y-te/gan-5-trieu-nguoi-mac-dai-thao-duong-nhung-chi-23-3-duoc-quan-ly-dieu-tri-i674238/)
- [Stroke: RES-Q registry analysis](https://orca.cardiff.ac.uk/id/eprint/178443/), [thrombolysis rate](https://pmc.ncbi.nlm.nih.gov/articles/PMC10589571/), [ambulance arrival](https://www.angels-initiative.com/angels-community/stories/planning-delivers-progress-vietnam-1)
- [Autism prevalence](https://journals.sagepub.com/doi/full/10.4081/jphr.2021.2460)
- [Health financing](https://www.commonwealthfund.org/international-health-policy-center/countries/vietnam)
- [Traditional medicine share of visits](https://vietnamnews.vn/society/1782611/ministry-targets-stronger-role-for-traditional-medicine-in-public-care.html)
- [K Hospital clinical research centre](https://benhvienk.vn/gioi-thieu-trung-tam-nghien-cuu-lam-sang-nd91180.html)

**Life AI**

- [Autism genotyping study, Scientific Reports 2024](https://www.nature.com/articles/s41598-024-52777-y)
- [Company profile](https://www.globenewswire.com/news-release/2025/11/21/3192632/0/en/LIFE-AI-CEO-Tuan-Cao-Named-to-Tatler-Most-Influential-List-2025.html)