BIOGUARD AIA product of Omni Care Oasis LLC

Research Use Only · Interpretation layer

The model is the easy part. Knowing when it must not answer is the product.

BioGuard AI™ profiles biomarker patterns across a 12-marker panel and explains which markers contributed to each assessment — working only on measurements that have already passed AssuranceView QC™. Its current model is fitted to synthetic data and has never been applied to a human specimen, so the part that is finished today is everything surrounding the model: what it may emit, what it is forbidden to emit, when it declines to compute, and who must sign before a result leaves the system.

Validation partnership What it refuses to do
BioGuard AI — a product of Omni Care Oasis LLC

Where this actually stands

Current model status

0Human specimens The model has never been applied to one.
12Markers in the panel Fixed order; a partial panel is a different input and is refused.
SyntheticTraining data Generated by train.generate_synthetic_cohort, 14 August 2026.
0.1.0-syntheticModel version Contract 2.0.0. Fingerprint 60d673d8091c8b04.
Stated plainly

This model's own metadata records clinically_validated: false. It was trained on a cohort this project generated, so any accuracy figure measured against that cohort describes how well the model recovered a pattern we invented. It is not evidence about cancer, and no such figure is published on this page for that reason.

BioGuard AI is not a screening test, not a diagnostic, and not a risk calculator. It is a governed interpretation layer awaiting the specimen data that would let anyone make a performance claim at all. Acquiring that data, under an approved protocol with a clinical partner, is the current objective.

What it does

Five capabilities, each marked with what is actually built

BioGuard AI™ is a research-stage analytical intelligence system. It is intended to support research and future clinical evaluation — not to independently diagnose cancer, and not to replace a qualified healthcare provider. Three of the five capabilities below are implemented and under test. Two are not, and are marked as such rather than described in the present tense.

Profile biomarker patterns

Returns a composite feature index across the 12-marker panel with a per-marker contribution for each.

Built
Explain which biomarkers contributed to each assessment

SHAP attribution, per marker, against the fitted model. Served in every mode and foregrounded in explain.

Built
Analyse only data that have passed AssuranceView QC™

Stronger than it sounds: the quality verdict is read from the recorded batch, not accepted as an assertion in the request. Data that failed, or whose verdict is unknown, abstains.

Built
Monitor changes over time

Not built. feature_trend is an approved output class in app/policy.py that nothing currently emits, and no endpoint accepts a prior result to compare against. Listed here because it is the next capability, not because it exists.

Roadmap
Track emerging cancer-related signals

Not built. Depends on the longitudinal comparison above.

Roadmap
Identify potentially concerning patterns associated with early-stage cancer

This is what FFT-E™ is being built to do, and it is the purpose the validation work serves. It is not what BioGuard AI emits today, and saying otherwise would contradict the code: flagging a pattern as concerning is a risk category, a patient classification or a screening result depending on how it is phrased, and all three are prohibited output classes in both app/policy.py and app/governance.py. The current output carries no threshold and no category. Changing that requires outcome data this project does not have.

Platform objective

Where this sits

BioGuard AI is the interpretation engine inside FFT-E

FFT-E™ is Omni Care Oasis’s blood-based multi-cancer early detection platform. Specimen stabilisation, integrity screening and fluidics are separate components with separate jobs; BioGuard AI is the last stage, and it only ever sees measurements that the stages before it have already cleared.

pH-Lock™

Specimen stabilisation at collection, holding the pH window the downstream assays were established in.

SIP™

Specimen integrity gate. Runs before the model, not after: a specimen outside the hemolysis or pH limits is never scored, because a plausible-looking number computed on a degraded sample is worse than no number.

OFM™ cassette

40-zone cassette with passive pulsatile flow, carrying the ELISA and qRT-PCR chemistries.

BioGuard AI™

Interpretation. Takes measurements that passed analytical QC and returns a composite feature index with per-marker attribution — or names the condition under which it will not.

The panel

Twelve markers in three tiers, across two chemistries

The tier decides how a marker is treated, not merely how it is labelled. Every marker declares the unit its range was established in, and a value supplied in a different unit abstains rather than being converted — the number is not wrong, it means something else.

12 markers · defined in app/markers.py

Core 4Measured on every specimen. Confirm mode will not return an output without their support.
PSAELISA · ng/mL
CEAELISA · ng/mL
HE4ELISA · pmol/L
miR-21qRT-PCR · dCt
Enhanced 5Added for breadth across the target cancers.
CA19-9ELISA · U/mL
AFPELISA · ng/mL
MMP-9ELISA · ng/mL
CA-125ELISA · U/mL
HER4ELISA · ng/mL
Zulu Package™Circulating microRNA, read by qRT-PCR.
miR-155qRT-PCR · dCt
miR-196aqRT-PCR · dCt
miR-10bqRT-PCR · dCt

Operating modes

Three modes, none of which return a decision

The modes change what evidence is demanded before an output is produced, not what kind of output it is. All three return a composite feature index with attribution and no threshold; none returns a screening call, a risk band or a classification, because those are prohibited output classes regardless of mode.

Attribution is SHAP, computed per marker against the fitted model.

screen

Favours sensitivity. Still returns a composite feature index and attribution — never a screening call, which is a prohibited output class.

confirm

Demands Core 4 support before an output is produced at all.

explain

Per-marker SHAP attribution for the index, so a researcher can see which markers moved it and by how much.

Default deny

Output classes it is forbidden to emit

The service permits output classes by enumeration, never by omission. An output class nobody has thought of yet is refused by default, because the check asks whether a class is on the permitted list rather than whether it is on a banned one. These eight are named explicitly so that a refusal can explain itself in the words a reviewer would use.

8 prohibited classes · enforced in app/governance.py · asserted by the test suite

diagnosis

A statement that a subject has or does not have a disease.

patient_facing_summary

Output addressed to a patient rather than to a researcher. Permitted only under an intended use that does not currently exist.

prognosis

A statement about expected course or outcome.

risk_category

A banding of a subject into low/moderate/high risk. This requires a threshold established on outcome data, which does not exist.

risk_score

A calibrated probability of disease. The composite index is not calibrated to outcomes and must not be presented as though it were.

screening_result

A positive or negative screening call.

treatment_recommendation

Any suggestion about clinical management.

triage_priority

An ordering of subjects for clinical attention.

Refusal to compute

Conditions under which it returns nothing at all

Abstention is not an error path. Each condition below describes an input the model was never evaluated against, and in every one of them the service returns the named condition instead of a number. A result produced outside the envelope would look exactly like one produced inside it, which is the whole reason the envelope has to be declared before anything is computed.

8 abstention conditions · enforced in app/interpretation.py

marker_out_of_range

One or more marker concentrations fall outside the range over which the model's performance was established. Reporting a value here would be extrapolation presented as a result.

matrix_mismatch

The specimen matrix differs from the one the model was evaluated on.

measurement_system_mismatch

The measurement system differs from the one that produced the model's training data. Recovery characteristics, precision and lot variation differ between systems; revalidation against the new system is required before its output can be interpreted.

missing_required_marker

One or more markers the model requires were not measured. The composite was evaluated on a complete panel; a partial panel is a different input.

no_envelope

No validation envelope was declared. The conditions under which this model may be applied are unknown.

population_outside_envelope

The subject falls outside the population the model was evaluated on.

qc_gate_not_passed

The analytical quality gate was not passed. Data that failed QC is not interpreted — interpretation of unreliable measurements produces an unreliable result that looks reliable.

unit_mismatch

A marker was supplied in a unit other than the one its range was established in. The number is not wrong; it means something else. Converting silently would hide a reporting-system difference that the laboratory needs to know about.

Human in the loop

When a result may not leave without a named reviewer

Escalation never de-escalates: where several conditions apply, the most senior role requested is the one required. The reviewer's conclusion is recorded and audited, including when the reviewer disagrees — a disagreement is preserved as a finding rather than resolved by recomputation.

6 escalation conditions

abstention_rate_elevated

This batch abstained far more often than the model's established rate, which usually means the input differs from what was validated.

all_markers_at_bound

Every marker sits at the edge of its evaluated range. The result is inside the envelope by definition and at its limit in practice.

first_use_of_model_version

This model version has not been used before. The first output from a newly released model is reviewed regardless of its content.

qc_failure_present

The analytical quality gate failed or was unknown. Someone must see that a result was withheld, because a withheld result and a result nobody requested look identical downstream.

reviewer_disagreement

A reviewer recorded disagreement with the output. The disagreement is a finding and is preserved, not resolved by recomputation.

version_disagreement

A previous interpretation of the same input under a different model version differs materially. Which is right is not a question the software can settle.

Permitted

The complete list of what it may produce

Three classes. Everything else, including anything not yet imagined, is refused until someone adds it here deliberately.

abstention

A refusal to compute, with the conditions that caused it named.

feature_attribution

A composite feature index with per-marker attribution, carrying no threshold and no category.

quality_assessment

An analytical quality verdict on measurement data, with the criteria applied.

Evidence record

How the behaviour above is held in place

None of this is a promise in a document. The refusals are code paths, and the suite asserts them. Measured 6 October 2026 against service 0.14.5.

527Tests passing 0 failing.
97%Mutation score 97 of 100 injected faults caught by a test.
96.54%Line coverage Of 3,121 statements. Reported, not gated.
all passQuality gates Coverage, module floor, mutation, failing tests.

Mutation score is the gate. Faults are injected into the governance and interpretation code and the suite must catch them; a suite that cannot is not evidence, whatever its coverage. Coverage is reported rather than gated because it measures which lines executed, not which behaviour was asserted, and gating on it rewards tests that assert nothing. Reproduce every figure with python measure_quality.py --mutation. The run log is append-only and hash-chained, which makes alteration detectable — not impossible.

What would change this

The missing input is specimens, not engineering

Every refusal on this page stays in force until a model is fitted to real measurements from a declared population, on a declared measurement system, with a validation envelope established from outcome data. Omni Care Oasis is seeking clinical and repository partners for exactly that step. Enquiries from institutions, repositories and reviewers are welcome.

Intended use — as the service states it

Omni AssuranceView QC and the BioGuard AI interpretation layer are for Research Use Only and are not for use in diagnostic procedures. The interpretation layer produces a composite feature index with per-marker attribution, for use by a qualified researcher, on measurement data that has passed an analytical quality gate. It does not diagnose, does not assign risk categories or calibrated probabilities, does not screen, and does not address a patient. Every output requires a named human reviewer before it is used or reported.

Served verbatim at GET /governance/intended-use · contract 2.0.0