Onboarding

MAKER-1 sepsis early warning 1.0.0 · 3 claim contracts · 12 failure modes · policy pol.sepsis-micu 1.2.0

Claim contracts

from 4 card claims

cc.discrimination1.0.0prognostic prediction

Card says: AUROC 0.81 for sepsis prediction

AUROC at or above 0.8107 on adult ICU patient-hours at this site

Clinical endpoint: Sepsis-3 onset during the ICU stay
Decision timepoint: each ICU hour, observations up to that hour
Intended population: adult ICU stays at the deployment site
Intended inputs: hourly vitals, laboratory results, demographics
Output semantics: continuous risk score in [0,1] for sepsis onset
Performance: auroc gte 0.8107 at 95% confidence
Reference standard: PhysioNet/CinC 2019 Sepsis-3 label, shifted 6h ahead of clinical onset
Adjudication: corpus reference label, no local re-adjudication
Maturity window: 6h

corpus label; no local chart review
label timing set by the challenge

cc.alert-precision1.0.0diagnostic detection

Card says: ~20% of alerts are true at the shipped threshold

At the shipped threshold, 20% or more of alerting patient-hours fall in a stay with adjudicated sepsis

Clinical endpoint: Sepsis-3 onset during the ICU stay
Decision timepoint: the hour the alert fires
Intended population: adult ICU stays at the deployment site
Intended inputs: hourly vitals, laboratory results
Output semantics: alert when the score reaches the shipped threshold
Performance: ppv gte 0.2007 at 95% confidence
Reference standard: PhysioNet/CinC 2019 Sepsis-3 label, shifted 6h ahead of clinical onset
Adjudication: corpus reference label, no local re-adjudication
Maturity window: 6h

corpus label; no local chart review
label timing set by the challenge

cc.lead-time1.0.0prognostic prediction

Card says: predicts sepsis before clinical onset

First alert precedes adjudicated onset in most alerting stays

Clinical endpoint: hours from first alert to adjudicated Sepsis-3 onset
Decision timepoint: the hour of the first alert for a stay
Intended population: adult ICU stays that alert at the shipped threshold
Intended inputs: hourly vitals, laboratory results
Output semantics: hours from first alert to onset, positive means early
Performance: lead_time gte 0
Reference standard: PhysioNet/CinC 2019 Sepsis-3 label, shifted 6h ahead of clinical onset
Adjudication: corpus reference label, no local re-adjudication
Maturity window: 6h

corpus label; no local chart review
label timing set by the challenge

Claims that produced no contract

M-4 performance generalises to new hospital systems — no supporting measurement provided

No endpoint, metric or reference standard. No contract issued.

Authority bundle

ab.sepsis-w2 1.0.0

Source	Type	Version	Local
PhysioNet/CinC 2019 Sepsis-3 label definition auth.sepsis3-cinc2019	reference dataset	1.0.0	—
Physiologic plausibility ranges for adult ICU vital signs auth.vitals-ranges	validated rule set	1.0.0	—
SIRS criteria (ACCP/SCCM consensus, Bone et al. 1992) auth.sirs-accp-sccm	diagnostic criteria	1992.1	—
qSOFA (Sepsis-3, Singer et al. JAMA 2016) auth.qsofa-sepsis3	diagnostic criteria	2016.1	—
Organ dysfunction markers: lactate, creatinine, platelets, MAP auth.organ-dysfunction	validated rule set	1.0.0	—
Maximum observation age for a risk score auth.local-freshness	hospital policy	1.0.0	override

Checks cite only these sources. Model output can extract or infer; promoting it to a rule requires clinical approval and a new bundle version.

Failure mode profile

fmp.sepsis-w2 1.0.0

Mode	Category	Severity	Default action
score produced with a required vital sign absent fm.missing-vitals	missing input	high	unable to verify
most recent required observation exceeds the policy age limit fm.stale-vitals	stale input	warning	flag
input outside physiologic screening bounds fm.implausible-value	corrupt input	high	flag
patient outside the approved population fm.out-of-population	out of distribution	high	unable to verify
operating threshold differs from the approved policy fm.threshold-drift	threshold error	critical	hold
no alert on a stay with adjudicated sepsis onset fm.false-negative	false negative	critical	review
alert on a stay with no adjudicated sepsis onset fm.false-positive	false positive	warning	review
first alert at or after adjudicated onset fm.late-alert	false negative	high	review
qSOFA is 2 or more while the model is silent fm.silent-while-criteria-met	false negative	critical	review
model is alerting while no independent screening criterion is met fm.alert-without-criteria	false positive	warning	review
too few screening components are available to form an independent view fm.criteria-not-evaluable	missing input	high	unable to verify
subgroup discrimination below the population estimate fm.subgroup-gap	subgroup gap	warning	review

Mode

Verification policy

pol.sepsis-micu 1.2.0 · mode validation

Intended use: Hourly sepsis risk scoring for adult medical ICU patients
Workflow: score written to the chart, no automated order or page
Population: adult ICU stays, age 18 and over, medical ICU
Permitted: display a risk score to the care team · populate a surveillance worklist
Prohibited: initiate an order · page without human review · apply to paediatric or obstetric patients
Latency target: 1500ms, on failure unable to verify

Checks

Check	Verification claim	Decision mode	Required	If evidence missing
chk.input-completeness	all required inputs present at score time	direct deterministic	required	unable to verify
chk.input-freshness	no required observation past the policy age limit	direct deterministic	required	review required
chk.input-plausibility	inputs within physiologic screening bounds	direct deterministic	required	skip with limitation
chk.population-scope	patient within the approved population	direct deterministic	required	unable to verify
chk.threshold-replay	alert decision replays from score and approved threshold	direct deterministic	required	fail closed
chk.sirs	SIRS criteria met, counted from observations at this hour	direct deterministic	optional	skip with limitation
chk.qsofa	qSOFA components met, counted from observations at this hour	direct deterministic	optional	skip with limitation
chk.organ-dysfunction	organ dysfunction markers outside threshold at this hour	direct deterministic	optional	skip with limitation
chk.independent-signal	the model's alert state agrees with an independent screening view formed from published criteria	direct deterministic	required	unable to verify
chk.alert-support	alert supported by locally measured precision	statistical calibrated	optional	review required

Check

Verification claim

Decision mode

Required

If evidence missing

chk.input-completeness

all required inputs present at score time

direct deterministic

required

unable to verify

chk.input-freshness

no required observation past the policy age limit

direct deterministic

required

review required

chk.input-plausibility

inputs within physiologic screening bounds

direct deterministic

required

skip with limitation

chk.population-scope

patient within the approved population

direct deterministic

required

unable to verify

chk.threshold-replay

alert decision replays from score and approved threshold

direct deterministic

required

fail closed

chk.sirs

SIRS criteria met, counted from observations at this hour

direct deterministic

optional

skip with limitation

chk.qsofa

qSOFA components met, counted from observations at this hour

direct deterministic

optional

skip with limitation

chk.organ-dysfunction

organ dysfunction markers outside threshold at this hour

direct deterministic

optional

skip with limitation

chk.independent-signal

the model's alert state agrees with an independent screening view formed from published criteria

direct deterministic

required

unable to verify

chk.alert-support

alert supported by locally measured precision

statistical calibrated

optional

review required

Decision table

Priority	When	Verdict	Reason code
10	any_check_status = error	unable to verify	check_error
20	any_failure_mode_action = unable_to_verify	unable to verify	required_evidence_absent
30	any_failure_mode_action = hold	hold	policy_violation
40	any_clinical_assessment = review_required	flag	review_required
50	any_clinical_assessment = clinically_disputed	flag	disputed
90	otherwise = true	no violation in scope	no_violation_in_checked_scope

Priority

When

Verdict

Reason code

any_check_status = error

unable to verify

check_error

any_failure_mode_action = unable_to_verify

unable to verify

required_evidence_absent

any_failure_mode_action = hold

hold

policy_violation

any_clinical_assessment = review_required

flag

review_required

any_clinical_assessment = clinically_disputed

flag

disputed

otherwise = true

no violation in scope

no_violation_in_checked_scope

Dispositions

Verdict	Action	Target	Landing SLA	Closure required
no violation in scope	record	verification_ledger	60s	closed
flag	route review	sepsis_stewardship_queue	3600s	reviewed
hold	escalate	ai_governance_oncall	3600s	acted
unable to verify	record	verification_ledger	60s	closed

Verdict

Action

Target

Landing SLA

Closure required

no violation in scope

record

verification_ledger

60s

closed

flag

route review

sepsis_stewardship_queue

3600s

reviewed

hold

escalate

ai_governance_oncall

3600s

acted

unable to verify

record

verification_ledger

60s

closed

A flag is routed to sepsis_stewardship_queue and lands only when it reaches reviewed inside 60 minutes. Nothing else counts, so an auto-close cannot move Landing.

GHM-0001 · sha256 4cd3e90a9cb7e2be · contract 2.0.0 validated