- Decision
- What action or allocation will this inform?
- Unit
- Organization × time—never a person
- Outcome
- What will be measured, and from which record?
- Owner
- Who decides, supplies data, and reviews?
- Success
- What changes, by when, and with what safeguard?
MP2 Field Lab · working engagement model
From command question to evidence-backed action.
Bring a bounded command question. Prepare approved unit-level data. Load it through a controlled intake. Analyze it transparently. Decide with context. Measure what happens next.
01 Engage
Define the decision and the question.
Start with the decision, not the algorithm. The command and MP2 define what must be understood, the organizational level, the time horizon, who owns the decision, and what useful learning would change.
- Input
- Decision need, unit and time horizon, outcome, current practice
- Output
- One bounded question, named owners, success measure, timeline
- Gate
- Lawful, feasible, unit-level, and tied to a real decision
02 Approve
Confirm authority before any transfer.
“Unclassified” and “anonymized” do not automatically mean shareable. The data owner and MP2 must confirm ownership, classification and CUI status, permitted use, access, retention, disclosure risk, and any required human-subjects review.
- Input
- Source inventory, data owner, intended use, handling needs
- Output
- Approved intake and handling plan with accountable owners
- Gate
- No transfer until authority, classification, access, and retention pass
Stop: classified information, person-level rows, clinical details, case narratives, identifiable operational information, or data intended for surveillance, discipline, diagnosis, or personnel action.
- 01OwnershipAuthorized data owner identifiedConfirm
- 02HandlingClassification and CUI review completeConfirm
- 03PurposePermitted use and research status recordedConfirm
- 04LifecycleAccess, retention, export, and deletion approvedConfirm
03 Prepare
Structure the approved data.
The original Unit Risk Forecasting Tool organized observations by company-, troop-, or battery-level week. MP2 translates that foundation into a platform-neutral, pseudonymous unit-period schema with exposure, approved aggregate measures, provenance, and explicit missingness.
- Input
- Approved aggregate extracts and documented source definitions
- Output
- Versioned unit-period file, data dictionary, provenance record
- Gate
- No names, UICs, case narratives, exact timestamps, or person-level rows
| unit_token | period_start | exposure | outcome | signal_01 | complete |
|---|---|---|---|---|---|
| UNIT-A17 | 2026-01-05 | 120 | 0 | 1 | TRUE |
| UNIT-A17 | 2026-01-12 | 118 | 1 | 2 | TRUE |
| UNIT-K42 | 2026-01-05 | 96 | 0 | 0 | TRUE |
| UNIT-K42 | 2026-01-12 | 97 | — | 1 | FALSE |
Stable synthetic tokens · aggregate counts · explicit missingness
04 Load
Validate, review, then accept.
A future controlled pilot would send approved files to a quarantined landing zone—never directly to analysis. Automated checks verify the file, schema, ranges, duplicates, missingness, and disclosure risk. A human steward accepts, returns, or rejects the submission.
- Input
- Versioned file, dictionary, approval record, source metadata
- Output
- Preflight report and frozen analysis-ready panel
- Gate
- Automated checks and human steward review must both pass
Current boundary: this public page has no account, file input, upload endpoint, data store, or live workspace.
- 01Local preflightSchema, types, ranges, missingness
- 02QuarantineSignature, malware, size, duplicate checks
- 03Steward reviewAuthority, disclosure risk, disposition
- 04Analysis vaultApproved, isolated, logged, time limited
Rejected data never enters analysis
05 Analyze
Build, test, and document the model.
Analysis starts with quality and descriptive checks, then uses the simplest method suited to the question. The public URFT base model used logistic regression to relate recent unit-level incident history to a current-week recorded outcome. It is a documented foundation—not a portable threshold or causal claim.
- Input
- Frozen panel, analysis plan, temporal holdout, code version
- Output
- Reviewed results, uncertainty, diagnostics, reproducibility record
- Gate
- Method fit, leakage, calibration, errors, and independent review
Every run preserves the data hash, code and environment version, parameters, diagnostics, analyst, reviewer, and release state.
- Question
- Registered before modeling
- Dataset
- Frozen file + SHA-256 fingerprint
- Method
- Specified model, features, and comparison
- Validation
- Forward-in-time holdout + diagnostics
- Review
- Analyst, independent reviewer, approved use
- DescribeRates, missingness, distributions
- PredictTemporal validation and calibration
- EvaluateDesign, comparison, sensitivity
06 Decide
Review the evidence and act with care.
Commanders, data owners, analysts, prevention professionals, and independent reviewers work from the same evidence canvas. The model contributes one signal; local conditions, uncertainty, ethics, available support, and implementation constraints remain visible beside it.
- Input
- Reviewed results, data limits, local context, support options
- Output
- Human-led action, accountable owner, measures, check-back date
- Gate
- Supportive and nonpunitive; never triggered automatically by a score
MP2 supports inquiry and learning. Commanders retain decision authority, and people retain responsibility for context, judgment, and care.
Trend, uncertainty, quality, provenance
Operations, climate, reporting, capacity
Benefits, risks, owner, safeguards
Measure, check-back, decision log
Review · challenge · decide · record
07 Learn
Measure the action and close the cycle.
The cycle ends only when the organization records what was implemented, measures aggregate outcomes, tests plausible explanations, captures limitations, and makes an explicit release and data-disposition decision. The next command question should inherit what the prior cycle learned.
- Input
- Implementation log, follow-up data, comparison plan, limitations
- Output
- Sustain, adapt, or stop decision; closeout and next question
- Gate
- Claim review, release approval, and verified retention or deletion
01ActImplement human-led support02MeasureTrack aggregate outcomes03EvaluateTest change and alternatives04AdaptRecord and improve
Kit Commander templates
Working products informed by the URFT workflow.
The historical URFT paired a structured unit-week dataset, repeatable analysis, leader-facing signals, command interpretation, tailored action, and evaluation. MP2 preserves that operating logic while replacing real identifiers and operational details with blank or synthetic public templates.
| Stage | Working product | Purpose | Format | Download |
|---|---|---|---|---|
| 01 · Engage | Command question brief | Decision, owner, unit, window, outcome, success | MD | Download |
| 02 · Approve | Data readiness + authority checklist | Purpose, classification, CUI, access, retention, review | MD | Download |
| 03 · Prepare | Synthetic unit-period template | Coded unit keys, aggregate counts, provenance, missingness | CSV | Download |
| 03 · Prepare | Data dictionary | Types, requirements, controlled values, handling notes | CSV | Download |
| 05 · Analyze | Analysis plan + model card | Claim type, split, diagnostics, assumptions, approved use | MD | Download |
| 06 · Decide | Evidence + action canvas | Evidence, context, decision, owner, safeguards, review | MD | Download |
| 07 · Learn | Evaluation + closeout plan | Comparison, timing, measures, export, deletion | MD | Download |
Gate Capability states
One workflow. Different trust zones.
Workflow, synthetic templates, method education, and canvas prototype.
Invite-only aggregate, non-CUI analysis in a separate governed environment.
Restricted work, if ever approved, stays inside a sponsor-authorized environment.
Classified data; individual surveillance, scoring, diagnosis, discipline, or personnel action.