SIF Balanced Scorecard

A leading / monitoring / lagging safety measurement toolkit for serious injury and fatality (SIF) prevention — a working implementation of the framework validated in Bayona, Hallowell, Bhandari & Raheemy (2026), "Balanced approach to serious injury and fatality prevention," Journal of Safety Research 98, 175–189 (open access).

▶ Try the live dashboard — no install needed.

Why

Most organizations still steer safety by TRIR — a lagging, frequency-based rate that weighs a paper cut the same as a fatality and says nothing about tomorrow's risk. The 2026 study measured 31 utility/construction companies (361M worker-hours, 999 field assessments) and showed that two short-term measures carry real signal about long-term SIF outcomes:

Lens	Question	Instrument	Key result
Leading	How safe will the job be?	PJSB quality — 15-item weighted pre-job brief scorecard (CSRA)	+1 pt PJSB ≈ +0.49 pt HECA (p = .04)
Monitoring	How safe is the job?	HECA — % of high-energy hazards (>1,500 J) with a Direct Control	+1 pt HECA ≈ −3% expected SIFs (p = .02)
Lagging	How safe was the job?	TRIR + SBLI (severity-weighted rate)	Weighs injuries by consequence, not just count

This repo implements all three instruments, the empirical risk model, the assessor-reliability gates, and two delivery surfaces:

1. Python package + Streamlit dashboard — for analysts.
2. Microsoft 365 workflow — a Copilot agent that scores pre-job briefs from Teams transcripts, SharePoint capture lists, Power Automate alerting, and a Power BI scorecard — for EHS teams to run every day with tools they already own.

Quickstart

pip install -e ".[dashboard,dev]"
pytest                                # 39 tests
python examples/replicate_paper.py    # fit the paper's models on synthetic data
streamlit run dashboard/app.py        # three-lens scorecard dashboard

Score a pre-job brief

from sif_scorecard import score_pjsb

result = score_pjsb({1: True, 2: True, 3: True, 4: True, 5: True,
                     6: False, 7: True, 8: False, 9: True, 10: True,
                     11: True, 12: True, 13: True, 14: True, 15: False})
print(f"{result.quality:.0%}")          # 79%  (45/57 weighted points)
print(result.missing_statements[0])     # "8. All life-threatening hazards..."

Project SIF risk from field observations

from sif_scorecard import HazardObservation, heca_score, expected_sifs_per_1000_fte

task = heca_score([
    HazardObservation("Suspended load - crane pick", direct_control_present=True),
    HazardObservation("Fall from 12 ft - leading edge", direct_control_present=False),
])
print(task.score)                                # 0.5
print(expected_sifs_per_1000_fte(task.score))    # ~0.56 expected SIFs

Gate your assessors before trusting their data

from sif_scorecard import cohens_kappa, assessor_gate

kappa = cohens_kappa(rater_a_items, rater_b_items)
print(assessor_gate(kappa))   # qualified only if kappa > 0.40 (study's rule)

What's in the box

src/sif_scorecard/
  pjsb.py         CSRA 15-item scorecard with official weights (max 57 pts)
  heca.py         High-Energy Control Assessment + Direct Control definitions
  lagging.py      TRIR and severity-based SBLI (Eq. 1-2)
  risk.py         Fig. 7 risk curve, PJSB->HECA->SIF pathway, coaching bands
  reliability.py  Cohen's kappa, ICC(2,k), the >0.40 assessor gate
  synthetic.py    Company-panel generator calibrated to the paper's Table 6
  models.py       Poisson / zero-inflated Poisson GLMs (Model 6 replication)
dashboard/        Streamlit three-lens scorecard + what-if projection
m365-implementation/
  QUICKSTART.md   No-code setup for EHS pros new to Copilot (one afternoon)
  copilot-agent/  Declarative agent that scores briefs from Teams transcripts
  sharepoint/     Capture list schemas
  power-automate/ Scoring, HECA<30% alerting, weekly digest flows
  power-bi/       DAX measures for the three-lens report
examples/         End-to-end replication of the paper's analysis
data/             Synthetic 31-company sample (no real data anywhere)

Does the replication actually work?

examples/replicate_paper.py generates a synthetic panel calibrated to the paper's descriptive statistics (Table 6), then fits the paper's count models. The zero-inflated Poisson recovers the headline coefficient almost exactly:

Relationship	Paper (Table 8)	Recovered on synthetic panel
SIF ~ HECA	−0.03 / pt (p = .02)	−0.029 / pt (p < .001)
FA ~ HECA	−0.02 / pt	−0.022 / pt
FT ~ HECA	−0.08 / pt	−0.037 / pt

And the implemented risk curve reproduces the paper's Fig. 7 anchors: HECA 0 → 2.5 expected SIFs per 1,000 FTE, 0.5 → ~0.56, 0.9 → ~0.17.

Operating thresholds encoded in the library

• HECA < 30% → coaching zone: "a SIF becomes a likely event."
• Baseline protocol: ≥15 PJSB + ≥15 HECA assessments, randomly sampled tasks/crews, within ≤3 months, before treating averages as stable.
• Assessor gate: Cohen's κ or ICC > 0.40, or the data doesn't count.

Caveats

The published relationships are correlational, company-level associations from North American utility and construction firms. They justify measuring and improving brief quality and Direct Control coverage; they don't certify any site as safe, and transfer to other industries is an open question. All data in this repository is synthetic.

Attribution

• Framework and empirical coefficients: Bayona, A., Hallowell, M. R., Bhandari, S., & Raheemy, Y. (2026). Journal of Safety Research, 98, 175–189 (CC BY-NC-ND 4.0).
• Pre-Job Safety Meeting Scorecard: Construction Safety Research Alliance (CSRA), University of Colorado Boulder.
• High-energy hazard / Direct Control concepts: Oguz Erkal & Hallowell (2023); EEI's The Power to Prevent.

Code is MIT-licensed. The scorecard instruments belong to their authors; this repo implements them for research and educational use with attribution.