Survey data README — column dictionary and generation method

Simulated data constructed for portfolio purposes. Brightline Health Services is a fictional organization. The file survey-data-simulated.csv contains no real responses; every value was constructed by the script described below, designed so the dataset demonstrates the discovery patterns of the case study.

What the survey is (in the fiction)

A pulse survey fielded during discovery (2026-09), sampling 200 respondents across the 12 regional clinics: 130 staff (people who receive feedback) and 70 people managers (people who give it). It complements the interviews and document review.

Column dictionary

Column	Applies to	Meaning
respondent_id	all	S001–S130 (staff), M001–M070 (managers)
sample	all	staff or manager
clinic	all	One of the 12 fictional regional clinics
role	all	Staff: Registered Nurse, Medical Assistant, Front Desk, Allied Health, Clinic Admin. Managers: People Manager
tenure_band	all	Respondent’s own tenure at Brightline: <1 yr, 1-3 yrs, 3-10 yrs, >10 yrs (managers: blank in this extract)
manager_tenure_band	all	Tenure of the respondent’s manager (staff) or the manager’s own tenure (managers): <1 yr, 1-5 yrs, >5 yrs
team_size_band	all	Size of the respondent’s team: 12-13, 14-15, 16-18 (spans per the org chart extract)
regular_1on1s	staff	Does the respondent have a one-on-one with their manager at least every two weeks? yes/no
likert_1on1_freq	staff	1–5: how often a one-on-one happens — 1 = never, 2 = rarely, 3 = monthly, 4 = fortnightly, 5 = weekly or more
likert_feedback_freq	staff	1–5: how often the respondent receives feedback on their work — same scale
likert_feedback_useful	staff	1–5: the pulse item — “I received useful feedback in the last two weeks” (strongly disagree … strongly agree). Baseline metric
likert_workload	staff	1–5: workload manageability (1 = unmanageable, 5 = comfortable)
mgr_1on1_cadence	managers	Self-reported cadence of one-on-ones the manager holds: Weekly or more, Every two weeks, Monthly, Rarely, Never
mgr_pct_1on1s_structured_feedback	managers	Self-reported share (0–100) of the manager’s one-on-ones that include structured feedback (specific behavior + impact + agreed next step)

How it was generated (full method)

The file is produced by scripts/generate-survey.mjs and is byte-identical on every run:

1. Deterministic PRNG. A mulberry32 generator with fixed seed 20261025 drives every random choice. Math.random is never used.
2. Group profiles. Each of the 200 rows receives a designed profile: clinic (weighted counts per clinic), role, tenure bands, team-size band.
3. Designed correlations. Each staff Likert item is assigned by sorted assignment: rows are ranked by a propensity score that encodes the discovery patterns (regular one-on-ones and smaller teams rank higher on feedback usefulness; manager tenure has only a weak effect; the Cedar Point rows are deliberately penalized), then the exact calibrated marginal counts are poured over the ranked rows.
4. Exact baselines. The marginals are calibrated so the locked case numbers hold exactly:
   • mean of likert_feedback_useful (staff) = 2.40/5 — the pulse baseline;
   • 22 of 70 managers (31%) report a fortnightly-or-better one-on-one cadence;
   • mean of mgr_pct_1on1s_structured_feedback ≈ 20% — “structured feedback in about 1 in 5 one-on-ones”;
   • 45 of 130 staff (35%) report regular one-on-ones (staff-side view of the same reality, with sampling noise).
5. Designed wrinkles (per the case plan: “add noise and contradictions; include one finding that doesn’t fit”):
   • Cedar Point: all 12 staff report frequent one-on-ones (mean 1:1 frequency 4.3/5) yet score lowest on feedback usefulness (1.5/5); its managers meet weekly but report almost no structured-feedback content. Meetings are not the same as feedback.
   • The Glenrose outlier: one brand-new manager reports a fortnightly cadence and 55% structured content — experience is not the deciding variable.
6. Reproducibility check. Run node scripts/generate-survey.mjs --verify — the script recomputes the statistics and exits non-zero if any locked value drifts.

What this dataset is for — and what it can’t do

• It can illustrate how the discovery evidence was assembled and analyzed (see the survey analysis).
• It cannot show anything about real organizations; it was constructed to support the case narrative. The correlations are designed, not discovered — that is disclosed so the analytic method (segmenting by structure, not blaming experience) can be judged on its own merits.