Capstone: quality metrics, culture and a quality plan

The previous lessons followed Larkspur Commons from its quality plan through inspections, tests, reviews, nonconformances and deficiencies. This capstone steps back to ask two questions. How does a team know whether its quality program is working, and how does it keep improving? And what makes people on site actually care about quality, rather than treating it as paperwork? The lesson ends by bringing everything together in a quality plan for Thornbury's next project, a building the team is about to start.

Measuring quality

Lesson 1 introduced first-time quality and the cost of quality. A small set of measures, tracked monthly and discussed openly, tells the team far more than a thick binder of inspection forms. Lena's quality dashboard for Larkspur Commons uses the measures below. Each has a clear definition so that results can be compared over time and between projects.

Numbers can be gamed. If foremen are judged only on the NCR count, they will stop reporting NCRs. That is why the dashboard pairs measures that pull in different directions: the NCR count goes with the self-reported share, and the first-time pass rate goes with the number of inspections done. A project with few NCRs and a falling self-reported share is not necessarily improving; it may be hiding problems.

Control charts for first-time quality

Weekly first-time pass rates go up and down. Some of that is noise, and some is a real change. A control chart helps tell them apart. For a proportion such as the pass rate, the usual chart is a p-chart: the centre line is the average pass rate, and the control limits are set about three standard errors either side, with the standard error depending on the number of items inspected each week. A point outside the limits, or a run of points on one side of the centre line, signals a real change worth investigating. PMP-131 introduced control charts; here is one applied to the contractor's inspections on the wood-frame floors.

# p-chart of weekly first-time pass rate, contractor inspections on the wood-frame floors (illustrative)
import math
weeks = [  # (items inspected, passed first time)
    (118, 99), (131, 112), (126, 110), (140, 121), (122, 102), (135, 120), (129, 116),
    (138, 104), (133, 119), (127, 116), (141, 129), (136, 126), (130, 121), (139, 130),
]
n_tot = sum(n for n, _ in weeks); p_tot = sum(p for _, p in weeks)
pbar = p_tot / n_tot
print(f"average {pbar:.3f}")
for w, (n, p) in enumerate(weeks, 1):
    se = math.sqrt(pbar * (1 - pbar) / n)
    lcl, ucl = max(0, pbar - 3 * se), min(1, pbar + 3 * se)
    rate = p / n
    flag = "BELOW LCL" if rate < lcl else ("ABOVE UCL" if rate > ucl else "")
    print(f"week {w:2d}: {rate:.3f}  limits {lcl:.3f}-{ucl:.3f}  {flag}".rstrip())

average 0.881
week  1: 0.839  limits 0.791-0.970
week  2: 0.855  limits 0.796-0.966
week  3: 0.873  limits 0.794-0.967
week  4: 0.864  limits 0.799-0.963
week  5: 0.836  limits 0.793-0.969
week  6: 0.889  limits 0.797-0.964
week  7: 0.899  limits 0.795-0.966
week  8: 0.754  limits 0.798-0.964  BELOW LCL
week  9: 0.895  limits 0.796-0.965
week 10: 0.913  limits 0.794-0.967
week 11: 0.915  limits 0.799-0.963
week 12: 0.926  limits 0.797-0.964
week 13: 0.931  limits 0.795-0.966
week 14: 0.935  limits 0.798-0.963

Week 8 fell below the lower control limit: that week, a new drywall crew started on level 5 and many of its first items failed, the situation lesson 5 describes. The crew's initial phase was repeated and the following weeks recovered. The last six weeks are all above the centre line, a run that is unlikely to be chance, which is a good signal: the process has improved, and the centre line and limits should be recalculated from recent data so that future problems are judged against the better standard. The chart did not tell anyone what to do, but it told them where to look.

Choosing a sampling plan

Lesson 3 showed that the chance of a sample finding at least one defect depends on the sample size and the defect rate. A sampling plan goes one step further: it sets the sample size and the number of defects allowed before the lot, such as a zone's firestops or a floor's windows, is rejected for full inspection. The operating characteristic curve shows how likely a plan is to accept lots of different quality. A good plan accepts good lots almost always and rejects bad lots almost always; the steeper the curve, the better it discriminates.

# Operating characteristic of two sampling plans for a zone of 300 firestops (illustrative)
from math import comb
N = 300
def p_accept(n, c, D):
    """Probability a random sample of n contains at most c defectives when the lot has D."""
    return sum(comb(D, k) * comb(N - D, n - k) for k in range(0, c + 1)) / comb(N, n)
plans = {"n=30, c=1": (30, 1), "n=60, c=2": (60, 2)}
print("defect rate   " + "   ".join(f"{name:>10s}" for name in plans))
for rate in (0.01, 0.02, 0.05, 0.08, 0.12):
    D = round(N * rate)
    row = "   ".join(f"{p_accept(n, c, D):10.2f}" for n, c in plans.values())
    print(f"   {rate:5.0%}      {row}")

defect rate    n=30, c=1    n=60, c=2
      1%            0.97         0.99
      2%            0.89         0.90
      5%            0.55         0.39
      8%            0.28         0.10
     12%            0.10         0.01

The larger plan, 60 items with up to two defects allowed, accepts a zone with 1 percent defects nearly every time and one with 2 percent about nine times in ten, but accepts a zone with 8 percent defects only about one time in ten, while the smaller plan would accept it about 28 percent of the time. For firestopping, where a bad zone means life-safety risk, Lena chose the larger plan, combined with the installer's full self-inspection and the building official's hold point. For paint touch-ups, a small plan is enough. Sampling plans are a way to spend inspection effort where it buys the most confidence.

Results on Larkspur Commons

By the time the last floor was handed over, the wood-frame and envelope phases could be compared with the podium phase from lesson 1. The comparison is not perfect, because the work was different, but the direction is clear.

# Podium phase versus wood-frame and envelope phases (illustrative)
phases = {
    "podium":                 {"cost": 14_800, "prev": 89, "appr": 183, "fail": 348, "ftq": 0.87, "ncr": 31},
    "wood frame and envelope": {"cost": 27_600, "prev": 214, "appr": 296, "fail": 402, "ftq": 0.91, "ncr": 38},
}
for name, d in phases.items():
    total = d["prev"] + d["appr"] + d["fail"]
    print(f"{name:24s} cost of quality {total / d['cost']:.1%} of work; failure {d['fail'] / d['cost']:.1%}; "
          f"prevention {d['prev'] / d['cost']:.1%}; first-time pass {d['ftq']:.0%}; NCRs per million dollars {d['ncr'] / (d['cost'] / 1000):.1f}")

podium                   cost of quality 4.2% of work; failure 2.4%; prevention 0.6%; first-time pass 87%; NCRs per million dollars 2.1
wood frame and envelope  cost of quality 3.3% of work; failure 1.5%; prevention 0.8%; first-time pass 91%; NCRs per million dollars 1.4

Prevention spending more than doubled in absolute terms and rose as a share of the work, and failure costs fell from about 2.4 percent of the work to about 1.5 percent. NCRs per million dollars fell and the first-time pass rate rose. The total cost of quality as a share of work fell too. These figures are illustrative and drawn from one invented project, but the pattern is the one quality programs aim for: spend more on getting it right, and less on fixing it.

Quality culture

Systems and measures matter, but in the end quality is made by people deciding, many times a day, whether to do the job right or just get it done. A quality culture is one in which doing it right is the normal choice, even when nobody is watching and the schedule is tight. It is built by leaders' behaviour more than by posters.

Quality and safety culture overlap almost completely. The same practices, planning the task, checking conditions before starting, stopping when something looks wrong, reporting problems without fear, learning from incidents and near misses, prevent both injuries and defects. Crews that are rushed and pressured cut corners on both. CON-355 covers safety leadership for site supervisors; most of what it says about leading safety applies to leading quality, and the two should be led together rather than as competing priorities.

Psychological safety, the belief that one can raise a concern or admit a mistake without being punished, is central to both. On Larkspur Commons, the moment that changed the drywall crew's attitude was not a rule but a conversation: when a young taper told Omar that she thought the crew had been skipping the second coat in closets to keep up, Omar thanked her in front of the crew, slowed the takt for that wagon by half a day, and the problem stopped. People keep reporting when reporting is rewarded.

Digital tools in quality

Digital tools have changed how quality is recorded and, increasingly, how it is checked. Thornbury uses an inspection application for checklists, photos and NCRs on tablets; reality capture (CON-315) to compare installed work with the model; and wireless sensors in the transfer slab that report concrete temperature, which the maturity method (CON-325) converts into estimated strength to support stressing decisions alongside field-cured cylinders. Drones photograph the roof and envelope for progress and for spotting obvious defects such as missing flashings.

Tools help, but they do not inspect. A photo of a firestop proves it existed; it does not prove it was the right system unless someone checked. The risk with digital tools is that records become easier to create than to read. Lena's rule is that every type of record must have someone who looks at it and a reason to look, which is why the dashboard draws directly from the inspection application. As of 2026, many contractors are also trying software that uses image recognition or artificial intelligence to flag possible defects in photos and scans. These tools are developing quickly and can help direct attention, but their findings still need a competent person's judgment before work is accepted or rejected.

Lessons learned

The final step in the plan-do-check-act cycle is to carry lessons to the next project. Thornbury holds a quality lessons-learned session at the end of each major phase, not only at the end of the project, while the people who know what happened are still on site. Each lesson is written as a specific change to a procedure, checklist, ITP or specification review item, so that it can actually be used.

Capstone: a quality plan for the next project

Thornbury's next project is Cedar Hollow, an invented five-storey seniors' residence of 96 suites with a single-storey concrete podium for parking and amenity space and four storeys of wood frame above, with a larger kitchen and dining area and more complex mechanical systems than Larkspur Commons. It is in the same region, with the same owner's consultants. Lena has asked Mei to draft the project quality plan, using what the team learned. The outline below is Mei's draft. Readers can use it as a model and adapt it to a project of their own.

The plan is shorter than Larkspur Commons's was at the same stage, because much of the detail now lives in Thornbury's updated procedures and checklists, which the plan references rather than repeats. The final check Mei runs on the draft is whether a new foreman, reading the parts that apply to his trade, could answer Lena's three questions from lesson 2: which inspections apply to my work this week, who signs them off, and what happens if something fails. If he can, the plan is ready to use.

# Quick check of the draft plan's objectives against Larkspur Commons results (illustrative)
targets = {"first-time pass rate": (0.92, ">="), "consultant deficiencies per suite": (4.0, "<="),
           "failure cost share": (0.015, "<="), "self-reported NCR share": (0.60, ">=")}
larkspur = {"first-time pass rate": 0.91, "consultant deficiencies per suite": 4.6,
            "failure cost share": 0.0146, "self-reported NCR share": 0.61}
for k, (t, op) in targets.items():
    v = larkspur[k]
    met = v >= t if op == ">=" else v <= t
    print(f"{k:34s} target {op} {t:<6} Larkspur result {v:<7} {'already met' if met else 'needs improvement'}")

first-time pass rate               target >= 0.92   Larkspur result 0.91    needs improvement
consultant deficiencies per suite  target <= 4.0    Larkspur result 4.6     needs improvement
failure cost share                 target <= 0.015  Larkspur result 0.0146  already met
self-reported NCR share            target >= 0.6    Larkspur result 0.61    already met

Two targets are already being met, at the level Larkspur Commons finished at, and two will need improvement. That is deliberate. A plan whose targets are all already met does not ask the team to get better, and one whose targets are far beyond what the team has shown it can do will be ignored. The targets for Cedar Hollow ask for a modest step up on first-time quality and on finished-suite deficiencies, the two areas where the lessons learned table suggests the biggest gains remain.