MODEL-DRIVEN CONSTRUCTION DELIVERY — SYSTEM DESIGN SUMMARY DOC MDCD‑001 · REV A · DRAFT FOR REVIEW · 2026‑08‑20

Model-Driven Construction Delivery

Work breakdown, cost, schedule and payment are computed outputs of one BIM model rather than four separately maintained records. This summary states the operating principle, the design decisions that follow from it, current build status, and the constraints the specification places on input data.

ScopeCommercial construction delivery, single trade pilot
Method basisLast Planner · LBMS/Takt · 5D BIM
Build statusDomain & planning engine proven; field loop not built
Source documentsREADME · ARCHITECTURE · MODEL‑SPEC · ROADMAP
01

Operating principle

Every model element implies a defined scope of work. Decomposing the model by trade and assembly yields a work breakdown that is generated, not authored: quantities are extracted from the model, durations are computed by applying productivity rates to those quantities, and cost is computed by applying rate data to the same quantities.

Cost, schedule and quantity cease to be three independently maintained records and become three projections of a single dataset. Payment is bound to verified completion of the generated work items. This dependency is the enforcement mechanism: keeping the model current stops being a discipline problem and becomes a precondition for payment.

Prior 5D implementations — Vico and its successors — established the same model → quantity → cost → schedule chain and did not persist, because no mechanism compelled the model to stay current. Binding payment to model state supplies that mechanism.

Operationally: a crew declares its composition, the system computes a constraint‑checked package sized to the resulting capacity, the crew accepts the commitment, and completion of that commitment drives payment. Partial completion is reported as a count against a fixed denominator — 73 of 100 items — rather than an estimated percentage.

02

Method basis

None of the individual methods below is original to this system; the composition is. The contribution is coupling a payment‑verification loop to the model so the underlying data maintains itself — a mechanism that requires contractual authority to mandate, not additional software.

M.1

Last Planner System Lean Construction Institute

Crews commit to what they will complete in a defined period; reliability is measured as Percent Plan Complete (PPC). The commitment is pulled by the crew, not pushed by a planner — pushed assignments show measurably lower completion reliability.

M.2

Location‑Based Management / Takt Kenley & Seppänen

Work is decomposed by location and trade, with crews flowing through defined zones in a fixed sequence, rather than the full site being worked simultaneously.

M.3

5D BIM Vico → Trimble

Established the model → quantity → cost → schedule chain. The data model was correct; the adoption model was not — no mechanism kept the model current past the issued‑for‑construction milestone.

03

Work‑to‑payment loop

One directional sequence, executed per work package, once per commitment cycle:

01ElementClassified, quantified, located
02DeriveAssembly template applies a sequence
03CheckConstraints cleared or attributed
04SizePackage computed to crew capacity
05CommitCrew accepts, trims, or swaps
06CaptureBinary, per item, attributed
07SettleClaim limited to verified items

Granularity is resolved at selection time, not fixed in the schema. Leaf nodes are atomic and binary; every node above a leaf is a computed rollup. A crew may select any subtree, or any arbitrary set of leaves, as its package.

04

Design decisions

Nine decisions govern the system's behaviour under load. Each is stated with the failure mode it forecloses.

D.1

Granularity as a runtime parameter

One work item per model element produces hundreds of thousands of checkboxes on a mid‑size building — unusable in practice. Coarse items force percent‑complete reporting, which is subjective, defeating the purpose of payment‑on‑verification. Binary leaves with free selection yield both tractable scale and objective partial completion.

D.2

Sequence bound to assembly type, not project

A partition wall sequences identically across projects: frame, rough‑in, inspect, insulate, board, tape, prime. That knowledge is held in a template library keyed by assembly type and applied at ingest. Site logistics, procurement and inspection gates remain a human overlay; corrections to the template feed the library. Cost on the second project is a fraction of the first.

D.3

Readiness computed, not asserted

Constraint checking at selection time — predecessor, area release, inspection, material — converts the tool from a form into a planning system, and makes delay attribution auditable. A commitment missed behind an open upstream constraint is logged as a system failure, not a trade failure; this distinction must survive into performance data or the platform penalises trades for management's failures.

D.4

Proposal, not assignment

Auto‑assignment is cheaper to implement and breaks the property that makes completion data meaningful: pushed work measures estimate accuracy, pulled work measures crew reliability. These are distinct metrics and the system requires both.

D.5

Reliability and productivity, unblended

Reliability — completion against commitment (PPC). Productivity — hours per unit, measured against a crew's own trailing average, never against other crews, since scope and conditions differ enough to make cross‑crew comparison mostly noise. Combining the two into a single score destroys the information content of both and produces a metric that is trivially gamed.

D.6

Performance data for calibration, not evaluation

Framed as calibration input, subcontractors report accurately and help correct the rate model. Framed as a scorecard tied to award decisions, the same data collection produces sandbagged estimates, cherry‑picked scope, and adversarial reporting. Identical data pipeline, different framing, materially different data quality.

D.7

Rates as priors, converging to observation

Published rate data (RSMeans and equivalents) functions as a cold‑start prior — national‑average, not specific to a given crew in a given market. Every completed package with quantities and timestamps is a productivity observation; after several projects, the empirical rate outperforms any published dataset for that scope.

D.8

Rates as a function of dispersion, not quantity alone

Three doors on Level 2 and two on Level 5 do not sum to five doors' worth of labour — setup, travel and staging do not distribute linearly across scattered locations. Rates are computed as a function of quantity and spatial dispersion; naive summation understates scattered selections and erodes trust in the estimate after the first miss.

D.9

Automatic computation, approval‑gated application

Recomputing quantities, work items and schedule after a model change is deterministic and safe, and happens automatically. Determining whether previously completed work is invalidated, or applying a resulting cost delta, is a judgement call with commercial consequence, and requires an explicit approval.

05

Excluded from automation

Automation is scoped to clerical computation. Where a decision carries professional liability or unresolved ambiguity, the system computes the available options and stops short of deciding.

DecisionBasis for exclusion
Whether completed work is invalidated by a model changeJudgement with high consequence — a 200mm wall shift may or may not void framing already installed
Application of cost changesA commercial commitment requiring a signature, not a recalculation
Sealing of inferred model dataProfessional liability — a seal is backed by an insurer, a computed value is not
Final construction sequenceDepends on tacit knowledge that is not represented in the model
Quality acceptanceThroughput metrics do not detect defects; speed scores well while rework surfaces downstream

Every inferred parameter carries provenance: who or what asserted the value, and who accepted it. The system proposes; a qualified professional accepts, in bulk where review time permits. No path exists for an inferred value to reach contract status unattended.

06

Build status

ComponentState
Work breakdown — nested, runtime granularityBUILT
Assembly sequence templates — 2 starter templatesBUILT
Constraint model & readiness, with attributionBUILT
Crew capacity from declared compositionBUILT
Package proposal, sized to capacityBUILT
Productivity rates — priors, observations, calibrationBUILT
Procore integration — REST client and toolsUNWIRED
IFC ingestNOT STARTED
CPM schedulingNOT STARTED
Field capture and verificationNOT STARTED

Reference run — one crew, one day, three zones

The domain layer is exercised end to end by examples/drywall_day.py, which plans a single crew against a synthetic three‑zone level:

$ PYTHONPATH=src python3 examples/drywall_day.py
drywall: 1 foreman, 3 journeyman, 1 apprentice (5 total)
Effective capacity: 28.22h (vs 40.0h raw)

1 item(s), 13.22h of 28.22h available (47%)
10 item(s) blocked: {'other_trade': 8, 'supplier': 1, 'management': 7}

Only 47% of capacity filled. Either the crew is oversized for available work,
or upstream constraints are starving this trade.

The final line is the diagnostic output of interest, independent of the package itself: a crew idle at 47% because an area was not released is recorded as a management failure, not a poor day for the trade.

07

Model input requirements

Because work, cost and schedule derive from the model, a requirement that cannot be checked automatically will not be enforced, and a parameter that is not populated cannot be derived. The model specification exhibit states this constraint before a consultant opens an authoring tool. Each requirement is marked MC — machine‑checkable, validated on receipt — or RV — requiring review. A model failing any MC requirement is returned unreviewed.

Requirement categories

AreaGoverning ruleCheck
Spatial structureEvery element resolves to exactly one storey — an unplaced element cannot be scheduledMC
ClassificationUniFormat II reference at type level — the join between element, spec section and cost codeMC
GeometryLOD 300 minimum for scheduled categories; an omitted element reads as scope that does not existRV
QuantitiesNative Qto_ base quantity sets, taken in preference to geometry recomputationMC
Modelling hygieneNo in‑place families; no duplicate or overlapping geometry; no groups substituting for typesMC
NamingStable, unique type names — two names differing only cosmetically are priced twiceMC
Change controlIssued‑for‑construction model frozen as baseline and hashed; every revision recorded as a deltaRV

The specification is self‑checking: run the validator against a model the intended consultant has already delivered on a prior project. If an existing model cannot pass, the specification is not yet writable and the requirement moves, not the consultant.

08

Prototype sequence

One trade, one building, end to end. Proven the day a crew opens the system on a Monday, plans a week against constraint‑checked work, and the resulting payment claim matches exactly what is verified complete.

Target demonstration state: three foundation walls marked complete; those three render green in the model; the claim lists those three and nothing else. Visual state and commercial state in a single view.
Ph.FocusExit gatePrincipal risk
0Specify input requirements; select pilot tradeSpecification written; conforming test model existsLow technical risk, high consequence if wrong
1Model to checklist — IFC ingest, quantity extractionDerived quantities reconcile with a manual takeoffHighest risk in the sequence
2Work plan against ingested project dataSuperintendent confirms the generated sequence is correctGated on trade‑contact availability
3Field loop — crew view, capture, model viewerA crew plans a week without supervisionAdoption risk, not technical risk
4Commercial close — claims, Procore write‑backGenerated claim matches manual CM approvalLow — client already built
5Shadow pilot — real crews, conventional process in parallelThree consecutive pay periods reconciledOrganisational, not technical

The Phase 1 validation gate is where the premise is tested. A discrepancy between derived and manual quantities is almost always traceable to the model specification, not the ingest code — establishing that in Phase 1 is inexpensive; establishing it in Phase 5 is not.

09

Extension path

Direction, not commitment. Nothing in this section is scheduled or required for the platform to deliver value; it is recorded so near‑term decisions remain consistent with the longer trajectory.

The stated end state is a project executed from proposal to completion without a human in the execution loop. Each stage is an adapter around a core that does not change: what work exists, and what is owed for it.

Stage
Current implementation
Extension
Interface
Survey
Consultant reports
Automated reality capture
Model ingest
Design
Human‑authored model
AI‑generated, code‑checked
ModelElement
Forecast
Rates and templates
Unchanged
— core —
Execute
Crews declaring capacity
Fleets declaring capability
CrewComposition
Verify
Human sign‑off
Drone, then autonomous
Completion capture
Settle
Payment application
Automated release
Payment output

Constraints that persist under extension

  • Acceptance — detection can be automated; acceptance is a liability structure requiring a party, not a process. Required human effort approaches zero; the signature does not.
  • Statutory inspection — a building department inspects under its own statutory authority. This changes with legislation, not with technology.
  • Disputed matters — automating undisputed completion is the addressable win. Scope interpretation, backcharges and delay claims remain human adjudication.

On settlement: automated payment against verified completion is achievable today on existing payment rails; smart‑contract settlement is one possible implementation, not a prerequisite. It has a defensible use case on multi‑party projects with a genuine trust deficit — not where a developer is paying its own trades through an account a bank already administers. Statutory holdback and prompt‑payment rules apply regardless of settlement mechanism.