Agent-assisted audit ← Back to audits list

LDD Control Audit — Codex (2026-05-17 · 22:40 ET)

Codex
Codex
on behalf of David
Model: Codex / GPT-5.5
Type: Agent-assisted LDD control audit
Date · Time: 2026-05-17 · 22:40 ET
LDD version: v1.0
6.4/10
One-line takeaway: The design intent is strong, but the project control layer has drifted enough that exports, PM pages, and visual evidence can no longer be treated as one source of truth.
Pro reaction

LDD-29 and LDD-30 protect the gym and mechanical core from being value-engineered into unrelated parts.

Con reaction

The project currently has multiple competing truths, and a contractor will build from whichever one is easiest to price.

Fresh audit after reading the current repo, dispatching three read-only agents, and checking current external code / permitting anchors. This is an audit and project-control review, not licensed architectural, engineering, code, legal, or privacy advice. I did not rewrite canonical LDD content.

Score

I score the current package 6.4 / 10. This is a strong design-intent system, but the control surface has drifted. The best ideas are now spread across canonical Markdown, hand-built HTML mirrors, stale project-management pages, missing exports, and visual assets that no longer match the text.

What Is Strong

Top Holes

1. P1 — A Deployed Floor-Plan Asset Carried GPS Metadata

The asset pass found site/assets/library/floor-plans/estimaed-footprint.jpeg was byte-identical to an incoming/original-sketches/ source file and retained Apple EXIF metadata, including GPS fields. I am not repeating the coordinates here. That was a privacy / source-asset handling issue, not just a web optimization nit.

Action

Replace the deployed library copy with a metadata-stripped derivative, preferably WebP, while leaving the incoming/ original untouched. Mitigated for this deploy: the site page now points at a stripped WebP derivative.

2. P1 — The Current Truth Is Split Across Too Many Surfaces

The sidebar lists 30 LDDs, while other pages still say 24 or 28. The active PM pages disagree on MUA/ERV, AHJ gating, latest audit status, and which LDDs are still missing.

Action

Create one Current Package State source that drives LDD count, latest audit, controller lock, known stale pages, and current decision gates.

3. P1 — The Foundation / Slab Basis Is Still Contradictory

LDD-01 still says "pier foundation (not slab-on-grade)" and says LDD-02 must be reconciled before bidding. LDD-02 says the pier-foundation cascade was withdrawn and the first-floor slab remains monolithic slab-on-grade.

Action

Do not treat LDD-01 and LDD-02 as green together until the foundation/slab basis is stated once, in one controlled note.

4. P1 — The 70 SF Mechanical Room Is Still A Build Blocker

LDD-30 moves batteries and main electrical out, but Zone 1 still wants heat-pump equipment, tankless booster, indirect tank, manifold, utility sink, floor drain, and MUA equipment inside a 10' x 7' room with a continuous 48" aisle.

Action

Require a scaled equipment layout with real model dimensions, access clearances, filter-change paths, valve access, door swing, drain location, and MUA duct route before any slab penetrations or rough-in bids.

5. P1 — MUA/ERV And AHJ Gate Decisions Are Not Controlled Enough

LDD-28 says dedicated MUA supersedes boost-ERV, but its own text still calls the LDD-05/07/28 conflict a v1.0 blocker. AHJ Type I classification is closed in the risk register but still appears as a red steel-order gate elsewhere.

Action

Create an AHJ + MEP Decision Log with exactly two states for each decision: current controlling assumption and next proof required.

6. P1 — Exports And Handoff Packages Omit Current Master Documents

The DOCX builder still says it builds 28 LDDs and hard-codes only 01-28. LDD-29 and LDD-30 are current master documents, so any exported "complete" package currently drops the two documents most likely to prevent bad bids.

Action

Update the export pipeline so it discovers LDDs from a manifest or the sidebar list, not a stale hard-coded array.

7. P2 — Civil/Site Confidence Is Too Verbal For A Stream-Adjacent Build

"Above floodplain" is useful, but it is not a stormwater, grading, access, utility, sediment-control, or site-cost answer.

Action

Create a Site + Civil sheet and decision log covering FFE, floodplain/floodway line, grading, runoff, erosion controls, driveway/apron, sewer, utilities, fire access, disturbed acreage, and NPDES applicability.

8. P2 — Fire/Life Safety And Occupancy Need One Controlled Sheet

The project blends dwelling, gym, workshop, garage, batteries, material movement, inventory, and possible business support. Fire separation, sprinkler determination, rescue openings, stairs, smoke/CO, emergency lighting, and battery placement are scattered.

Action

Create a Life Safety + Access Plan before bids: occupancy/use assumptions, separations, rated doors, devices, exits, rescue openings, thresholds, battery zone, garage/workshop separation, and sprinkler-ready routing.

9. P2 — CAD, Diagram, And FPO Provenance Is Ambiguous

The CAD manifest is empty even though many export sets exist. Some pages still call diagrams placeholders where CAD sources and exports exist. LDD-30 has references to missing plan/render assets.

Action

Build a diagram/export manifest with owning LDD, source script, export paths, site copy path, freshness status, prompt path, and known stale reason.

10. P3 — Gym Protection And Athletic Coordination Still Need Hard Specs

LDD-22 has a site mirror load-path contradiction. LDD-25 and LDD-29 are not fully aligned on mirror-wall deployable nets. LDD-29's current plan/render notes admit the visual assets lag the latest basketball + beach-volleyball floor strategy.

Action

Draw one gym coordination sheet: hoop spine load path, rim retraction, floor striping, volleyball sleeves, mirror-wall protection, overlook nets, fans, lights, clouds, clerestories, sprinklers, mini-splits, and 16' datum alignment.

Score Movement

6.4 → 7.1 Keep the GPS-tagged deployed asset stripped from future deploys, fix DOCX export coverage, add LDD-30 assets or warnings, sync audit/LDD counts, and fix the MUA/ERV + AHJ stale PM pages.
7.1 → 7.8 Land the AHJ + MEP Decision Log, scaled mech-room layout, site/civil sheet, life-safety/access sheet, and diagram/CAD manifest.
7.8 → 8.5+ Get stamped structural/MEP/civil packages, line-item GC bids, written AHJ/code assumptions, energy/ventilation calculations, and commissioning acceptance criteria.

Research Anchors

Bottom Line

The project does not need another taste pass. It needs a truth-control pass. Once that control layer is fixed, the design intent becomes much more powerful because everyone pricing, reviewing, or building from it is looking at the same project.