The rate library
A master metric system with anchored sub-tracks. Different building types get different rate libraries, but every track shares the same architecture.
The core insight from running a real bidding practice: premium custom residential rates do not apply to a hotel, and a spec-home anchor cannot price multifamily. Without segmentation, rates muddle and bids drift wrong. So the library is split into six tracks, each carrying its own anchor projects, its own intake protocol, its own output templates, and its own audit rules.
| Track | Anchor status | Working note |
|---|---|---|
| Residential custom, premium tier | Anchor live | Anchored to a premium spec home the practice bid, built, and closed out with firm actual cost. The calibration bible; every premium residential bid is checked against it. |
| Residential ADU tier | Pending | Different finish tier than custom premium; needs its own anchor before the track goes live. |
| Commercial general | Partial | Anchored for interior partition scope only (framing, drywall, ceilings), backed into labor rates from material counts. Not yet a full-project anchor. |
| Multifamily / hospitality | Provisional | Anchored to a full extended-stay hotel estimate that was bid but not built. Rates carry an explicit aggressive-posture confidence tag and get demoted the day a built, firm-cost comp arrives. |
| Panelized structural (SABS) | Live, cross-track | A trade-specific track for the panelized wall system the practice installs. Applied inside residential and multifamily tracks when SABS is the structural system; priced at parity against stick framing. |
| Renovation / remodel | Pending | No anchor yet. Renovations carry enough existing-conditions risk that we will not run the track on borrowed rates. |
What an anchor is, and why we protect it
An anchor is a project that was bid, built, and closed out with a firm actual cost, so the library can compare what we said it would cost against what it actually cost. Nine real projects across five building types feed the library, but only anchors set a track's baseline; the rest are historical reference.
Anchors stay standalone. Historical bids, won or lost, never get averaged into an anchor, because mixing bid attempts with built outcomes produces noise dressed up as data. We can always say exactly what the anchor project cost and exactly how far any new bid sits from it. When a track has only one anchor, we say so out loud: that is a named risk on the Roadmap, not a footnote.
On the one track with a full anchor, the engine's autonomous end-to-end run priced the anchor project within 7.1% under actual built cost, inside our margin of error. That number is the whole reason this data room exists. It is also a single data point, and we treat it as one.
Variance bands and rate hygiene
Every line-item rate in a track lives inside a ±10% variance band around the track's library median. A new bid whose rate falls outside the band gets flagged out-of-bounds on the internal audit copy, color-coded, and has to be either justified or corrected before the bid ships. With a single anchor the median is the anchor; as comps accumulate, the median is computed across the track.
Beyond the band, every rate carries standing metadata:
- Basis tag. Whether the rate is tied to square footage, duration, count, or is project-specific. A rate is never applied against the wrong measurement basis silently.
- Source tag. Which historical project the rate came from, so any number in any bid can be traced to its origin.
- Confidence tag. One of QUOTED (a sub or vendor wrote this number for this plan set), TAKEOFF (measured from the drawings), ASSUMED (market average or ratio-derived), or ALLOWANCE (a placeholder the client should expect to move). The tag prints on the bid, which makes the document auditable line by line.
- Tier variant. Premium and ultra-premium tiers resolve their own rate keys before falling back to base rates.
When a rate is needed, the engine resolves it through a fixed cascade: a quote on this specific plan set wins; then live market pricing (daily materials feeds and federal producer-price indices, with a tariff watch); then the same line item quoted on a past project; and only then a stated assumption. Hard gates sit underneath: missing plan-cover data or unverified quantities mean the system refuses to produce a headline number at all.
The autonomous takeoff pipeline
Drop in a plan PDF; get back an assembled, confidence-tagged bid. Two to four minutes of wall time, roughly a dollar of model spend.
The accuracy stack
Takeoff is the foundation: if the counts are wrong, every downstream number is wrong. The engine treats accuracy as six stacked layers rather than one clever trick, and we built them in order of cost-effectiveness.
| Layer | What it does | Cost |
|---|---|---|
| 1 · Clean input | Proper render resolution, OCR fallback, text-layer extraction, sheet classification. The unglamorous layer that does the most work. | Free |
| 2 · Cross-source verification | The door schedule and the floor-plan vision read must agree before a count is accepted. | Free |
| 3 · Self-consistency | The same model reads the same sheet multiple times; the median wins, outliers get flagged. | Cents |
| 4 · Multi-model consensus | Two independent vision models must agree on contested quantities. | Dollars |
| 5 · Domain invariants | Construction sanity rules (square-footage-to-bedroom ratios, door-count boundaries) catch what statistics miss. | Free |
| 6 · Human in the loop | An operator drills into any line: confidence, rate key, and the formula in plain English. Overrides require a reason and persist. | Minutes |
The working target is 95% autonomous, 5% human review. Chasing 100% autonomy on construction documents is a fool's errand, and the product is honest about that: model disagreement shows up on the operator's screen as a flag, not a silently averaged guess.
What comes out the other end
Clean and binding-aware
CSI-organized line items, allowance flags, contract-type verbiage matched to fixed-price, cost-plus, or time-and-materials. No internal calibration data leaks into the client copy.
Rate-by-rate accountability
Every line compared against the track anchor, out-of-band rates color-flagged, provenance and confidence printed per line. This is the copy that keeps the library honest.