Bid Tracker · Admin only

Calibration

Per-field correction patterns over the rolling window · Admin + C-Suite only
Last refresh: 4m ago
Invites in window
47
PM corrections
128
avg 2.7 per invite
Review Inbox rate
17%
8 of 47
Fields > 30% correction
3
bid_due, pre_bid, prevailing_wage

Per-field correction rate

% of (invite × field) instances where the PM changed AI's value · sorted by rate descending · >30% flagged red
Field
Correction Rate
Rate
AI Conf
pre_bid_meeting
41%
0.68
bid_due_date
38%
0.74
prevailing_wage
32%
0.74
rfi_deadline
28%
0.72
square_footage
22%
0.78
delivery_method
18%
0.81
project_name
14%
0.85
owner
11%
0.83
bond_required
8%
0.88
gc_canonical
6%
0.91
city
4%
0.94
state
2%
0.98

Overconfidence detector

AI mean confidence (x) vs actual correction rate (y) · ideally points cluster along the diagonal · top-right quadrant = high confidence + high correction = AI is overconfident

AI mean confidence →
↑ Correction rate

Red dots are overconfidence outliers — AI says "I'm fairly sure" but PM corrects them >30% of the time. Tune the specialist prompt or add tools that fetch these explicitly.

Validation failure reason histogram

Why invites land in Review Inbox
incomplete
11
opportunity_ambiguous
7
low_confidence_required_field
5
gc_not_found
4
manual
3
agent_loop_exhausted
2

Per-GC patterns

Which GCs land in Review Inbox most
GC Invites Inbox % Top failure
Miller-Tippens 3 33% opportunity_ambiguous
Walters Morgan 4 25% incomplete
Crossland Construction 9 22% opportunity_ambiguous
Nabholz 7 14% incomplete
Flintco 8 12% low_confidence_required_field
Manhattan Construction 12 8% incomplete
Bravo Construction 4 0% —

Source: bid_invite_field_corrections + bid_invite_classification_traces joined on invite_id. Refresh runs every 5 min on a background task.