Scoring Methodology — Capture Desk DMV

Overview

Opportunity Scout scores each uploaded solicitation and pre-solicitation on a 0–93% Win Likelihood scale. This is not a statistical win probability — it's a relative priority rank designed to surface the best-fit, lowest-friction opportunities for a small business at an early stage.

Scoring runs entirely server-side in the Cloudflare Worker. The formula is not visible in page source. All scores are computed fresh on each scan and can be refined further by Claude AI analysis of the actual solicitation document.

Sources & taxonomy

Every opportunity carries a source tag: state-md (Maryland eMMA) or state-va (Virginia eVA). Records are UNSPSC-coded; the code is stored in a classificationCode field. Opportunities are loaded via upload (JSON or a fixed-column CSV template) and persist in the Worker's KV store; live portal adapters are stubbed pending API access.

Because records carry a UNSPSC code rather than a matchable profile-code table, scope alignment is judged from the title keyword patterns (see Step 3), with a flat classification credit standing in for a profile-code hit.

Win Likelihood = Base(25) + Eligibility + Value + Set-aside + Remote + Keyword − No-match penalty + Classification credit + Complexity penalty − Red flag penalties → clamped 8%–93%

Step 1 — Exclusion Filter

Before scoring, any opportunity whose title matches an exclusion pattern is flagged "off-profile" and given a score of 5% (excluded from the ranked view by default). Exclusions are hard filters — they indicate scope that is structurally outside your capabilities.

Category	Example keywords
Aviation / aerospace	aircraft, aviation, airframe, rotor, helicopter
Weapons / ordnance	weapon, ordnance, ammunition, missile, artillery
CBRN / hazardous materials	nuclear, radiological, hazmat, biological agent
Heavy civil construction	dam, bridge, freeway, dredge, excavate
Medical devices	implant, prosthetic, surgical instrument, diagnostic kit
Food service / agriculture	cafeteria, catering, livestock, crop
Electronics manufacturing	circuit board, PCB, semiconductor, cable assembly
Marine vessels	vessel, ship, submarine, destroyer

Step 2 — Base Score & Adjustments

Base

Every non-excluded opportunity starts at 25% — roughly the average competitive win rate for a small business bidding open solicitations.

Eligibility

Condition	Adjustment
Eligible now (SB / SBR / SWaM / SBE / open competition / unrestricted)	+25%
Future eligible (WOSB, SDVOSB, 8(a), HUBZone — cert gap)	+5%
Other / check required	+0%

Contract value

Contract value	Adjustment	Rationale
Unknown / $0	+5%	Unpriced — assume small
≤ $25K	+20%	Micro-purchase range — easiest to win
$25K–$100K	+16%	Small-purchase threshold
$100K–$250K	+10%	Small enough to execute easily
$250K–$500K	+4%	Moderate size
$500K–$1M	−4%	Increasing competition and scrutiny
> $1M	−14%	Large contracts attract established contractors

Complexity

Complexity (Low −0% / Medium −5% / High −14%) is decoupled from raw dollar value. Large contract size alone does not imply execution difficulty for a low-friction service.

Category	Complexity rule
Janitorial, Grounds/Lawn, Software Resale, Admin Support	Low up to $1M; Medium above $1M — never High from size alone
Real Estate, Consulting, Mental Health, unclassified	Standard value tiers: Low <$150K, Medium $150K–$1M, High >$1M
Any category with a Systems Engineering, Enterprise Scope, or COOP red flag	+1 complexity level (capped at High)

Complexity also drives sort order — the ranked list is ordered by WinLikelihood² ÷ Complexity², so a High-complexity opportunity ranks well below a Low-complexity one at the same win likelihood.

Set-aside & location

Condition	Adjustment
SB set-aside or unrestricted	+10%
Statewide / remote performance	+4%

Maryland SBR (Small Business Reserve), Virginia SWaM (Small, Women-owned & Minority-owned), and generic municipal SBE (Small Business Enterprise) set-asides are all treated as small-business-eligible now.

Step 3 — Keyword Boost & No-Match Penalty

The opportunity title is scanned for keywords matching your service profile. The first matching pattern applies its boost and labels the opportunity (e.g. "Janitorial", "Mental Health").

Category label	Trigger keywords (sample)	Boost
Mental Health	mental health, counseling, therapy, behavioral health, clinical, psycholog	+20%
Janitorial	janitorial, custodial, cleaning, housekeeping, sanitation	+18%
Grounds/Lawn	lawn, landscape, grounds, mowing, turf, vegetation	+18%
Consulting	consulting, advisory, management support, program support, acquisition support	+15%
Software Resale	software, license, subscription, SaaS, technology resell	+14%
Real Estate	real estate, property, space search, facility finder, lease	+14%
Admin Support	administrative, admin support, office support, clerical, records management	+12%

No-match penalty: −12%
If no keyword pattern matches the title, the opportunity scores a −12% penalty. Vague titles (e.g. "Miscellaneous Support Requirement") are likely misclassified or out-of-profile. Unknown scope defaults to skepticism.

Classification credit

Records carry a UNSPSC code, not a matchable profile-code table. Instead, an opportunity whose title matches one of the keyword patterns above receives a flat +8% classification credit plus +15 signal points. This stands in for a direct profile-code hit; it applies once, on a keyword match.

Step 4 — Red Flag Deductions

Title patterns that signal structural disadvantage are scanned. Total deduction is capped at −30% regardless of how many flags trigger. Up to 2 flag labels are shown in the opportunity row.

Flag	Trigger keywords (sample)	Deduction
O&M	operations and maintenance, O&M	−15%
Incumbent Signal	incumbent, recompete, follow-on contract, option exercise	−12%
IDIQ/Vehicle	IDIQ, MATOC, BPA, blanket purchase agreement, indefinite-delivery	−12%
Sys Engineering	systems engineering, systems integration, systems architect	−12%
COOP	continuity of operations, COOP	−10%
Enterprise Scope	enterprise-wide, enterprise solution, portfolio management	−10%
Health IT	medical records, EHR, electronic health record, clinical data	−8%
Logistics	logistics support, supply chain, inventory management, warehousing	−8%

Score Signal

Each score is accompanied by a Signal indicator (High / Med / Low) that tells you how much data supports the score — not whether the score is good or bad.

● High signal Strong keyword match (category boost triggered) plus classification credit. Score reflects real information about scope alignment.

● Med signal Partial signal. A keyword match is present but weak, or the title is otherwise ambiguous.

● Low signal No keyword match. Score is a conservative structural estimate — treat it as a flag for closer review, not a reliable likelihood.

Claude AI Analysis

On demand (click "Analyze ✦"), Claude reads the solicitation documents attached to the uploaded record and returns a structured assessment including:

Overall fit: strong / moderate / weak
Staffing estimate and key roles
Required certifications, licenses, or background checks
Period of performance and contract type
Evaluation / award basis and past performance requirements
Incumbent risk, win factors, and red flags

AI analyses are cached for 90 days. After a scan, any previously analyzed opportunities have their results automatically restored so you don't re-pay for Claude on the same solicitation. All analyzed opportunities are stored in the Analysis Archive.

AI vs. algorithmic score: The algorithmic score is fast and free — it's designed to filter out noise and surface candidates worth closer inspection. The AI analysis is the deeper read. The goal is to close the gap between the two by continuously improving the algorithm based on patterns seen in AI outputs.

Known Gaps & Limitations

What the score does not capture:

Market competition: There is no comprehensive public historical-award dataset for state and local procurement, so competition density is not scored. A future market-data source could feed the competitorCount input the scorer already supports.
Amendment history: Heavily amended solicitations may indicate incumbent negotiation or scope changes — not yet scored.
Solicitation type: IFB vs. RFP vs. RFQ vs. Sources Sought indicate different procurement vehicles and effort levels — only partially extracted (from the type field).
Geographic mismatch: Specific non-remote locations that require local presence are not penalized unless the title flags it.
Contract type (FFP vs. T&M): Fixed-price contracts are generally more favorable for small businesses. Contract type is only visible in the solicitation document, not the record metadata — captured in AI analysis, not algorithmic scoring.
Bonding & insurance thresholds: Bid/performance bond and insurance requirements can be a hard barrier. Captured in AI analysis, not scoring.
Past performance requirements: Many solicitations require several years of directly relevant references. Captured in AI analysis but not in algorithmic scoring.