What the data actually says

“Full and open” competition usually draws one bidder

Across every agency in scope, more than half of awards labeled FULL AND OPEN COMPETITION received exactly one offer — $39.0B where the competitive label describes the procedure, not the outcome.

• ›SSA is the worst offender — only visible once SSA was added to scope.
• ›The finding gets stronger after cleaning: unfiltered it reads ~6 pts lower, because the artifact rows are all high-offer.
• ›Related: unrestricted (no set-aside) work draws fewer bidders than small-business set-asides. The direction survives cleaning; the magnitude does not.
• ›Needs no scoring model — it's a straight count.

Offer counts are contaminated by vehicles and sentinels

number_of_offers_received will embarrass you if quoted raw. Parent IDIQ vehicles report offers on the vehicle, not the order, and at least one award type uses 999 as a placeholder. Real order-level competition is 2–3 offers.

number_of_offers_received BETWEEN 1 AND 50
AND NULLIF(contract_award_type, '')
    IS NOT NULL

The original scores didn't discriminate, so they were rebuilt

In the original points model, 54.1% of the 3,658 recompete candidates tied at incumbent_strength = 65, the score's structural maximum. With ties broken by dollars, the radar was close to “sorted by contract size.” Since 2026-09-16 both scores are backtested probabilities instead.

• ›The lifetime term is least(30, ln(lifetime) × 2), which caps at e¹⁵ ≈ $3.27M. A $3.3M vendor and a $9.2B vendor score identically.
• ›Correlation between strength and offers received is effectively zero, and single-offer rate falls as strength rises — backwards for a defensibility score.
• ›recompete_score has the same problem: it cannot exceed 60, and the least(100, …) wrappers never bind.
• ›Saturation worsened after the scope expansion (52.5% → 54.1%) because vendor totals aggregate across NAICS.

• ›incumbent_strength is now P(incumbent wins the follow-on). recompete_score is P(a follow-on happens) × (1 − that), which is the chance a new vendor wins the work.
• ›Both come from logistic scorecards fitted on ended contracts with inferred follow-ons and point-in-time features. The models were trained on contracts that ended before 2021 and tested on later ones.
• ›Biggest retention factors: 0–1 rival vendors at the office (×10 odds), sole-source (×4.2), 4+ prior awards there (×3.4). Incumbent size is not a factor.
• ›The same model still scores 0.76 on a placebo window, so the matching method explains part of the signal. Treat scores as a ranking, not exact odds.

Mid-size specialists retain better than giant primes

The score treats bigger incumbents as stronger. The backtest says otherwise: vendors with $18.4–82.1M in-scope lifetime obligations kept their work most often, and the largest primes kept it least.

• ›Replicated: same shape — peak at Q2, trough at Q5 — before and after a 54% data increase. That survival is the strongest evidence here.
• ›Successors are inferred (same office + PSC, closest in dollars). The 2.7× lift over placebo shows the match captures something real, but treat the level as soft.
• ›Look-ahead bias: lifetime obligations are as-of-today, not as-of-contract-end. Read the ordering as the finding, not the exact percentages.

HHS, VA, SSA and Commerce are clean public-trust lanes; DHS is mixed

Clearance requirements decide who can realistically compete — and who can be hired. Four of the five agencies are essentially all public-trust work. DHS has to be split by component.

• ›Commerce is pulled whole because the API filters by top-tier agency only — it brings in Census, NOAA and USPTO, all clearance-free.
• ›Applying the DHS split (keep USCIS and FEMA, drop CBP, ICE, TSA, USSS, Coast Guard) removed four previously-surfaced firms whose work was unreachable without a clearance.