The government tells you who it's paying, months before the earnings report. We tested whether that's worth money.

When a federal agency signs a contract, it does not wait for the vendor's quarterly earnings call to tell you.▸ Civilian agencies must report every contract action to the government's procurement database within three business days (FAR 4.604), and USAspending.gov publishes it — recipient, amount, date — through a free public API. Revenue that will not appear in a 10-Q for one or two quarters is sitting in a government database now, named and dated. The question a research desk should ask is not whether that is interesting. It is whether it is priced.

So we built the map.▸ Three public data streams, each earlier in the information chain than a securities filing: daily short-selling volume from FINRA, public attention from Wikipedia's pageview logs, and federal contract flow from USAspending. Every test registered before looking — signal, horizon, universe written down first — because eight tests means eight chances to fool yourself, and the significance bar must rise to match: |t| ≥ 2.73, the Bonferroni correction at eight trials on top of the raised standard the literature already demands (Harvey, Liu & Zhu, 2016).

Seven of the eight failed.

The failures are the context that gives the pass its meaning.▸ Short-sale flow produced a mean spread of +485 bps a month at t = 3.07 that dissolved under one median — volatility masquerading as alpha, dissected in an earlier piece. Wikipedia attention showed nothing on large caps at any horizon. Contract flow measured against the S&P 500 came close twice and cleared nothing.

The one that crossed: rank the 32 publicly traded federal contractors in the panel each day by their trailing 21-day civilian contract flow, z-scored against each company's own year-long baseline.▸ Measure each stock not against the market but against the median contractor in the same panel — so a defense-sector rally cannot masquerade as signal; the ranking has to work within the industry. Top quintile minus bottom quintile: +195 bps per 21 sessions on the mean, +146 on the median — the two summaries agree, which is precisely what the short-sale mirage could not manage — and +155 bps net of a double round-trip cost haircut. Rank IC +0.064, t = 2.91, on 859 evaluation days.

Two hygiene details carry the honesty.▸ Every transaction enters the panel only seven days after its action date — later than the three-day rule requires, deliberately. And every Department of Defense row is thrown away: DoD withholds its awards from public view for 90 days (FPDS data-availability notice; USAspending, About the Data), so a backtest that could see them would be reading data the public did not have. The signal is built only from what a stranger with a laptop could have downloaded that morning.

A desk that has buried seven signals does not celebrate the eighth.▸ It runs the diagnostics that would expose it.

The decomposition found something worth saying out loud: the passing variant's margin over the bar comes from 69 extra sessions of early 2023 that its sector-neutral label unlocked.▸ On the 790 days both label constructions share, the t-statistic is +2.52 — below the bar. In those 69 early sessions the signal was three times as strong as its full-sample average. Nothing about that is disqualifying — every cut of the data is positive: both horizons, all three universes, early window and late, mean and median. The artifacts this desk has caught did not look like that; they inverted somewhere. But a pass whose margin lives in its oldest 8% of days has not earned confidence. It has earned measurement.

What an in-sample pass proves is limited: one history, one panel, with the multiple-testing penalty as the only guard.▸ It says nothing about the next 859 days. So the desk's response is not a position — it is a forward clock. Every night, the current ranking of all contractors is written to an append-only, timestamped log before the outcomes exist; every week, matured snapshots are scored with exactly the in-sample metric. The benchmark to beat is written down: IC +0.064, +195 bps per 21 sessions. The first matured readings arrive in mid-August 2026. Until the forward numbers hold up, the standing rule applies: no capital — paper trading included.

Scope: 814,796 federal contract transactions for 35 recipients (2022-07-10 → 2026-07-07); the passing test evaluates 32 publicly traded contractors over 859 days at a 21-session horizon, sector-neutral label (return minus same-date contractor-panel median), entry at the next session's close, 7-day publication lag, DoD rows excluded. In-sample; forward validation began 2026-07-10. This is research forensics, not investment advice.