Arcvue
Opportunity intelligence—screening that shows its working
The system is never allowed to be confidently wrong in the dark.
Arcvue Opportunity Intelligence screens federal solicitations against your past performance and capability—deterministic rules decide what is removed, a model only ranks what survives. This page follows Arcvue screening one solicitation through nine steps, from four thousand notices to the thirty-six a person sees.
Specimen Systems LLC · an example 120-person contractor · all figures synthetic
A solicitation arrives, and it is one of four thousand nobody has time to read.
Nothing is inherited yet. Federal opportunity feeds—SAM.gov, GSA eBuy, and anything else with a portal—produce more than any business development function can evaluate, so the filter is keyword rules plus somebody's memory of which agencies have been worth bidding.
The question this page follows is not whether Arcvue picks good opportunities. It is why is this one still here at each stage—what has actually been established about it, as opposed to assumed, guessed, or inherited from a score nobody can explain.
The design principle everything below follows from: Arcvue is never allowed to be confidently wrong in the dark. Being wrong is survivable. Being wrong invisibly, with a plausible-looking number attached, is not.
Synthetic. An invented pipeline. Every one of the 1,600 removals is retained with the rule that made it, and is reversible. A funnel that cannot show you what it discarded is asking to be trusted, which is the one thing a screening system should never ask for.
Sixteen hundred are removed, and every removal is a fact rather than a judgment.
The first pass over the four thousand from stop 01 decides only what can be decided against the company profile without an opinion. Two of those decisions end a pursuit: a response window already too short to answer, and a title carrying one of the firm's own exclusion keywords. Those are the kills, and there are no others.
The eligibility facts do something more useful than kill: they route. A set-aside this firm cannot hold, a vehicle it is not on, a clearance floor it cannot meet, a size standard it is over—none of those ends the opportunity, because not being able to prime something is not a reason to stop looking at it. They move it to the teaming lane, where the question becomes who you would have to partner with.
None of these is a matter of degree. A firm either holds the vehicle or does not, and a response window either fits or does not. That is what makes them safe to act on automatically—and it is exactly why this stage is the only one Arcvue performs without a human ever being involved.
A no-bid is a status, not a deletion—and the record keeps the rule that set it.
A screen-out at stop 02 is not a disappearance. It carries the pass it happened on, the rule or rationale, and a timestamp—and the set is queryable. You can ask what was screened out last quarter and why, and get an answer rather than an absence. A person who disagrees with the machine puts it back, and that is the part that makes the rest of it safe.
Which matters for two reasons that pull in different directions. It is an audit trail: a removal you cannot explain is indistinguishable from a bug. And it is training signal: the removals are the negative examples, and a system that deletes them has thrown away half of what it needs to get better. Reversible, too—a rule that turns out to be wrong is a rule you can undo, on records that are still there.
A low score takes an opportunity off the list, and it can never do that quietly.
Nothing is deleted, and everything says why it went.
Models read every surviving opportunity and score it—fit against capability, past performance relevance, the competitive field, and whether the work already has a holder. A score below the line takes it off the working list. What the model cannot do is take one off without recording why, and it cannot delete anything at all.
The reason for the constraint is that a model is confident in the same voice whether it is right or wrong. A deterministic gate that removes wrongly is a bug you can find by reading the rule; a model that removes wrongly and says nothing is a silence. So every removal carries the score and the reason that produced it, and the twenty-one hundred below threshold stay queryable. It is a junk folder, not a shredder.
The first thing a capture lead asks is whether somebody already has this, and the screen asks it too. A named incumbent and an unnamed one are different facts here—a recompete with a name attached is a position you can size, while one without a name is uncertainty, and scoring the two identically would be the easy thing rather than the right one. Neither is a reason to stop looking. An incumbent is a reason to rank an opportunity differently, or a reason to look at teaming.
The vast majority parse. Twelve do not, and they say so rather than guessing.
Federal solicitation documents are not uniform. Some are scanned. Some put the requirements in an amendment posted three days later. Some are a Word file with the substance in a table that extracts as one unbroken line. Arcvue reads the vast majority of them. It does not read all of them, and it is specific about which is which.
A record it cannot read is marked needs_data and shown as such. It is not given a score. Twelve of the final three hundred arrive that way—and what happens to them is the ordinary thing: most are opportunities the firm would pass on regardless, and they are dismissed in a glance. The few that are genuinely of interest take a couple of minutes of human input against a named document.
Which is why the flag is worth more than a score would be. It names the document and what is missing from it, so the two minutes go somewhere specific. A number in its place would have sent the same record to the same queue carrying no instruction at all—and would have made a document nobody could read look exactly like one the model had rated and found ordinary.
A NAICS code is a label somebody else chose, which is why it is a signal here and never a filter.
The ranking at stop 04 reads each opportunity against a profile of the firm: the NAICS set it works in, its size status, its facility and personnel clearances, the vehicles it holds, and the competencies it can evidence. That profile is the customer’s own record rather than something inferred about them, which is why it can be argued with.
NAICS and PSC enter that match as signals the model may weigh, never as filters that exclude. A NAICS code is assigned to a procurement by a contracting officer; it is not derived from the work inside it, and it is frequently the closest available label rather than the right one. A screen that filters on NAICS has inherited somebody else’s clerical decision and is treating it as a qualification—dropping work the firm could do, and keeping work it cannot.
An eligibility fact is certain; what follows from it is not. A set-aside this firm cannot hold is a fact about who may prime, not a fact about whether the work is worth pursuing. Certainty about a label is not certainty about an opportunity, and a screen that treats a code as a gate has confused the two.
The funnel is tuned to fail toward showing you too much.
Every threshold in Arcvue could be set two ways, and the choice is not neutral. Bias toward precision and the queue is short and occasionally missing the one that mattered; bias toward recall and it is longer and complete.
Thresholds here bias to recall; ranking provides the precision. Borderline anything surfaces to a human queue rather than being resolved quietly, because the two errors are not symmetric—a marginal opportunity shown to you costs thirty seconds, and one withheld costs a bid you never knew existed. You can always ignore a row. You cannot review a row you were never shown.
A new scoring weight does not count until it has been tested against what you actually won.
The ranking at stops 04 and 07 rests on weights, and a weight is an opinion until it has been checked. Changing one is easy and feels like tuning; the risk is that it quietly reorders a queue that nobody re-examines because the queue looks the same shape.
So a change is re-scored against the historical bid set first: known wins must rank high and known no-bids must rank low. The calibration report goes to the operator for sign-off, and until they sign, the new weights drive nothing. The test is the firm's own history, not a benchmark—which is the only version of "calibrated" that means anything to a specific business.
Thirty-six reach a person, and the other 3,964 are still there.
Four thousand became thirty-six—0.9% of the input reaching a human decision, and nothing was discarded to get there. Sixteen hundred were removed on facts, each keeping the rule that removed it and each reversible. Twenty-one hundred were ranked down and remain visible. Twelve arrived flagged rather than scored.
So the funnel is narrow at the end and transparent the whole way along. Those are two different properties, and the second one is what lets a capture lead act on the first: a queue of thirty-six is only worth reading if you can ask what is not in it.
The thirty-six do not land in one shared queue. Each one is routed to the division that would run the work, following how your business is organized, so it reaches the lead best placed to judge it. Each lead clears their own queue on their own schedule, rather than waiting on somebody else to get to theirs.
And what happens next is priced. An opportunity that survives this is the one the Arcvue pricing workspace builds a bid for, from four independent sources on one rate—and whose outcome comes back into the record so the next screening pass starts with more than the last one did.