Scoring you can audit, on your own corpus.
PRAPI scores journalist queries against your brands and improves from real outcomes. Every change to that scoring is captured in an audit log you can read: what changed, when, and why. It runs on your corpus, not a shared black box, and sharpens as your data grows. With your sign-off, never silently.
Most scoring is a black box. This one keeps receipts.
The PR tools you have used score you with a model you cannot see and cannot question. When a number changes, you do not know why. PRAPI takes the opposite stance: every weight, threshold, and prompt that shapes your scoring is versioned, and every change is written to an append-only log with the evidence behind it. The intelligence is a feature you can inspect, not a claim you have to trust.
Runs on your corpus
Scoring is computed against your brands, your briefs, your outcomes. Every captured row carries a privacy tier, so your data sharpens your scoring and stays inside your boundary.
Auditable by design
An append-only Learning Log records every change to a weight, threshold, or prompt: the reason, the observation count behind it, and who approved it. Nothing moves without a receipt.
Capture, calibrate, audit, recalibrate.
The same loop runs across every scored module: editorial scoring, voice validation, pitch matching, the Source Score. Outcomes go in; a logged, reviewable change comes out.
Capture
Every action emits a downstream outcome that joins back later: a query scored, a draft sent, a pitch cited, an editor’s cite-or-pass decision. Capture is inline and durable; it never slows the request.
Calibrate
Scheduled jobs read those outcomes and measure where the scoring is miscalibrated, the precision and recall at each threshold, which signals actually predicted a citation. Calibration is asynchronous, so learning never taxes a live pitch.
Audit
Each review writes one Learning Log entry: the module, what was examined, the recommendation, the evidence, and the confidence. This is the part you can read. It is the difference between a system that improved and a system that can prove it.
Recalibrate, with sign-off
A proposed change is applied with operator approval, before and after state logged. As a module earns a track record across many cycles, more of that can run hands-off, but the audit trail is the same either way: no silent edits to how you are scored.
What is running today, and what is ahead.
We would rather be precise than impressive. Here is the honest state.
Live now
Outcome capture across the scored modules. Scheduled calibration jobs. The Learning Log audit trail. And a customer-facing Source Score that updates daily from real snapshots.
Ahead
The corpus is still being seeded, so the largest sharpening is in front of us, not behind. As your outcomes accumulate, the calibration loop has more to learn from and the scoring tightens. The log will show each step when it happens.
The live proof point: Source Score
A public contributor profile shows the machinery working: a Source Score from 0 to 100 that blends citation frequency, contribution volume, directory longevity, and topic breadth, with the exact weights version it was computed with, the snapshot date, and a 30-day trend line. It updates daily. It is the same capture-and-audit loop described above, surfaced to the person it scores.
How does PRAPI learn?
Every outcome is captured: which queries you scored, which drafts you sent, which pitches earned a citation, how an editor decided. Scheduled calibration jobs analyze those outcomes and measure where the scoring is over- or under-confident. The result of every review is written to a Learning Log you can read, and the scoring recalibrates from it as your corpus grows.
Is the scoring a black box?
No. That is the point of this page. Most PR tools score you with a model you cannot inspect. PRAPI keeps an append-only audit log of every change to a score weight, threshold, or prompt: what changed, when, the evidence behind it, and who approved it. You can see why a number moved.
Does it learn on a shared model or on my data?
Your corpus. Scoring is computed against your brands, your briefs, and your outcomes. Calibration carries an explicit privacy tier per row and respects those boundaries, so your data sharpens your scoring, not a competitor’s.
Does the system change my scoring automatically?
Not silently. Calibration runs on a schedule and proposes changes with the evidence attached; a change is applied with operator sign-off, and the before and after state is logged. As a module builds a track record across many cycles, more of that can run hands-off, but every change stays in the audit log either way.
What can it honestly do today?
Today the engine captures outcomes, runs the calibration jobs, and logs every change, and the Source Score on contributor profiles updates daily from real snapshots. The corpus is still being seeded, so the biggest sharpening is ahead, not behind. We would rather show you the audit log than claim magic.
Where can I see it working?
A public contributor profile shows a live example: a Source Score from 0 to 100, the weights version it was computed with, the snapshot date, and a 30-day trend. That is the same capture-and-audit machinery, surfaced to the person it scores.
Scoring you own, on data you own.
The learning loop runs on the same engine, brief.md, and quality gates as every other PRAPI surface. Your brand context is a portable file you control, and the scoring built on top of it keeps a record you can read.