A daily agent that hands over a reading queue, and says what it could not judge

The queue

Ranked by the strict reading. Nothing here is open until a person reads it — the machine axes together describe a paper that is recent, on-topic, by named authors and carries a DOI, which is several hundred papers a week. The gap column is how many points are resting on judgment nobody has applied yet.

Denominator

A shortlist without a denominator is a recommendation. With one, it is a measurement.

General news aggregators are deliberately excluded. A news feed delivers coverage-of-coverage, which is the material this rubric exists to reject — ingesting it would be paying to generate work for the filter.

The axes

Calibration — the test that can fail

The three dossiers already published were chosen by hand, months before this rubric existed. If the rubric scores them below its own open line, the rubric is wrong: the run exits non-zero and refuses to rank anything.

Current state: PASS

This was not the first design. The first version let a candidate reach open on machine axes alone, and on the first live run three exoplanet papers were ranked as dossiers on the strength of having recent dates and DOIs. The band rule was changed rather than the weights, because the weights were not the thing that was wrong.

Two parser bugs, and the threshold they quietly invalidated

Recorded here rather than fixed silently, because the interesting part is not the bugs — it is that the output looked healthy for as long as they lasted.

The Nature feed supplies 75 of the 155 daily candidates. Every one of them was arriving with an empty date and an empty identifier. Not because Nature withholds them: the feed states both, as dc:date and as dc:identifier. The date parser tried three formats and Nature's plain 2026-08-05 was not among them, and the identifier was being scraped by regex from the link and description, where the DOI does not appear — while sitting in plain sight one element away.

Under the strict reading an unmeasured axis costs its full weight, so each of those items was docked 35 points of 100: fifteen for freshness, twenty for retrievability. Half the daily pull was being penalised for a missing format string. Nothing in the output said so, because a genuinely weak candidate and an unparsed one produce the same low number.

So the line was re-derived from something that cannot drift with a parser: the four machine axes can award at most 65 points, and a candidate must now earn three-quarters of them — 49 — before it costs a person twenty minutes. That admits 2 of 155 today, or about 14 a week against a lane that can genuinely read fifteen; the agreement is offered as corroboration and not as the reason, because choosing a threshold for the count it produces is how a rubric ends up only ever agreeing with the person who wrote it. The reason is that an absolute line can report a weak day. A percentile admits the same share of whatever arrives and therefore cannot.

The federal-register lane was not affected and was not adjusted. Its documents score zero on freshness because the newest is dated 2024-12-11 — the UAP apparatus has filed nothing in twenty months. That is a measurement, not a bug, and it is the single most interesting thing this pipeline currently reports.

A third bug, and what the corrected axis then admitted

The parser fixes above repaired what the sources delivered. This one was in the scoring itself, and it had been inflating precisely the candidates the lane least wanted.

lane_fit tested each of the 1,012 vocabulary terms with a plain substring match. Nothing in the output showed what that meant in practice, so it was worth printing the matches and looking: sting was matching inside existing, pear inside appearing, lore inside explorer, craft inside spacecraft, and occult inside radio occultation — a standard astronomy term. A paper on interplanetary radio occultation was scoring as on-lane partly because the word contains the word occult.

Each false hit paid the same floor weight a genuine one-off subject earns, and the axis summed every match rather than the strongest ones — so eight accidents maxed an axis worth 20 points. Its own comment had always said "three central subjects present is a strong match", but the code divided by a literal 3.0 while only one term in the corpus carries full weight. No candidate could reach the ceiling through central subjects; it could only get there by accumulating generic words. The axis was measuring abstract length.

What the corrected axis then admits is worth stating plainly rather than tidying away. The three dossiers this lane actually chose to publish average 0.44 on lane_fit. The three candidates the same axis ranks highest today average 0.69, and 2 of them score above every published dossier. They are ordinary planetary science — crater ray lengths, an instability in the Martian ionosphere.

Not corrected here, on purpose. Re-weighting an axis against three published examples is fitting to three examples, which is the failure this whole page argues against. The test that settles it is dull and takes months: score each dossier as it is opened, and check whether lane_fit separates the opened set from the passed-over one at n large enough to mean anything. Until then it is recorded as a known defect with a stated size, which is worth more than a confident re-weighting.

Where this fits

Triage decides what to look at. The claim rubric decides what a claim is worth once you have looked. The dossiers are what comes out the far end, each with a call and a set of options. Full rule: RUBRIC.md, version 1.0.