Metrics, Methods & Change Log

Everything this tool measures, how every number is earned, and a dated record of every change ever made to the scoring rules. Nothing about the method is hidden.

What this is — and is not

This is a Value Detector, not an AI detector. AI detectors guess who typed the words and tell you nothing about whether a piece is worth reading. This tool ignores authorship entirely and measures the thinking: the evidence, the originality, the honesty, and what a reader gains. Every grade must be backed by verbatim quotes and specifics from the piece — a score without proof is treated as an error by the engine's own final consistency check.

The nine scored measures

Seven measure quality. Two measure undesirable qualities and are inverted, because low is good. All nine average into the overall number.

What a “5” means — the anchor ladder

5 is average for what is actually published on that specific subject — not an imagined ideal article. If little published writing on a subject reaches a piece's depth, the piece scores high. A piece is never held down for lacking things (original data collection, interviews, public-records requests) that almost nothing in its field has.

Every measure has a printed ladder, and the grader must award the number when the piece meets the description. Example — quality of the evidence:

Two grading methods, every time

Rants, filler, and media filler

What makes a rant

A rant is defined by what it lacks and why it exists — never by its intensity. Force, anger at government, moral certainty, and blunt language are not rant indicators; the most important writing in the American tradition has all of them. A piece scores as a rant when most of these hold:

• Its purpose is to discharge the author's grievance rather than to inform, prove, or persuade.
• Antagonists are unnamed or vaguely gestured at, so no claim can be checked.
• Emotional intensity substitutes for specifics — no dates, documents, quotes, statutes, or named sources.
• It circles and repeats rather than developing an argument from premises to conclusion.
• Digressions serve the author's feelings, not the thesis.
• Nothing survives the reading — no fact, framework, or action.

Filler and the media-filler check

Filler is word count that adds nothing — restatement, generic observations anyone could write without knowing the subject, passages you could delete without losing anything.

Media filler is the content-farm pattern common to local radio-station and aggregator sites, and it gets its own report. Indicators: the whole article reduces to one fact stretched over hundreds of words; no original reporting (rewritten from another outlet, a press release, an agency statement, or a social post); padding devices (restating the headline, “here's what you need to know,” rhetorical questions, a bolted-on local angle); engagement bait and listicle bulk; and low information density.

The report states plainly, in one sentence, everything the article actually tells you, alongside a count of distinct verifiable facts, the approximate word count, whether any original reporting occurred, and what it appears rewritten from.

Invective and wit

Earned vs. empty invective

Name-calling is a legitimate rhetorical tool with a long pedigree — Mencken, Ivins, and Buckley all did it well. What matters is whether the epithet earns itself.

The test in one line: could a reader say who is meant and why the label fits, using only the piece? If yes it is earned; if no it is empty.

Humor and sarcasm — credited only when earned

Wit is not automatically a virtue, and its absence is never a fault. Earned wit illuminates: it makes an argument land harder than plain statement would, exposes an absurdity the reader can now see, compresses a real point into a memorable image, or reframes something familiar. That is genuine original thought, and it raises “new ideas of its own” and “worth your time.” Unearned wit — decoration, in-group snark, a sneer standing in for a point — earns nothing. No points are awarded merely for attempting humor, and none are deducted for being serious throughout.

Judged by its own form

Every piece answers the same nine questions; no form earns a free pass. But the form informs what a strong answer looks like, and the results always state plainly how the form affected the grades.

• Polemic — judged on whether it argues well: principled grounds, internal consistency, vivid evidence, honest handling of facts. Never penalized for being one-sided or morally certain.
• Reportage — judged on sourcing, verification, and fairness.
• Essay or devotional — judged on insight and honesty of reflection, not citation count.
• Citizen guide / analysis — judged on accuracy and usefulness. A piece with a strong viewpoint that also equips readers with mechanics, definitions, and guidance is a citizen guide, not a polemic.
• Satire — judged on whether the exaggeration lands and illuminates.

Source verification

Citations and footnotes are not merely counted — each is checked: does the cited work exist, is it attributed correctly, does it support the claim it is attached to? Glossary definitions are checked for accuracy. Verified apparatus raises scores; fabricated, misattributed, or misused citations lower them and are flagged with the specific problem. Results show the count checked and any flags.

Displayed documents count as shown work. The analysis reads text only, so meeting minutes, letters, and records a piece displays as images cannot be viewed by the grader — they are credited as primary evidence shown to the reader, never treated as missing. Image alt text and figure captions are preserved during extraction so the grader can see they exist.

Disclosed limit: verification is checked against the model's own knowledge, not a live fetch of each source. Only claims with positive reason to be wrong are flagged; an unfamiliar citation is never flagged merely for being unfamiliar.

Liberty & consequences

When a piece involves government action, regulation, or rights, it receives a dedicated analysis in the negative-liberty tradition — liberty as freedom from government interference, powers enumerated and limited, rights preceding government rather than granted by it. It names:

• the specific right or liberty at stake;
• exactly what the government actor did — who, what, when;
• whether it was lawful under the authorities the piece cites;
• the concrete consequences — to those affected, to public trust, and to precedent if the action stands unchallenged;
• the accountability arc: what correction occurred and who forced it, crediting citizens or the author when the piece shows they did.

A government action does not become acceptable because it was reversed. The violation, the harm while it stood, and the fact that reversal required a citizen to force it are all part of the consequences.

The idea graph

Separately from scoring, the tool maps the piece's ideas: 10–16 concepts as circles, 12–20 directed relationships as arrows, and the meaningful omissions.

Center of gravity: the highest-weighted concepts must be the piece's central claim and its stakes — what happened, who did it, what law or principle it violated, and what came of it. When a piece documents officials reversing course because they were called out, that accountability arc is a core concept, never omitted. Practical asides (lawful alternatives, procedures) carry low weight.

Fairness rules

The optional second opinion

A “Verify with a second AI” button sends the piece to an independent model (Google Gemini) that re-scores every measure and independently lists what it considers the usual ground for the topic — blind to the first analysis. Agreements and disagreements are shown side by side; disagreements over 2 points are flagged rather than averaged away. Concepts both models call usual ground are marked ✓; where they disagree, marked ?.

Honest caveat: the two checklist/comparative methods are the same model reasoning two ways — not fully independent. The second-opinion button is what closes that gap.

Storage, determinism, and cost

Scoring runs at temperature zero: the same article produces the same grades on every scan. Every scan is stored permanently in a database keyed to the article, so a piece is analyzed once and served instantly forever after — repeat visits cost nothing and never re-run the AI. An Update link on any result forces a fresh analysis and replaces the stored one. Stored results are labeled with the engine version that produced them, and share links carrying tracking parameters resolve to the same stored article.

Known limits

• Verification checks sources against the model's knowledge, not a live lookup.
• The grader reads text only; images and embedded documents are credited but not viewed.
• Articles are analyzed up to roughly 24,000 characters; longer pieces are truncated.
• The two built-in grading methods share one model — use the second-opinion button for true independence.
• Anchors reward accumulated apparatus, which can under-serve very short work whose value is density rather than volume. Documented here rather than hidden.

Change log — every scoring change, dated

Each engine version is recorded with what changed and why. Several rules exist because bias was identified in live results and corrected; those are stated plainly rather than quietly patched. Stored analyses are labeled with the version that produced them.

Tool and infrastructure changes

← Back to the Article Values Graph™