warrantos tells

warrantos tells is the opinionated sibling of warrantos slop.

slop hunts for chat residue: scaffold that visibly leaked out of an assistant session and shipped by accident: assistant openers, identity disclaimers, stray TODO placeholders. Those patterns are close to objective. A sentence that opens with "Certainly, the analysis below" is almost never something a person typed on purpose into a policy brief.

tells hunts for something harder to pin down: prose that is already clean of residue but still sounds machine-written. Contrastive negation, hedge stacking, em-dash punctuation, a small set of filler phrases, and a drumbeat of formulaic paragraph-openers. None of these are proof that a model wrote the sentence. Humans write every one of them, some of them often. tells flags them anyway because they are disproportionately common in unedited model output, and a reviewer who sees ten of them in one document has a reason to slow down and read closely.

Usage

warrantos tells docs/ README.md         # scan paths for AI-writing tells
warrantos tells --json docs/            # machine-readable output
warrantos tells --badge docs/           # emit a badge URL
warrantos tells --fail-over 0 docs/     # CI mode: non-zero exit above the threshold
warrantos tells --include-fences docs/  # also scan inside fenced code blocks

Same flags and semantics as warrantos slop: [PATH...] (files or directories, default the current directory), --json, --badge, --fail-over THRESHOLD, --include-fences. Fenced code blocks are skipped by default because they usually quote deliberate examples, including the example strings in this document. A TELL SCORE runs 0.0 to 10.0 on the same density formula as SLOP SCORE, and the exit contract is identical: 0 unless --fail-over is exceeded (1) or a path does not exist (2).

tells reuses the slop scanning engine rather than reimplementing it: file discovery, fence handling, display-path resolution, the score formula, and the badge base all come from warrantos.provenance.slop. Only the rule set is new.

Why this scanner is opinionated where slop is objective

slop's pattern list is the canonical _AI_RESIDUE_RULES set shared with the Layer 7 G1 prose-boundary gate: a slop finding and a check violation are always the same rule, and there is little room for house taste in an assistant self-identification phrase showing up in a shipped document.

tells has no such external anchor. Every rule below is a house-style call. Some newsrooms use em dashes freely. Some technical writers stack hedges on purpose to be precise about uncertainty. A team that dislikes "Furthermore" as a paragraph opener will set a lower --fail-over threshold than a team that does not care. tells does not claim these choices are universal; it applies one particular set of them, consistently, so a reviewer gets a repeatable signal instead of a vibe.

Rule families

Each finding reports its file, line, matched text, category, and rule id.

1. contrastive-negation

The "not X, but Y" / "it's not X, it's Y" family, with the pivot required within roughly 60 characters of the trigger phrase so an unrelated long sentence that happens to contain both halves does not fire.

Real example (fenced so the scanner treats it as quoted, not live prose):

This is not just a policy, it is a whole new operating model, but the
paper does not say how.

Deliberately not flagged: bare "rather than" and "instead of" are far too common in ordinary prose to be a useful signal on their own, and the near miss below stays silent because there is no "but" pivot within range: it is just a sentence with two clauses, not the contrastive-negation shape.

Not only did the committee meet, it voted on the motion immediately.

2. hedge-stacking

Two or more hedges from one fixed word list, inside one sentence (sentences are split on period, exclamation mark, question mark):

may, might, could, perhaps, possibly, arguably, potentially, seemingly,
somewhat, appears to, tends to

Real example:

The measure might possibly reduce costs, though implementation could
arguably take longer than planned.

One hedge in a sentence is ordinary, defensible caution about an uncertain claim and does not fire:

The measure could reduce costs over the medium term.

The stack, not the hedge, is the tell.

3. dash-punctuation

An em dash used as punctuation anywhere in a line, or a spaced en dash (with whitespace on both sides). Number ranges written without spaces are left alone: the spacing is what marks the dash as punctuation rather than a range separator.

The scheme failed - not because of design, but delivery.       (em dash fires)
The review ran 2020 - 2021 across three agencies.               (spaced en dash fires)
The review ran 2020-2021 across three agencies.                 (no spaces, does not fire)

This project's own documentation treats "no em dashes" as house law (see the workspace writing rules); here it is one opinion among five, weighted the same as the others, because a general-purpose scanner cannot assume every user shares that preference.

4. filler-lexicon

A short list of near-unambiguous AI filler phrases, case-insensitive:

delve into
rich tapestry
stands as a testament
in today's fast-paced
in today's rapidly evolving
it is important to note that
it's worth noting that
in the ever-evolving
game-changer
unlock the (full) power / potential
seamlessly integrates
at the end of the day,
let's dive in

Real example:

Let us delve into the detail of the funding model.

5. formulaic-transition

Sentence-initial "In conclusion,", "Furthermore,", "Moreover,", "Additionally," -- only when two or more of them (in any combination) appear in the same document. A single one opening a paragraph is normal, human paragraph construction and is not reported. The second and any later occurrence in the same document are reported: the drumbeat is the tell, not the individual word.

Honest limits

tells measures style markers, not authorship. Every pattern above has a plausible human origin: careful hedging, a fondness for em dashes, a writer who leans on "Furthermore" out of habit. A high TELL SCORE is a prompt to open the document and read the flagged lines, never proof that a model wrote them, and a low score is not proof that a human did. Sentence splitting is a simple ./!/? scan with no abbreviation handling, so an abbreviation like "e.g." can occasionally split a sentence early; this mainly affects hedge-stacking and formulaic-transition counts and is a known, accepted trade-off for staying stdlib-only and fast. Rules are tuned for precision over recall throughout: a false positive on legitimate prose is treated as worse than a miss, so tells will under-report before it will nag.