lenses 4 verify adversarial artifact markdown

Method.

Every finding quotes the reviewed content. This page is the pipeline, the evidence rules, and what a run does not see. The pre-deploy report documents this site as it stood before those findings were fixed. The previous version of this site has its own archived report; its receipts live inside it.

The pipeline

  1. Crawl, render, observe

    Headless Chromium loads each page, same-origin and bounded, respecting robots.txt during discovery. It keeps the rendered DOM and a behavior log: network requests, storage writes, console errors. Single-page apps are read after JavaScript runs. Source mode skips the crawl and reads a repo instead.

  2. Deterministic checks

    Exact-match passes run first: crawl errors, placeholder hrefs, draft remnants, head-metadata gaps, duplicate titles. No model involved.

  3. Four lenses

    Four AI reviewers, each holding one lens: cold stranger (first visit, thirty seconds of patience), skeptical buyer (assumes shilling until shown otherwise), detail nitpicker (pre-launch sweep), register keeper (every sentence must be for the visitor). Each lens reads independently. Agreement is recorded, not assumed.

  4. Adversarial verification

    A separate pass re-reads every candidate against the cited content and tries to refute it. Misquoted, overreaching, or fine-in-context candidates die. Confirmed findings get corrected line numbers, a severity check against fixed anchors, and a fix note for a coding agent. The runner then mechanically re-finds every receipt; a finding whose receipt cannot be re-found is dropped.

  5. The report

    Canonical markdown: stable SG-NNN IDs, verbatim receipts with paths and line numbers, fix scopes, a rollup by failure class. HTML is rendered from that markdown, with copy-for-agent buttons.

Evidence rules

No receipt, no finding
Every finding quotes your content verbatim. A receipt the runner cannot re-find is dropped, not shipped.
Lens consensus
Each finding shows how many of the four lenses caught it independently. Lenses 4/4 means every reviewer tripped on the same defect.
Severity anchors
Trust-killer: an active, visitor-verifiable trust breach, such as a claim contradicted by shipped behavior. Embarrassing: a human editor would refuse to ship it. Nit: polish. When a finding sits between tiers, it takes the lower one.
Fix scopes
Each finding carries this-line, sweep-all-instances, or restructure, plus an imperative fix note. The report is written to be pasted into a coding agent without extra context.

What a run sees

Seen: rendered pages in a real browser; network requests and their payloads; storage writes; console errors; source, when you point SlopGuard at a repo.

Not seen: anything behind a login; form submissions, purchases, or any interaction beyond navigation; external facts, which are flagged for you to check rather than asserted.

A free self-serve scan at /scan/ is kept about 24 hours, then deleted. A queue run, or a run we perform ourselves, is kept as history so runs can be compared; that history stays private unless you publish it, is never sold or used for training, and is deleted on request.

Standing benchmark

SlopGuard is scored against a plain “review my site” chat prompt on identical targets: confirmed-finding precision and trust-killer recall. If receipt-verified findings do not clearly beat a free chat prompt, SlopGuard has no reason to exist.

Round 1 · 2026-08-09, this domain’s previous site. A real ChatGPT session with browsing quoted the site accurately and caught the copy’s repetitive cadence. It surfaced none of the five verified trust-killers in the archived report, praised the affiliate disclosure that run proved false on the page where it sits, and reviewed a different company’s website on its first attempt. The session transcript is not published, so treat this round as our account, not a receipt. Round 2 runs on the queue’s first external site.

Reflexive rule

This Hexento surface gets a SlopGuard run before it ships. The 2026-08-09 rebuild run found 22 verified defects, 6 of them embarrassing; the report is public and those findings were fixed before that ship. A rerun on the fixed pages found 14 residual findings, none trust-killers; that report is public too. The 2026-08-13 pass on these rebuilt intake pages found 21 more verified defects, one trust-killer among them; that report is public and the findings were fixed before this ship. Each report documents the site as it stood when its run ran. If we skip our own tool, the tool dies.