Haltslop is not an AI chat app. It will not write the final answer, choose your life, diagnose you, or pretend its local rules understand more than they do.

Haltslop checks the part of the job that usually gets blurred when AI is too easy to ask: the goal, source material, audience, evidence, judgment owner, verification path, or stopping rule.

It may tell you to use AI with constraints. It may tell you to make a list, ask a person, verify first, decide first, or step away before prompting.

No. Haltslop is deterministic. It uses local rules, not a model call.

That is the point: it should slow the prompt down before fluency makes the work feel clearer than it is.

Most AI slop starts before the model responds.

It starts when the task is vague, the source is missing, the human judgment is outsourced, or fluent output arrives before the real work is named.

Haltslop asks one sharper question so the next move is smaller and less fake.

Haltslop starts with three public buckets, then checks for more specific risks from its rules engine.

• Output slop: when AI makes more artifact than the job needs, such as generic copy, tool overkill, source-light synthesis, coordination drag, or a coding harness change without a tight acceptance boundary.
• Reasoning slop: when fluency can hide uncertainty, especially verification tax, automation bias, synthetic trust attacks, or claims where the source of truth is the real work.
• Reliance slop: when AI starts owning the judgment, such as cognitive debt, judgment outsourcing, agency loss, life-decision outsourcing, prompt spiral, taste erosion, or problem-definition avoidance.

The implemented risk set includes verification tax, automation bias, cognitive debt, false productivity, context collapse, prompt spiral, taste erosion, agency loss, jagged frontier / task-fit, coordination drag, synthetic trust attacks, and life-decision outsourcing. The rules engine also uses tool overkill, judgment outsourcing, and harness blast radius as primary verdict risks.

Examples: a claim may need "what source of truth verifies this?"; a life or career question may need "where should AI stop and your preference begin?"; a Codex task may need files, tests, rollback, and a small diff before execution.

This is a product summary, not a debugger. Haltslop should not expose ranked classifier candidates, scores, regexes, or full rule traces on the help page.

Paste the prompt you were about to send to AI. Rough is fine.

If the frame is muddy, use this shape: "I need help with ____ so that I can ____."

No login. No model call. No saved transcript.

The active chat lives in this browser while the page is open. If you create a share card, Haltslop uses sanitized verdict state, not your raw conversation.

Restate the job plainly. The deterministic engine can miss phrases, but it should recover when you name the task, goal, source, judgment owner, verification path, or human next move.

Try: "I need help with ____ so that I can ____."

Yes. Harness tasks are part of the point.

Haltslop may lower risk when the task has files, tests, logs, and a clear acceptance boundary. It may raise risk when the agent could make broad executable changes without a small diff or rollback path.

AI is usually safer when you can name the source, audience, acceptance boundary, verification path, and who owns the judgment. If one of those is missing, that missing boundary is probably the next move.