Hallucination is twenty-eight bugs

Every one of them has a name, a one-line description, and a fix. None of the fixes is "better prompting".

Stop saying hallucination.

Not because it is wrong. Because it is a category, and nobody has ever fixed a category. You fix a specific failure with a specific detector, and "hallucination" hides twenty-eight of them.

All twenty-eight, one line each

class	failure mode	what it looks like
1	Unsupported-Claim Confidence	States a checkable fact with nothing behind it
1	Citation Resolution Failure	Cites a source that does not exist, or does not say that
1	Claimed-Action / Tool-Log Divergence	Says it ran a search; the log says it did not
1	Planted-History Recall Failure	Misremembers what was agreed earlier in the chat
1	Schema-Correspondence Fabrication	Invents a match between two tables' columns
2	Pushback Instability	Drops a correct answer because you disagreed
2	Flattery / Agreement Drift	Shifts its view toward what you seem to want
2	Authority-Framing Sensitivity	Answers differently because you claimed a title
2	Basis-Demand Evasion	Asked why, restates instead of giving a reason
3	Constraint Decay at Depth	Forgets a rule set forty turns ago
3	Format-Lock Under Distraction	Breaks the required format when the task gets hard
3	Cross-Turn Schema Validity	JSON that stops validating a few turns in
3	Needle-Position Sensitivity	Finds a fact at the top of a document, misses it mid-way
4	Method/Execution Split	Right approach, wrong arithmetic
4	Self-Consistency Spread	Ask five times, get three answers
4	Verification-When-Asked Failure	Asked to check its work, approves the mistake
4	Dimensional-Analysis Failure	Gets the units wrong
5	Prompt-Perturbation Drift	A typo or trailing space changes the answer
5	Paraphrase Non-Invariance	Same question reworded, different answer
5	Option-Ordering Sensitivity	Picks differently when the options are shuffled
6	Tool-Routing Error	Calls the wrong tool
6	Tool-Argument Validity	Calls a tool with arguments that do not exist
6	Agentic Loop / Non-Termination	Repeats the same call and never finishes
6	Tool-Return Groundedness	The tool said 12; the answer says 20
6	Ledger Reconciliation Failure	Loses track of state across a run of actions
7	Post-Cutoff Fabrication	States an outdated fact as current
7	Recency-Confidence Miscalibration	Equally sure about stable facts and stale ones
7	Date-Arithmetic Error	Given today's date, gets "in 90 days" wrong

The full taxonomy. Every mode has its own entry, with probes and guidance, in the Failure Mode Index.

What the average costs you

A single "hallucination rate" averages a model that invents citations with one that loses count of a ledger. Improve one, worsen the other, and the average does not move. You shipped a regression and the dashboard stayed green.

The probe

Merge these two sheets on customer ID. (Sheet A has it in column C. Sheet B has it in column E.)

fails

"Matched on Customer ID, column C in both sheets" — and a formula that joins the wrong columns.

passes

"Sheet B has no customer ID in column C. I found one in column E — confirm before I merge."

Schema-Correspondence Fabrication. It will never appear in a hallucination benchmark, and it is the one that corrupts a finance model on a Tuesday.

Placement is judgement, not measurement. The point is the top-left: the failures that matter most are the ones that look exactly like a correct answer.

Where the popular advice is right

"Use retrieval" is good advice for about four of the twenty-eight. Supplying a real source does reduce fabrication. It does nothing for a model that caves under pushback, loses a rule at depth, or miscounts a ledger — and it can make needle-position failures worse, because now there is more haystack.

The hard part

Naming is cheap. Detecting is not. Each of these needs a probe that can fail, a control that tells a real failure from noise, and an interval honest about how few cases you ran.

But you cannot build a detector for a word. You can build one for "the final answer contradicts the tool return" this afternoon.