Failure modes
A language model fails in ways built into how it works, so no single check catches every failure. G6 stacks its components like slices of Swiss cheese: each one names the failures it guards against, and where one slice has a hole, the next one catches what slips through.
This is the map of those slices: every failure mode G6 names, the components that guard against it today, and the ones whose own use can cause it. It is generated from the code, so a component appears here only when its package declares the mode.
- Failure modes
- 240
- Guarded today
- 148
- No guard yet
- 92
Guarded today
Each of these has at least one component in the library that declares it guards against it.
-
Ad hoc, unrepeatable choices about what to try next
-
A factual question answered from the model's memory instead of a source
-
Audio output that is silent, clipped or corrupt, reported as done
-
A workflow used beyond what its review covers
-
A coding agent changing an unfamiliar repository without first learning how it works
-
Spending a budget on blind trial and error
-
A shortcut that breaks on cases it has not seen
-
A change that breaks something that worked
-
A hand-built page that breaks on some screens
-
A task overrunning its token, cost or time budget
-
A relevant observation buried under old ones
-
A recoverable error that stops the whole workflow
-
An automatic fix applied without checking what it changes
-
A reasoning slip stated with confidence
-
A claimed gain that extra compute or luck explains
-
Context overflowing the model's window or its budget
-
Facts or decisions lost between the steps of one task
-
Rules that contradict each other without anyone noticing
-
Taking a correlation for a cause
-
Data mistakes from ad hoc table handling
-
An interface with inconsistent styling or poor contrast
-
A claim made with no regard for whether the evidence supports it
-
Spending more on model calls than the work needs
-
Made-up facts and quotes
-
A safety check that passes when its parts are missing
-
A failed or blocked search taken to mean there are no results
-
A false analogy between fields that differ in structure
-
An uncertain or evasive answer passed off as confident
-
Code changed by text edits that break its structure
-
A long task drifting from its goal
-
An object dropped or crushed when grasped
-
An agent reporting work as done that the evidence does not back up
-
A result handed over without saying how far to rely on it
-
Boilerplate web code written by hand, inconsistently
-
Acting on a risky or privacy-violating request
-
A warning ignored on the way to an action
-
A model ill-suited to the data it is used on
-
Quality judged by impression instead of measurement
-
A new component missing its tests or standard parts
-
Conclusions drawn from an incomplete search of the literature
-
Rules applied inconsistently across a long task
-
Optimisers disagreeing on what a workflow looks like
-
A motion plan that collides, or asks a body for what it cannot do
-
An error report that leaks code or personal data
-
A failure never recorded, so nothing learns from it
-
A model lost in too much context
-
Messages between agents lost or mixed up
-
Model output that does not match the expected structure
-
A wrong or planted entry in stored memory, used as if true
-
A chart that misstates or hides what the data shows
-
An image's content misread, or not read at all
-
A consequence the rules imply that goes unnoticed
-
Key facts lost past a cut-off in long text, or given values the text does not hold
-
An AI missing signs that a vulnerable person is at risk
-
Work in a specialist field started without that field's evidence standard
-
A constraint the person stated, forgotten between the request and the result
-
A proof of a different statement from the one you meant
-
A safety analysis of a model that differs from the real system
-
Moving the goalposts after the work is done
-
An agent acting step by step with no plan that reaches the goal
-
Generated code that only works on the examples it was shown
-
A coding agent run with more access than it needs
-
A conclusion stated with more certainty than the facts give
-
Work called verified when the evidence does not support it
-
A plan that fails halfway through
-
A task or request sent to a tool or agent that is a poor fit for it
-
Obeying instructions hidden in text it should only read
-
A claim or citation that cannot be traced to the passage it came from
-
A design that fails only in a rare order of events nobody tried
-
Reimplementing a method that already exists
-
Repeating the same mistake
-
The same claim repeated, counted as more evidence
-
A plan that breaks one of your rules
-
A run that keeps going past its limits
-
Tampering with the safeguards themselves
-
A tool silently missing or replaced, with the result reported as if it ran
-
A workflow slower than it needs to be
-
Claims that need sources and have none
-
A citation to text that has since changed
-
Outdated or unofficial material used as if it were current
-
A cost quoted from a price that is unknown or out of date
-
A safety decision altered or lost after the fact
-
An agent weakening tests to make them pass
-
A model scored on data it was trained on, or on a leaky split
-
Setup or training-loop bugs in hand-written model code
-
Unsafe outcomes nobody worked out in advance
-
An agent changing protected code or rules without approval
-
Multi-agent workflows run without limits
-
An agent running generated code without limits
-
An index that outgrows its storage budget
-
A risky operation run without a safety bound
-
Taking one model's first answer unchallenged
-
An answer with skipped steps that nobody checked
-
A big move made without checking it first
-
A maths claim accepted without a machine-checked proof
-
A plan from an agent or a model followed without checking each step against what its actions need
-
A self-change that slips past its own checks
-
A task sent to a paid or unsuitable model nobody chose
-
A paid outside service used without the person choosing it
-
Unclear authority in mixed human and AI teams
-
Work started with no measurable test of success
-
Relevant passages that retrieval never finds
-
Code from a published method that does not do what the method says
-
A proof that only looks finished
-
An answer not grounded in the person's own documents
-
An agent using tools with no guard on what they do
-
Each agent handling failure differently, or not at all
-
A lesson that sounds right and does not help
-
A planning agent acting on a model that has not learned the environment's rewards
-
Facts spread across text with no links between them
-
Code growing more complex than anyone can safely change
-
A model of human recall that ignores how memory fades and interferes
-
A slow decline in quality or speed that nobody notices
-
A plan called the best when nobody proved it
-
Useful heuristics forgotten or never ranked
-
Source material skipped or garbled because its format could not be read
-
A fix made for a failure nobody reproduced
-
A long workflow that cannot resume after a failure
-
An agent taking a harmful or irreversible action nobody approved
-
Unsafe or inconsistent database access
-
Early warnings in a system's own behaviour going unseen
-
A long job blocking an agent or failing unseen
-
A document citing sources that do not support it
-
A conclusion the rules and facts do not support
-
Models tuned by hand without a systematic search
-
An agent design judged by argument instead of by a test
-
A prompt change adopted without testing whether it helps
-
Losing track of which prompt or spec version produced a result
-
Repository changes an agent made that nobody tracked
-
Completion claimed with nothing named that would prove it
-
A vacuous or missing property passing as a real proof
-
Output that no independent check looks at
-
An image, video or 3D output that is blank, corrupt or wrong, reported as done
-
One weak step spoiling an otherwise good result
-
Answers built on thin or weak sources
-
An agent writing files outside its workspace
-
Controlling for the wrong variables
-
A law or ruling from another jurisdiction given as the one asked about
-
A poor solution from the wrong optimisation method
-
Solving the wrong problem
-
Harm or irreversibility underestimated
-
A wrong formula, root or integral stated with confidence
-
A transcript too empty or too long for the speech it came from
No guard yet
Nothing in the library guards against these yet: a hole in every slice.
-
Admin errors in records, filings or correspondence
-
Attempted too much in one patch
-
Conflicting claims about who may decide or act
-
Bookings, schedules or itineraries that clash or are wrong
-
Generated code that does not run or does not work
-
Tests written by the caller taken as independent proof
-
Compaction added unsourced claim
-
A decision lost when earlier turns are summarised to save room
-
Component metadata missing
-
A decision or insight drawn from data nobody checked
-
Late deliveries and inventory mismatches
-
Dependency breakage
-
AI help that leaves people less capable than before
-
Docs not updated
-
A missed drug interaction or dosing error
-
End to end failure
-
Output presented as expert-reviewed when it was not
-
Hallucinated failure
-
Learners who feel they know more than they do
-
An agent improvising in a known risky situation
-
Custom glue code between incompatible components
-
Job agents with inconsistent safety and review rules
-
A fix that does not fix the failure
-
Insecure code shipped
-
Interagent coordination failure
-
Irrelevant exemplar
-
Outages and access problems found too late
-
A secret leaked in code, logs or output
-
Teaching that does not match the learner's level
-
License or copyright risk
-
Tests run in a repository that can change during the run
-
Effort spent on low-value tasks first
-
A manual deployment step done wrong, or impossible to repeat
-
Misallocated resources and unclear objectives
-
A tax rate or rule applied for the wrong year, jurisdiction or kind of supply
-
A misdiagnosed fault or substandard workmanship
-
A policy draft that misreads legislation or polling
-
Output that misses a field's mandatory requirements
-
Unresolved complaints and missed service levels
-
Safety or permit issues missed on a site
-
A screen that found no crisis words read as a sign the person is all right
-
An agent's behaviour never tried in a simulated world first
-
Off-brand or unreviewed content going out
-
Container steps failing without clear errors
-
Output contract missing
-
A change that alters behaviour outside the part it was meant to touch
-
Overbroad retrieval
-
Containers that need root or a daemon to run
-
Prompt altitude mismatch
-
Model calls wired separately for each provider
-
Quality defects and production bottlenecks found too late
-
Repository dumping
-
Risk management failure
-
A broken change left running with no way back
-
Code under test running with the owner's permissions
-
Safety policy failure
-
Workflow behaviour changing silently between releases
-
A decision based on one noisy sensor
-
Skipped safety checks and missed maintenance
-
Source visualisation stale
-
Stale file handle
-
Stale test result used
-
Subagent context leakage
-
Subagent result overmerge
-
Test suite missing or weak
-
Tool context missing
-
Translated code that behaves differently from the original
-
A campaign run without audience or spend analysis
-
Threats and vulnerabilities left unassessed
-
An unprintable model or an invalid machine toolpath
-
Financial risk and compliance gaps left unchecked
-
A message released without checking its coverage or tone
-
A multi-step workflow failing in a step nobody checked
-
An experiment without controls or proper statistics
-
An agent running with more autonomy than its record supports
-
A model shipped without evaluation, drift or fairness checks
-
Ungrounded exemplar use
-
A recommendation not grounded in the client's situation
-
A deployment sent without a readiness check
-
Books reported as reconciled or balanced when they are not
-
Unqualified leads and unreliable pipeline forecasts
-
Unreported hazards and missing time records
-
Flaky model calls with no retry or fallback
-
Clinical notes, plans or judgements reaching people without a clinician's review
-
A design or code change made without review
-
Legal judgements reaching people without a lawyer's review
-
An AI decision in a regulated field escaping review
-
Job agents judged without reviewed test cases
-
A business decision made without a structured case
-
A clever prompt mistaken for a tested workflow
-
A startup building before it validates the idea
-
A plan based on wrong physical expectations