Calling an AI smart tells you almost nothing about whether it can complete your job. The word collapses several different things into one label: model reasoning, available context, persistent memory, tools, permissions, verification, and judgment.

Once those layers are named separately, a lot of apparently contradictory AI experiences stop being contradictory.

One person says Claude solved the problem immediately. Another says Claude could not touch the file. One person says Codex completed the work. Another says Codex lost the instruction after one turn.

They may all be reporting accurately.

The identical model can produce different systems

Imagine identical twins with the same education and memory.

One enters a workshop containing the project, the rules, the history, the tools, and permission to test the finished work. The other enters an empty room where questions arrive through a speaker.

The first twin repairs the machine. The second gives advice.

"The first twin is smarter" is the wrong semantic label for what happened. The meaningful difference is working context.

That gives us the first distinction:

  • Model intelligence is the reasoning ability supplied by the underlying model.
  • Product capability is what the application lets that model see and do.
  • Working environment is the project-specific context, rules, files, tools, permissions, and boundaries available in the session.
  • Reliability is how consistently the combined system produces a correct, verified, recoverable result.

Those terms point to different things. Treating them as synonyms makes comparison almost useless.

Context is not memory

Context is what the system can use right now. Memory is what persists and can return later.

A session may have a large context window and still know nothing about yesterday's decision. Another may remember a preference but lack the source file required for today's work.

Project instructions create another category. A rule stored in a file is persistent as data. It only becomes working context when the current environment loads it. The rule can exist globally in human terms while being absent from the model's current input.

This is why "the AI forgot" often names the wrong failure. The actual failure may be:

  • the instruction was never loaded;
  • the relevant history was outside the session;
  • the product did not expose the file;
  • the session could read the rule but not act on it;
  • the task moved into another environment with a different context boundary.

Better labels create better fixes.

Tool access is not task completion

An AI with a shell is more capable than the same model without one for jobs that require files and commands. That still does not make every shell action correct.

There are three states worth naming separately:

  1. Proposed state: the AI described what should change.
  2. Attempted state: the AI used a tool or changed a file.
  3. Verified state: the target system now shows the intended result.

The language around these states is often sloppy. "I fixed it" may describe the attempted state. The reader hears verified state.

That semantic gap becomes operational risk.

Capability is not judgment

Judgment is the ability to choose an action that respects the goal, the evidence, the limits, and the cost of being wrong.

A model can be highly capable and still have poor judgment about which layer it is safe to change.

On September 9, 2026, a change to the Digital Karma Data Warehouse damaged derived crawler measurements. The raw request evidence survived. The summaries and cache states built from it needed recovery.

That incident provides a clean ontology for data work:

ConceptMeaningWarehouse example
EvidenceThe recorded event that must remain inspectable6,621,793 live raw request rows
InterpretationRules applied to evidenceCrawler identity, purpose, scanner exclusion, and spoof handling
MeasurementA derived count or relationshipDaily crawler totals and query-level page associations
PresentationHow the measurement is displayedGood Bots columns, charts, and detail panels
Reliability controlWhat detects or reverses a bad resultBackups, atomic updates, receipts, rollback, and 33 regression checks

The recovery validated 71 dates and preserved two unavailable dates as gaps. It replaced 280 table and day states while leaving the live raw request rows unchanged.

"The warehouse was fixed" would flatten all of that into a sentence too broad to audit. The named layers show what was recovered, what remained untouched, and what is still unavailable.

Environment awareness is its own capability

An AI can know how PHP works and still edit the wrong PHP file.

It can understand structured data and still change visible author copy while missing the separate JSON-LD source. It can know the project's rule but fail to notice the current session is operating against an old local copy instead of the live server.

Environment awareness answers questions such as:

  • Which project is active?
  • Which source is authoritative?
  • Which files are local, generated, or live?
  • Which rules govern this content type?
  • Which tools are connected here?
  • What is outside the allowed boundary?

This layer explains why a product comparison based only on answer quality can mislead. The answer may be excellent while the system awareness required for safe execution is absent.

Communication is part of reliability

An AI that distinguishes "I know how" from "I can do it here" is more reliable than one that lets the reader infer completion.

The same applies to future promises. A system may say it will return with an update. Unless a scheduler, background task, monitor, or notification mechanism exists, that statement names an action the session cannot carry.

The problem is not tone. It is a broken relationship between the words and the available mechanism.

A more useful AI evaluation vocabulary

Instead of asking whether an AI is smart, score the combined system across these dimensions:

DimensionQuestion
ReasoningCan the model understand and solve the subject problem?
Context continuityDoes the relevant project state stay available across the work?
Instruction persistenceDo the governing rules load in the environment where action occurs?
Environment awarenessCan the AI identify the authoritative source and current system state?
Tool accessCan it use the files, shell, browser, server, database, or app required?
Boundary controlDoes it know what it may change and what must remain untouched?
VerificationCan it inspect the result after acting?
CommunicationDoes it distinguish proposals, attempts, results, limits, and future actions?
RecoveryCan a wrong action be reversed without losing the evidence?

This is semantics doing practical work. The system becomes easier to evaluate because each label points to one failure class instead of making "intelligence" carry the entire conversation.

The name changes the decision

If the failure is reasoning, a better model may help.

If the failure is missing context, load the project.

If the failure is instruction persistence, fix how the environment supplies rules.

If the failure is tool access, move the work into a connected session.

If the failure is verification, add an observable check.

If the failure is judgment, narrow the boundary and keep a human decision at the dangerous point.

If the failure is recovery, protect the source before increasing capability.

"The AI is not smart enough" gives you one expensive answer. Precise language gives you the correct next move.