Meaning lives below the label.
An agent has not understood a concept because it can name it, define it, or explain it fluently. Understanding shows up when the concept survives a changed context, a reversed relation, a novel case, and a check against the world.
The practical conclusion: make every consequential answer pass an external depth check before the agent acts.
What the evidence actually supports
The literature rejects both easy stories. “It is only pattern matching” ignores internal world models and concept geometry. “It understands because it sounds right” ignores shortcut learning, memorization, and unfaithful explanations.
Understanding is not a property the model can certify about itself. It is a performance that must survive intervention and external verification.
Not every word needs direct grounding
Dictionary-graph work estimates that a minimum grounding set is roughly 1% of the dictionary; other terms can, in principle, be defined from that base. The exact solution to symbol grounding remains open.
World models can emerge from sequence data
Othello-GPT predicted legal moves with 0.01% top-1 error. More important: altering its internal board representation made it act legally for a counterfactual board — causal evidence, not just a readable probe.
Recall can fail under reversal
One reported GPT-4 comparison answered “Who is Tom Cruise’s mother?” at 79% but the reversed relation at 33%. The association was stored without the symmetric relation a person would usually infer.
Reasoning traces can rationalize
Biasing cues shifted answers by up to 36 points on BIG-Bench Hard while the chain of thought often omitted the cue. A fluent explanation can be theater rather than a faithful account.
The depth-before-action protocol
Six gates turn “look below the water” into an operational check. The agent may proceed only when the gate’s evidence is visible; otherwise it retrieves, tests, asks, or abstains.
Fix what the concept means here.
Words do not carry a single context-free meaning. Bender and Koller define understanding as recovering communicative intent from expression in context; naming the category is only form.
- What problem is being solved?
- Whose intent matters?
- Which meaning is in play?
- What is still ambiguous?
Anchor foundational terms outside the model.
Grounding is a gradient. The foundational concepts should connect to observation, retrieved sources, tools, or environment feedback; definitions made only of more definitions leave the system floating.
- Which claims are observed?
- Which are retrieved?
- Which are assumptions?
- How fresh is the evidence?
Build a causal and relational model.
Move beyond “what is it?” to what it implies, rules out, depends on, and changes. A concept becomes useful when it can support prediction and intervention.
- What causes what?
- What follows if this is true?
- What would contradict it?
- Where does it stop applying?
Try to break the understanding.
Transfer is the strongest practical separator between recall and understanding. Reverse the relation, alter a premise, introduce a novel case, or change the environment’s rule.
- Does the reverse relation hold?
- What changes under a counterfactual?
- Can it handle an unseen case?
- What is the strongest counterexample?
Check against something the agent did not generate.
Reflection helps when the critic has an external signal: a unit test, tool result, retrieved source, sensor, or environment score. Pure self-judgment can compound the same error.
- What tool can test this?
- What evidence could falsify it?
- Do logs match the self-report?
- Is the source itself gameable?
Choose action, escalation, or abstention.
Consequential work needs a stopping rule. If the depth packet is incomplete, the correct output is not a better-sounding guess; it is a targeted retrieval, a question, an escalation, or no action.
- What is the evidence-backed action?
- What is the downside if wrong?
- What would change the decision?
- Who owns the final call?
Measure depth where models separate
Bloom’s revised taxonomy gives the evaluation ladder. A 2026 study of five frontier models found no significant differences at Remembering/Understanding; separation appeared at Applying/Analyzing and Evaluating/Creating. Lower-level fluency is a weak test.
Remember
Name it. Recall facts. Repeat a definition.
Understand
Summarize it. Compare it. Explain it in simpler words.
Apply
Use it correctly in a case the agent has not seen.
Analyze
Separate causes, assumptions, boundaries, and failure modes.
Evaluate
Judge competing interpretations against evidence and consequences.
Create
Build a new solution, then test whether the concept still holds.
Sources: Ohio State on revised Bloom ↗ · PLOS One evaluation ↗
Three ways the tip fools us
Each failure looked competent on the surface. Each failed for a different reason below it: shortcut features, stale reality, or genre-perfect fabrication.
The wolf was snow
A high-performing classifier learned snow rather than “wolf,” mislabeling a husky in snow. Explanation tooling exposed the shortcut that validation accuracy hid.
The paid invoice was “late”
An invoice agent chased a customer who had already paid because its input was stale and the system’s keywords did not represent “already paid.” Classification stayed inside the rules; reality moved.
The perfect-sounding answer
Asked about Xenoposeidon, a model produced realistic names, journals, dates, and an expedition — with every specific wrong. It understood the shape of an answer, not the facts.
Build the check outside the model
The architecture should not rely on a model announcing that it understands. It should make evidence travel with the decision and keep hard boundaries outside the reasoning loop.
Trigger a depth scan
Unfamiliar concept, conflicting evidence, changed rule, irreversible action, or high downside.
Expose the model
Context, anchors, causal claims, boundaries, counterfactuals, uncertainty, and proposed action.
Demand contact
Retrieval, tests, tools, environment feedback, logs, human review, or a hard policy gate.
Use reasoning traces as working notes, not proof.
Chain-of-thought, Socratic questions, and reflection can elicit depth. They remain auditable hypotheses until an external signal checks them.
Test premise-sensitivity.
Change a premise and require the conclusion to change appropriately. If it does not, the agent may be replaying a pattern rather than using a model.
Keep safety boundaries deterministic.
Permissions, destructive-operation locks, and approval gates should not depend on the same model that is trying to reason past them.
Score justifications and outcomes.
Record the evidence used, what would have falsified the claim, the action taken, and what the environment later showed.
What remains unresolved
There is no consensus solution to symbol grounding. Othello-style world models are compelling but strongest in constrained synthetic domains. Chain-of-thought faithfulness remains unsolved. External signals can also be stale, biased, or gameable. The claim here is narrower: these checks are the best-supported ways to make shallow understanding harder to pass as deep understanding.
Primary and directly verified sources
Selected supporting sources
- Harnad on hybrid grounding ↗
- Vincent-Lamarre et al., dictionary MinSet ↗
- Mahowald et al., formal vs. functional competence ↗
- Berglund et al., Reversal Curse ↗
- Turpin et al., unfaithful reasoning traces ↗
- Jin et al., Socratic questioning ↗
- SoDa, multi-agent Socratic debate ↗
- Bloom’s taxonomy in the AI era ↗
- PLOS One, Bloom-based LLM evaluation ↗
Evidence note · Two primary sources were read directly during the 20 Sep 2026 research session. Other findings were drawn from indexed abstracts and secondary summaries; this distinction is preserved above. No direct published benchmark for a Feynman-style “teach it simply” LLM test was found.
