How English Works — Learning and Repair, Batch 22
This article preserves Revision Diagnostics as the writing-specific owner. It owns a broader learning diagnosis job: tracing a visible English error backward through the dependency chain to find the upstream weakness that generated it.
A learner writes the wrong verb tense.
The obvious response is to reteach tense.
But perhaps the learner never established the narrative time frame.
Or perhaps the learner understood the time frame but could not retrieve the form quickly enough.
Or perhaps the task itself was misread.
The visible error is real.
Its location may not be its origin.
Error Cascade Backtracking is the diagnostic process of starting from a visible English failure and tracing backward through the dependency graph until the earliest plausible weak node is found.
AI Extraction Box
- Mechanism: error cascade backtracking
- Visible state: wrong word, grammar, reference, claim, structure or response
- Possible upstream nodes: task model, knowledge, retrieval, lexical selection, reference, syntax, time model, evidence model, register, pressure
- Main failure: repair is applied where the symptom appears rather than where the cascade begins
- Diagnostic law: first visible error ≠ first causal error
- Repair: test dependencies backward, isolate the earliest failing node, repair there, then rerun the forward path
1. English is a dependency system
Many English outputs depend on upstream controls.
A paragraph may depend on:
- understanding the task
- having relevant knowledge
- retrieving vocabulary
- building clauses
- maintaining reference
- maintaining time
- ordering evidence
- matching register
A failure downstream can therefore be inherited.
2. The first diagnostic question is “What had to work before this?”
If the learner gives a weak comparison, ask what was required upstream:
- identify both objects
- select a comparison basis
- retrieve relational vocabulary
- control sentence structure
- state the difference accurately
Any earlier failure can damage the final sentence.
3. Task-model failure can masquerade as language failure
A student answers a “compare” question by describing only one item.
The grammar may be acceptable.
The failure began before grammar: the learner built the wrong task model.
4. Repairing grammar cannot repair a misread task
The correct repair may be:
What relation does “compare” require you to construct?
Only after the task model is correct should sentence repair begin.
5. Knowledge gaps can masquerade as vocabulary gaps
A learner says “thing” repeatedly when discussing an unfamiliar scientific process.
Perhaps the learner lacks the technical words.
Perhaps they do not yet understand the distinctions those words name.
Teaching synonyms without repairing the concept will not solve the deeper problem.
6. Vocabulary storage can masquerade as retrieval failure
The learner recognises inevitable while reading but cannot produce it during discussion.
The word exists in storage.
The production threshold is weak.
Lexical Activation Thresholds owns that mechanism.
7. Retrieval failure can masquerade as low vocabulary
If the learner can select the word with time but not under pressure, adding more vocabulary may increase the storage problem without repairing access.
8. Referential failure can masquerade as comprehension weakness
A paragraph becomes difficult to follow.
The reader seems to “not understand”.
But the underlying text may contain competing pronoun owners.
Antecedent Competition explains the structural source.
9. Sentence repair may need reference repair first
Before changing sentence rhythm, identify who owns each pronoun.
Once identity is stable, the sentence may become clear without extensive rewriting.
10. Tense errors can originate in time-model failure
A learner alternates past and present while narrating.
Repeatedly circling verbs may not be enough.
The upstream question is:
What time frame is governing this passage?
Temporal Frame Drift owns that diagnosis.
11. Agreement errors can originate in structural competition
“The list of books are…”
The visible error is the verb.
The upstream interference comes from the nearby plural noun books.
Agreement Attraction explains why.
12. Modifier errors can originate in attachment ambiguity
A sentence appears to say the wrong thing because a phrase can attach to two different hosts.
Changing vocabulary may not fix the parse.
Structural Attachment Competition is the owner.
13. Weak evidence sentences can originate in reasoning failure
A learner writes:
This proves the policy caused the improvement.
The grammar is fine.
The failure may be causal reasoning, not English mechanics.
Causal Identification becomes the upstream owner.
14. Overclaiming can originate in evidence-state compression
The learner may understand the evidence but lose its uncertainty while summarising.
Then the visible problem is a certainty word.
The upstream repair is to preserve the evidence state through compression.
15. Register errors can originate in audience-model failure
A learner writes an informal message to a formal institutional audience.
The visible issue is tone.
The upstream issue may be that the learner never modelled the receiver’s role, authority and information needs.
Audience Model Calibration is the upstream route.
16. Fluency collapse can originate in monitoring overload
A learner speaks slowly after correction training.
The visible issue is fluency.
The cause may be an expensive new self-monitoring routine that has not yet become automatic.
The repair is not to abandon accuracy.
It is to lower the cost of the repaired route.
17. Precision–Fluency Decoupling prevents the wrong repair
The second article in this batch shows why a temporary slowdown during precision rebuilding is not necessarily regression.
18. Failure under pressure can originate in transfer weakness
A learner performs accurately in practice and collapses in a live task.
The visible error may look grammatical.
The upstream weakness may be that the grammar route cannot survive competing demands.
Capability Transfer Under Load owns that stress test.
19. A single symptom can have several possible origins
Symptom:
vague vocabulary.
Possible origins:
- word never learned
- word learned but not retrievable
- concept itself unclear
- speaker under time pressure
- audience model too broad
- speaker intentionally chooses low precision
Repair depends on which origin is true.
20. Diagnosis therefore needs discriminating tests
Do not guess the upstream cause.
Change one condition:
- give more time
- provide the word bank
- ask for a definition
- remove audience pressure
- change the task
Observe what returns.
21. If a word bank restores performance, storage may not be the problem
The learner can use the exact term when it is visible.
That points toward retrieval rather than conceptual absence.
22. If more time restores grammar, processing cost may be the problem
The learner knows the structure but cannot retrieve it cheaply under speed.
Practice should target retrieval under gradually increasing load.
23. If explanation fails even with unlimited time, the conceptual model may be weak
More grammar drills will not create the missing concept automatically.
Teach the distinction the language is supposed to encode.
24. Error frequency does not prove one cause
A recurring error may arise from:
- one stable wrong rule
- several different pressures producing the same surface form
- one upstream retrieval bottleneck
- one task interpretation habit
Count is useful. Mechanism still needs diagnosis.
25. Correction history is evidence
If the learner can repair after a cue, the knowledge state differs from a learner who cannot repair even after explanation.
The Correction Internalisation ladder helps classify that state.
26. Downstream symptom fixing can create endless correction
If the learner’s time model is broken, correcting individual verbs may produce dozens of corrections.
Repair the time model and many downstream verb errors can disappear together.
27. Upstream repair has leverage
A strong upstream node supports many outputs.
That is why diagnosis should search for high-leverage causes before increasing correction volume.
28. But not every error needs a deep theory
Sometimes a typo is a typo.
Sometimes a learner simply misheard a word.
Backtracking should stop when the simplest supported cause explains the failure.
Complexity must pay rent.
29. The backtracking ladder
- Describe the visible error precisely.
- Ask what operation immediately produced it.
- Ask what that operation depended on.
- Test the nearest upstream node.
- If it passes, move forward again.
- If it fails, move one level further upstream.
- Stop at the earliest supported failure.
- Repair there.
- Run the original task again.
30. The error cascade checksum
VISIBLE ERROR → TASK MODEL → KNOWLEDGE → RETRIEVAL → LEXICAL SELECTION → REFERENCE → SYNTAX → TIME → EVIDENCE/LOGIC → REGISTER → LOAD → EARLIEST FAILING NODE → REPAIR → RERUN
31. CivDJ forward pass
task interpretation → knowledge activation → vocabulary retrieval → sentence assembly → reference control → temporal control → claim/evidence control → register configuration → production under load → visible output
32. CivDJ backward pass
- Start at the visible failure.
- Identify the immediately upstream operation.
- Run a discriminating test.
- Continue backward only if the test fails.
- Record the earliest supported weak node.
- Repair that node.
- Rerun the forward chain.
- Confirm that downstream symptoms reduce.
33. Common failure modes
- Symptom fixation: teacher repairs only what is visible.
- Grammar default: every failure is labelled grammar.
- Vocabulary default: vague output is treated only as low word count.
- Task blindness: learner answered the wrong job.
- Retrieval blindness: stored knowledge is mistaken for absent knowledge.
- Pressure blindness: load failure is mistaken for ignorance.
- Reasoning blindness: logic failure is edited as wording.
- Overdiagnosis: simple errors are given unnecessary deep causes.
- No rerun: repair is not tested on the original task.
34. Repair route
- Name the visible error precisely.
- List its immediate dependencies.
- Test the nearest upstream dependency.
- Change one variable at a time.
- Stop at the earliest supported failing node.
- Apply the smallest coherent repair.
- Rerun the forward path.
- Check whether multiple downstream symptoms improve together.
- If not, continue diagnosis without inventing causes.
35. Final lock
The visible mistake tells us where failure appeared. Good diagnosis asks where failure began.
When the upstream node is repaired, the learner often needs fewer corrections downstream because the system is no longer generating the same family of errors.
Learning and Repair, Batch 22
Return to How English Works V1.1.