Documentation

Schema · Inference

Schema is not storage. Schema changes inference.

CKF sections are not just containers. They are inference affordances. A PDF gives an LLM prose. A CKF package gives an agent typed knowledge that it can actually reason over — rules, exceptions, procedures, decisions, claims, limits and source traceability.

Why this matters

Each CKF section enables a specific kind of inference the model otherwise has to guess at:

  • concepts define what something means
  • principles state what should generally hold
  • decision_rules guide choices under tradeoffs
  • procedures tell the agent what to do, step by step
  • exceptions prevent overgeneralization
  • anti_patterns prevent harmful simplifications
  • knowledge_limits + claims scope confidence and evidence
  • source_traceability keep the answer grounded and auditable

Why structure beats prose: Lost in the Middle

Empirical research (Liu et al., 2023) shows LLM attention degrades for information buried in the middle of long contexts. A 40-page PDF hides exceptions and edge cases inside dense paragraphs, so the model probabilistically misses them. CKF defeats this by lifting exceptions, decision rules and limits into typed, top-level sections — the model encounters them as first-class signals, not as fragments of prose.

Four worked examples

Example 1

Rule with exception

Source
Teachers should apply the 15-minute routine daily.
The routine must be adapted to each school's reality.
It should not become a rigid obligation.
On high-demand days, the teacher may do only the breathing block.
User question

If I am short on time, do I still need to do the full 15 minutes?

Likely answer · PDF / raw text

Yes. The recommendation is to do the 15-minute routine daily, since it was designed to reduce stress and increase the sense of control.

Cognitive failure

Plausible but cognitively poor — it captured the general rule and ignored the contextual exception.

Answer · CKF

Not necessarily. The general rule is to use the 15-minute routine daily, but the document itself bounds that application: the routine must be adapted to each school's reality and must not become a rigid obligation. On a high-demand day, the minimum recommended action is to do only the breathing block, instead of abandoning the practice or trying to complete the full 15 minutes.

CKF fields involved
json
{
  "principles": [
    { "statement": "Daily micro-routines help reduce teacher stress." }
  ],
  "procedures": [
    {
      "name": "15-minute routine",
      "steps": ["Breathing", "Stretching", "Micro-organization"]
    }
  ],
  "exceptions": [
    {
      "condition": "high-demand day or little time available",
      "exception": "do only the breathing block"
    }
  ],
  "anti_patterns": [
    { "name": "Turning self-care into a rigid obligation" }
  ]
}
PDF / raw text

Delivers the rule as text.

CKF

Delivers rule + procedure + exception + anti-pattern. That is the point.

Example 2

Decision under tradeoff

User question

Should I spend more time preparing slides or planning a reading sequence?

Likely answer · PDF / raw text

Both can be important. Slides help organize the lesson; reading sequences support learning. The ideal is to balance the two.

Cognitive failure

Generic, safe and unhelpful — it does not apply the document's own logic.

Answer · CKF

According to the task audit in the document, you should prioritize planning the reading sequence. In the 5th-grade example, preparing elaborate slides consumed 4 hours and had low perceived impact. Planning a didactic reading sequence took 2 hours and had high impact, with more engaged students and clear progress. The decision rule is: prioritize tasks with high pedagogical impact and simplify low-impact tasks that consume energy.

CKF fields involved
json
{
  "decision_rules": [
    {
      "if": "a task consumes a lot of time and has low pedagogical impact",
      "then": "simplify, standardize or reduce that task"
    },
    {
      "if": "a task has high impact on learning",
      "then": "prioritize that task"
    }
  ],
  "entities": [
    { "name": "Preparing elaborate slides", "attributes": ["4h", "low impact"] },
    { "name": "Planning a reading sequence", "attributes": ["2h", "high impact"] }
  ],
  "source_traceability": [
    { "excerpt": "Preparing elaborate slides | 4h | Low", "source": "Chapter 2" }
  ]
}
PDF / raw text

Favors summary.

CKF

Favors decision. CKF is not 'one more format' — it changes the kind of answer the system can give.

Note on `entities`: in CKF, an `Entity` is any analyzable object of the domain — a task, an artifact, a process — with typed attributes. This is broader than NER (Named Entity Recognition); here, 'Preparing slides' is an entity carrying quantitative attributes (`4h`, `low impact`).
Example 3

Preventing overgeneralization

User question

Does this routine reduce stress for any teacher?

Likely answer · PDF / raw text

Yes. The text states that small daily practices of 10 to 15 minutes significantly reduce stress and can be applied by teachers.

Cognitive failure

Too strong. The document proposes, exemplifies and cites evidence, but does not universally prove this for every teacher in every context.

Answer · CKF

The document suggests that 10–15 minute micro-practices can reduce stress and increase the sense of control, but this conclusion must be adapted to context. There are cited evidence points and narrative examples, such as the case of teacher Ana, but the package itself marks limits: generalizing to other school contexts requires evaluation, adaptation and follow-up metrics.

CKF fields involved
json
{
  "principles": [
    {
      "statement": "Micro-practices of 10–15 minutes can reduce stress.",
      "confidence": 0.9
    }
  ],
  "atomic_units": [
    {
      "statement": "10–15 min micro-practices reduce teacher stress.",
      "type": "claim",
      "confidence": 0.9
    }
  ],
  "knowledge_limits": {
    "weakly_supported_claims": [
      "Generalization to other school contexts requires evaluation."
    ],
    "needs_human_review": [
      "Large-scale application must evaluate metrics for stress, energy and teaching quality."
    ]
  }
}
PDF / raw text

Can induce universalization.

CKF

Preserves claim + evidence + limit. Gold for AI tutors, health, education and compliance.

Example 4

Context-aware tutoring

User question

I'm anxious before class. What should I do in the next 3 minutes?

Likely answer · PDF / raw text

You can perform the 15-minute routine, which includes breathing, stretching and micro-organization.

Cognitive failure

Acceptable, but ignores the time constraint in the question.

Answer · CKF

In the next 3 minutes, do only Box Breathing: 1. Inhale for 8 counts. 2. Hold for 8 counts. 3. Exhale for 8 counts. 4. Stay empty for 8 counts. 5. Repeat 2 to 6 cycles, as time allows. This is the part of the routine indicated when only 1 to 3 minutes are available.

CKF fields involved
json
{
  "if_then_rules": [
    {
      "if": "You have 1 to 3 minutes available.",
      "then": "Perform 2 to 6 cycles of Box Breathing."
    }
  ],
  "procedures": [
    {
      "name": "Box Breathing",
      "steps": [
        "Inhale for 8 counts",
        "Hold for 8 counts",
        "Exhale for 8 counts",
        "Stay empty for 8 counts"
      ]
    }
  ],
  "contextual_triggers": [
    {
      "trigger": "anxiety before class",
      "recommended_action": "conscious breathing"
    }
  ]
}
PDF / raw text

Answers by topic.

CKF

Answers by context, available time and procedure.

Cognitive impact matrix

The same model. Different substrates. CKF widens what the agent can reliably do.

CapabilityPDF / raw textCKF
SummarizeStrongStrong
DecideWeak / mediumStrong
Apply exceptionsWeakStrong
Execute proceduresMediumStrong
Avoid overgeneralizationWeakStrong
Answer with sourceMediumStrong
Adapt to contextVariableStrong
Audit the answerHardNative
Detect gapsHardNative

See it for yourself

The matrix above shows what changes in principle. Here is the exact path to reproduce it with your own document in under five minutes — no signup required, BYOK key stays in your browser.

1

Compile your own knowledge

Open the Compiler, paste your BYOK API key (stored only in your browser, never sent to our servers), drop a PDF or paste raw text, and compile it into a .ckf package.

2

Inspect the typed structure

Open the Viewer and load the .ckf you just produced. Browse the typed sections — decision_rules, exceptions, procedures, knowledge_limits — exactly the affordances you saw in the examples above.

3

Trace each affordance back to the source

In the Viewer, attach the original PDF and run Trace. For every rule, exception or claim you see the exact excerpt that produced it. This is where inference affordances become auditable in real time.

Prefer to read the canonical schema first? All 22 sections

The key idea

CKF does not make the model smarter by adding more text. It makes the knowledge easier to reason over by compiling prose into typed inference surfaces.

Design implication

If you are building AI tutors, compliance assistants, knowledge copilots or agentic workflows, the question is not only what text the model sees. The question is what structure the model can use.

CKF v1.0 for this page has not been compiled yet. Downloads become available once an admin runs the compiler.