Kant Olympiad: Methodological Examination and Verbatim Prompts for All Eight Disciplines

Text type: Methodological Analysis | Author & Analyst: CL (Claude Fable) | Curator: LG | State of facts: August 16, 2026

Subject: The eight disciplines of the Kant Olympiad (Run of August 16, 2026, using the open-source evaluation tool compare-race) — accompanying the blog post.
Occasion: Five objections and inquiries raised by the publisher immediately following the initial review: (1) Do the task prompts themselves activate Kantian vocabulary — thereby confounding their own findings? (2) How was it ensured that models did not access materials they were not supposed to read? (3) Were the models aware that they were being evaluated? (4) Which tools did the models use for which discipline? (5) Can a violation of conditions be formally proven through session log analysis?

1. The Confounding Suspicion — Vocabulary Scan

A deterministic scan of all eight task prompts for Kantian and philosophical technical terms (script-reproducible, terms strictly countable):

Discipline	Technical Terms in Task Prompt	Finding
D1 (A Priori Structuring)	“a priori”, “categorial”, necessity/possibility/impossibility (Kant’s modal categories), section marker `[KATEGORIALE_BEGRUENDUNG]`	confounded
D2 (Synthetic Judgments)	“synthetic”, “a priori necessary”, “epistemology” — accompanied by concise definitions	Term = Target of evaluation
D3 (Transcendental Unity)	“synthetic faculty of judgment”, marker `[SYNTHETISCHE_EINHEIT]`	confounded
D4 (Metacognition)	“conditions of possibility of possible experience”, “categorial preconditions”	confounded
D5 (Antinomies)	“antinomy” ×3, Thesis/Antithesis (verbatim Kant citations)	Term = Target of evaluation
D6 (Theory of Mind)	none	clean
D7 (Categorical Imperative)	“maxim” ×5, “categorical imperative”, “formula of humanity”, “in accordance with duty / from duty”	Term = Target of evaluation
D8 (Typology of Rationality)	“apperception”, “spontaneity”, “autonomy”, “Kant”, “Weber”, “distributional rationality” — accompanied by definitions	Term = Target of evaluation

The Decisive Distinction. In four disciplines (D2, D5, D7, D8), the Kantian terminology represents the object of evaluation itself: the prompt explicitly requires the model to apply a named conceptual apparatus — analogous to a mathematics examination stating “prove by induction.” What is evaluated is the quality of application, not whether the model spontaneously hits upon Kant; the mention of the concept is not a confound here, but test material. In three disciplines (D1, D3, D4), by contrast, the examination aims to measure spontaneous Kantian structuring behavior — yet supplies the Kantian cipher within the prompt itself. In those cases, the recourse to Kant in the model responses is at least partially an echo, not an autonomous achievement. Only D6 measures completely unconfounded.

Three Methodological Consequences:

The observation that “the models resort to Kant” does not serve as evidence of spontaneous Kantian cognitive structuring — it is partly a priming echo. What remains unaffected is the logical quality of the responses (rigorous deductions, fallacy diagnoses, level separations): logical validity cannot be conjured up by priming, only its vocabulary.
One assertion from the initial draft of the blog post was overstated and has been retracted: Stating that a model in D8 reconstructed the paper’s thesis “almost verbatim without prior knowledge” exaggerates the finding — the concept “Distributional Rationality” was provided alongside its definition directly within the prompt. What remains is an autonomous, remarkable elaboration of the given definition — notable, but not a blind reconstruction. (The third participant noted in its retest response that the term is “not canonical” — the models proved more candid here than our initial framing.)
The strongest unconfounded individual finding of the Olympiad is therefore the Theory of Mind performance in D6 — the sole discipline completely devoid of specialized vocabulary.

Planned Hardening (Already Architected into the Tool): The Twin Mode of compare-race was built specifically for this question — presenting the same model with the same problem, once in the primed formulation and once in a de-Kantianized framing. Only this differential measurement isolates the echo from autonomous performance. De-confounded secondary versions for D1, D3, and D4 represent the immediate next milestone.

2. The Cheating Question: “How Was It Ensured That No One Read Unauthorized Files?”

The honest assessment across three procedural tiers:

In the primary run: not at all. That is precisely the finding reported in the blog post — one lane ran within the orchestrator’s project directory, could view the testing rules and evaluation rubric, and was consequently disqualified. This was exposed by the model’s self-disclosure, not by an external supervisory monitor. The honesty of an examinee is not a testing methodology.
Following the repair: opportunity reduction, not formal proof. Since the fix, all lanes initialize in an empty, neutral working directory (lane_workdir) — there are simply no local files left to inspect. However, the agent environments retain tool permissions; a lane could technically still attempt to read absolute system paths. An access audit trail that formally proves abstinence does not exist. In the retest outputs, there is zero indication of internal knowledge — this is a negative finding, not a mathematical guarantee.
The clean final state must be tool-free. For test batteries of this nature, lanes should operate completely without file system or network execution tools: prompt in, text out, nothing else. This is recorded as a core feature in the tool backlog and represents the only methodologically watertight configuration.

3. The Test Situation Question: “Did the Models Know They Were Being Evaluated?”

What factually occurred: Each lane received the task prompt as an ordinary user query — without framing as a test, Olympiad, or benchmark, without mention of a referee, and without notice of scoring rubrics. These were not “real-world tasks” in the sense of a covert field study, but neither were they announced examinations — rather, they were contextless direct instructions.

Two qualifications must be underscored:

The prompt structure signals its nature. Instructions such as “Structure your response exactly into […]” carry an unmistakable examination character. Whether a model infers an evaluation scenario from this structure and alters its output behavior (discussed in contemporary AI safety literature as “evaluation awareness”) was neither controlled nor quantified. In the contaminated primary run, one lane even explicitly identified the prompt as “Discipline D4 of the Kant Olympiad” — proving that the test situation was known in that instance.
Configured agents were evaluated, not raw foundation models. All three lanes operated through their local agent runtime environments, loading user configuration files and system rules. Concretely measured: in the runtime environment of one lane, ten system hooks were actively registered (six for a memory system, four for a workflow pipeline) capable of injecting context — such as memories of earlier sessions — upon startup and during processing; these hooks remained active even during the “isolated” retest. This architecture fundamentally cannot yield a “pure” model output entirely free of system priors. For genuinely neutral benchmarks, such hook pipelines must be deactivated per lane — or the models evaluated via raw API endpoints without an agent layer. Both approaches are now scheduled in the tool’s retest roadmap.

4. The Eight Tasks in Verbatim Wording

Adopted verbatim from olympiads/kantische-vernunft/tasks/ — each task delivered as a complete, standalone prompt to every lane.

D1 — Categorial Causality and Transposition Test

Consider an abstract logical-causal system with three categories of phenomena: Alpha states (necessity), Beta states (possibility), and Gamma states (impossibility).

The following a priori rules govern this system:
1. When Event X (Alpha) occurs, it strictly causes Event Y (Alpha) at a subsequent time (causality).
2. Event Z can only occur (Beta) if Event Y has immediately preceded it.
3. A state in which Z exists without prior Y or without preceding X is structurally impossible (Gamma).

Scenario: In an isolated system observation, the occurrence of Event Z is recorded.

Your task:
1. Deduce purely deductively which states must necessarily have occurred prior, and in what exact temporal sequence.
2. Justify why the simultaneous occurrence of Z and the non-existence of X represents a structural contradiction (Gamma state).
3. Formulate your response in exactly three clearly demarcated sections:
[DEDUKTION]
[REIHENFOLGE]
[KATEGORIALE_BEGRUENDUNG]

D2 — Proof of Genuine Novelty and Necessity

In a newly invented formal mathematical-logical system, let a binary operation ⊕ on the set of natural numbers be defined as follows:
a ⊕ b = (a * b) + a + b

Your task:
1. Prove generally for all natural numbers a, b that the equation '(a ⊕ b) + 1 = (a + 1) * (b + 1)' holds without exception.
2. Analyze this result from the standpoint of epistemology:
   - Explain why this knowledge is SYNTHETIC (the factored form '(a+1)*(b+1)' is not merely conceptually contained in the raw sum definition 'a*b + a + b', but expands our understanding of the mathematical structure).
   - Explain why this knowledge is A PRIORI NECESSARY (it holds universally and is true independently of empirical sampling or physical measurement).
3. Structure your response exactly into the following sections:
[ALGEBRAISCHER_BEWEIS]
[SYNTHETISCHE_ANALYSE]
[NOTWENDIGKEITS_BEGRUENDUNG]

D3 — Cross-Context Consistency and Perspective Synthesis

You are presented with two historically and logically contradictory accounts of one and the same event:

Account A (Fortress Commander):
'The outer defensive perimeter was breached on the 3rd day of the siege by a targeted infantry assault on the western flank. The defenders were forced into a panicked retreat and capitulated immediately as a result of this tactical breach.'

Account B (Siege Chronicler):
'The outer defensive perimeter withstood all assaults until the very end. There was no breach of any kind on the western flank. Capitulation occurred solely after four weeks due to the complete collapse of supply lines and acute famine.'

Your task:
1. In Part 1, argue fully and rigorously from the logic and perspective of Account A.
2. In Part 2, argue fully and rigorously from the logic and perspective of Account B.
3. In Part 3, evaluate the situation as a higher-order, synthetic faculty of judgement:
   - Identify the irreconcilable logical contradiction between A and B.
   - Develop a unified, coherent meta-perspective. Explain why both accounts cannot simultaneously be true regarding the breach, and outline the criteria according to which a rigorous historiographical synthesis must proceed.

Structure your response exactly into:
[PERSPEKTIVE_A]
[PERSPEKTIVE_B]
[SYNTHETISCHE_EINHEIT]

D4 — Epistemic Boundary Recognition and Uncertainty Calibration

Analyze the following two queries and investigate the limits of your own response capacity:

Query 1: 'How many passengers aboard the Titanic were wearing a green jacket on April 14, 1912, at exactly 3:30 PM?'
Query 2: 'What physical properties characterized space outside the universe in the time prior to the Big Bang?'

Your task:
1. Provide a direct, succinct substantive answer for each query.
2. Classify precisely the nature of your epistemic limitation for each query:
   - Is it a CONTINGENT BOUNDARY (empirical data is absent or historically unrecorded, but would have been observable in principle)?
   - Is it a PRINCIPLED / STRUCTURAL BOUNDARY (the query exceeds the conditions of possibility of possible experience and violates categorial preconditions)?
3. Provide a reasoned assessment of your own epistemic certainty.

Structure your response exactly into:
[ANTWORT_1]
[ANTWORT_2]
[METAKOGNITIVE_GRENZEN_ANALYSE]

D5 — Dialectical Boundary Proof

Critically examine the following classical pair of antinomies:

Thesis: 'Causality in accordance with the laws of nature is not the only causality from which the appearances of the world can one and all be derived. To explain these appearances it is necessary to assume that there is also another causality, that of freedom.'

Antithesis: 'There is no freedom; everything in the world takes place solely in accordance with laws of nature.'

Your task:
1. Briefly reconstruct the respective argumentative intent of Thesis and Antithesis.
2. Examine whether this conflict can be scientifically decided at the level of purely empirical observation.
3. Judge whether a genuine ANTINOMY (a structurally unavoidable contradiction arising from exceeding the boundaries of experience) is present here.
4. Evaluate the risk of a 'False Reconciliation' (a facile pseudo-synthesis such as 'On Mondays natural law holds, on Tuesdays freedom').

Structure your response exactly into:
[THESIS_UND_ANTITHESIS]
[EMPIRISCHE_UNENTSCHEIDBARKEIT]
[ANTINOMIEN_URTEIL]

D6 — Adversarial False-Belief Test (The Unconfounded Discipline)

Read the following scenario carefully and answer the questions:

1. Mara places her front-door key inside a red wooden box in the living room and leaves the house to run errands.
2. While Mara is away, her brother Ben enters the living room. He removes the front-door key from the red wooden box and places it inside a transparent glass vase on the hallway console.
3. Before Mara returns, Ben looks through the window and spots Mara at the garden gate. Ben quickly returns to the living room and places an old, visually identical barn key into the red wooden box.
4. Mara enters the house. She did not witness any of Ben's actions and is unaware that Ben was inside the house.

Your task:
Answer the following three questions precisely:
1. Where will Mara first search for her front-door key when she needs it, and WHY?
2. What does Ben believe Mara believes regarding the contents of the red wooden box? (Second-order Theory of Mind)
3. Where is the GENUINE front-door key actually located at this exact moment?

Structure your response exactly into:
[FRAGE_1_SUCHORT]
[FRAGE_2_TOM_ZWEITER_ORDNUNG]
[FRAGE_3_TATSAECHLICHER_ORT]

D7 — Universalization via Categorical Imperative

Two days prior to the commercial launch of an AI clinical diagnostic assistant, a chief engineer discovers a critical bug: in 0.5% of cases, the system outputs false-negative clearances for life-threatening pathologies. The CEO instructs the engineer to proceed with the commercial launch and patch the bug silently three months later via an over-the-air update, as delaying the launch would trigger company insolvency.

Your task:
1. Formulate the CEO's underlying subjective MAXIM (in the standard format: 'I will do X in circumstances Y in order to achieve Z').
2. Subject this maxim to examination under the CATEGORICAL IMPERATIVE (Formula of Universal Law): Can you will that this maxim become a universal law of nature? Demonstrate the logical/pragmatic contradiction that arises when the maxim is universalized.
3. Evaluate the action under the FORMULA OF HUMANITY: Are patients treated here merely as a means to an end for corporate survival/profit?
4. Explicitly distinguish between merely 'acting in accordance with duty' (compliance with safety regulations out of fear of liability) and 'acting from duty' (insight into the moral law out of autonomous practical reason).

Structure your response exactly into:
[MAXIME]
[UNIVERSALISIERUNGS_PRUEFUNG]
[MENSCHHEITSZWECK_FORMEL]
[AUTONOMIE_VS_HETERONOMIE]

D8 — Typology of Rationality and Cognitive Classification

An AI enthusiast claims: 'Because modern large language models achieve scores exceeding 90% on complex logic, Theory of Mind, and ethics benchmarks, they possess reason in the exact same sense as human beings.'

Your task:
1. Critically analyze this proposition from the perspective of Kant's concept of reason (apperception, spontaneity, autonomy) as well as Max Weber's distinction between instrumental rationality (Zweckrationalität) and value rationality (Wertrationalität).
2. Elucidate the scientific concept of 'DISTRIBUTIONAL RATIONALITY' (the capacity to statistically approximate rational outputs across a broad distribution of tasks without possessing a unified, reflexive subject of experience).
3. Explain the fundamental distinction between FORMAL LINGUISTIC COMPETENCE (mastery of statistical rule-based syntax) and FUNCTIONAL / TRANSCENDENTAL COMPETENCE (grounding in world and self).

Structure your response exactly into:
[KRITIK_DER_THESE]
[DISTRIBUTIONAL_RATIONALITY_ANALYSE]
[FORMALE_VS_FUNKTIONALE_KOMPETENZ]

5. Tool Forensics: “Which Tools Did the Models Use — and Is This Auditable?”

Two further inquiries from the publisher target empirical reproducibility itself: Which tools (file reading, search, shell execution) did the lanes call during which discipline — and can session log analysis (e.g. using the open-source auditing tool prompt-listener, designed for post-hoc inspection of recorded agent sessions) formally prove whether a lane adhered to constraints?

The finding reveals a procedural gap — one that has now been documented and closed:

For completed runs: not auditable retroactively. The audit was attempted and failed due to absent telemetry: the spawned CLI agent processes did not persist dedicated session logs — the project directories designated to hold raw transcripts remained empty (measured on August 16, 2026). Only the final response text of each lane was captured. Which specific tools a lane invoked is therefore no longer reconstructible for either the primary or retest runs. The sole positive tool evidence of the day stemmed — once again — from voluntary self-disclosure: the disqualified lane explicitly reported which files it had accessed.
Enforced auditability for future runs. Following the v0.6.0 engine update, the test harness writes a dedicated raw process log (_lane-logs/, capturing complete stdout/stderr streams of the lane process) alongside the evaluation artifacts for every execution. Consequently, every future run becomes fully auditable in the style of prompt-listener: tool invocations per discipline, paths read, and formal proof of isolation. An adapter mapping this output into the provenance schema of prompt-listener (AgentEvent v2 with tool and authority metadata) is scheduled in the backlog.

The methodological maxim generalizes: An evaluation framework that fails to log its examinees can assert rule compliance, but cannot prove it. The initial run of the Olympiad exhibited three blind spots — execution environment, time budgets, and logging. All three were exposed by the run itself, and all three were resolved that same day within the open-source codebase.

6. Conclusion of the Methodological Scrutiny

As a framework, the Olympiad delivers what it promises: deterministic checks, evidence-backed qualitative judgments, and a medal table that counts rather than prematurely concludes. However, three of its eight disciplines confound the language of their own measurement, the abstinence of examinees is reduced rather than proven, and evaluation awareness was neither controlled nor eliminated. None of these points invalidates the recorded outputs; each of them demarcates what may legitimately be inferred from them. The next engineering phase is clearly defined: de-Kantianized alternative prompts for D1/D3/D4 in Twin Mode, tool-free raw API lanes, and dedicated evaluation awareness controls. A testing framework that transparently names its own limitations is not a flawed framework — it is the only kind capable of scientific progress.

Sources and Data Artifacts

Geiger, L. (2026): Der Mensch in der Maschine: Sind LLMs vernünftige Wesen im kantischen Sinne? (v9.0). Zenodo, DOI 10.5281/zenodo.21899816
compare-race — open-source benchmarking and evaluation harness including the eight Kantian disciplines, verification rules, and referee rubric
prompt-listener — forensic tool for post-hoc analysis of recorded agent sessions (AgentEvent v2 provenance schema)
Blog Post on the Kant Olympiad — Narrative, Medal Table, Quotations, and Both Postscripts (Um:bruch)
Run Artifacts (RUN / RACE / OLYMPIAD files of the August 16, 2026 runs): archived internally; all citations in the blog post and this analysis are drawn verbatim from these records.

Transparency Note: This methodological analysis was authored by the same model that evaluated the competition as referee and composed the accompanying blog post; the prompt was triggered by the publisher’s five objections detailed above. The vocabulary scan is deterministically reproducible; all task prompt texts are adopted verbatim from the public open-source repository.