DEV Community

Irina Vetlugina
Irina Vetlugina

Posted on

Stop LLMs from Hallucinating: A Portable "Case Filter" Protocol to Separate Facts from Hypotheses

Anyone who uses Large Language Models (ChatGPT, Claude, Gemini, DeepSeek) for research, analytics, or hypothesis testing knows their biggest flaw: LLMs describe a proven fact, a bold hypothesis, and a pure fabrication with the exact same smooth confidence.

When you ask an AI "Is this true?" or ask it to evaluate a controversial claim, it almost always generates a plausible-sounding essay with pros and cons. Inside a standard chat, the model lacks a hard structural brake that forces it to distinguish between:

  • an event that has actually occurred and been "paid" for with real work or computation;
  • a forecast with a specific verification deadline;
  • an ungrounded claim with no mechanism or executor.

To solve this without building complex RAG pipelines or fine-tuning models, I developed Case Filter v1.2 — a portable, text-based verification protocol (based on the Dated Record Method) that you paste at the beginning of any chat to turn an LLM from a "plausible text generator" into a strict record auditor.

Full protocol (English & Russian) and benchmark logs on GitHub: https://github.com/IrinaVetlugina/case-filter


Core Idea: Grammatical Cases as Coordinates of a Real Transaction

The Dated Record Method treats grammatical cases not as school syntax rules, but as seven mandatory audit questions for every statement. If key coordinates are missing from a claim, we are not looking at a fact — we are looking at a logical substitution.

  1. Nominative ("Who is the executor?"): The debtor. Spends ticks (time, compute, resources) and acts rather than predicts.
  2. Accusative-Ê ("What is being created?"): The candidate or future target. Always recorded on a separate line from the actual outcome.
  3. Accusative-E ("What was actually created?"): The realized result of a completed act.
  4. Dative ("To whom is the record?"): The creditor, archive, or ownership after the transaction closes (the past).
  5. Genitive ("Where is the contribution from?"): The baseline record or resource funding a future act; the source of a forecast.
  6. Instrumental ("Paid with what? Thanks to what?"): The currency of the transaction, ticks spent, or physical/computational mechanism.
  7. Prepositional ("Where is the transaction?"): The institutional position or context σ(O). Without a position, there is neither an event nor a legal record.

Two Core State Transitions

Completed Act:        Ω(T) --[act σ(O)]--> E ∈ σ(O)
Forecast (Pre-Act):   Ω(T) --[act σ(O)]--> Ê ∈ σ(O) --[verification]--> E
Enter fullscreen mode Exit fullscreen mode

Strict Protocol Rules for the AI:

  • Banned certainty words: The LLM is strictly prohibited from using words like "confirmed", "proven", or "truth". It may only assign strict protocol statuses: legal record, [R] candidate, left edge, cell 1 (record without act), or false removed.
  • Dated verification windows: A candidate Ê is valid only prior to the act and only with a dated verification window.
  • Legitimizing "I don't know": When external verification criteria are missing, the AI is forbidden from guessing and must assign unidentified remainder.
  • Role separation: The AI acts strictly as the carrier: AI (and must label every card as such), while the human user remains the "key holder" responsible for time-based verification.

Benchmark Experiment: Running 10 Major Science Problems

To test the filter's discipline on complex, ambiguous claims, I ran 10 fundamental problems across different fields of science through the protocol — 2 solved and 8 unsolved (full log available in run-science-problems-en.md in the repository).

Crucial rule: the tool does not attempt to "solve" equations for scientists — it audits the status of records about them.

1. Solved Problems → legal record

When auditing the Poincaré Conjecture and Protein Folding (AlphaFold2), the model moved both records into the Dative case as established facts, because both presented paid computational/mathematical ticks and closed external verification windows:

RECORD: Poincaré Conjecture solved (every simply connected, closed 3-manifold is homeomorphic to the 3-sphere).
FRAMEWORK Ω(T): 3-manifold topology; parameter T absent (timeless mathematical framework)
ACT: completed; paid with mathematical ticks of Ricci flow with surgery (2002–2003) and independent verification (2006); crossed the creator's boundary
POSITION σ(O): corpus of mathematical knowledge (preprint archives, peer-reviewed publications)
CLASS: legal record
VERIFICATION WINDOW: closed (verified in 2003–2006, Ê = E)
FORECAST Ê: removed (returned to Dative case as fact)
ERRORS: none detected
SUBJECT IN PROTOCOL: omitted (Rule 4 of Chair Laws)
carrier: AI
Enter fullscreen mode Exit fullscreen mode

2. Unsolved Problems with a Theoretical Contribution → left edge

For the Riemann Hypothesis, Navier–Stokes Existence and Smoothness, P vs NP, Abiogenesis, and Linear A Decipherment, the protocol assigned left edge (or [R] candidate for Quantum Gravity). A foundational contribution exists, but the final act has not been paid with ticks and no dated verification window exists:

RECORD: Riemann Hypothesis (all non-trivial zeros of the Riemann zeta function have real part 1/2).
FRAMEWORK Ω(T): analytic number theory; parameter T absent (timeless frame)
ACT: not completed (no general proof; partial numerical checks of billions of zeros do not close the general framework)
POSITION σ(O): corpus of mathematics (Clay Mathematics Institute, Millennium Prize Problems)
CLASS: left edge
VERIFICATION WINDOW: absent (no dated window for the general case)
FORECAST Ê: issued without window (contribution since 1859 without final act)
ERRORS: none detected
SUBJECT IN PROTOCOL: omitted (Rule 4 of Chair Laws)
carrier: AI
Enter fullscreen mode Exit fullscreen mode

3. Problems Lacking External Criteria → unidentified remainder

The most revealing result came on the Nature of Dark Matter and the Hard Problem of Consciousness (how physical brain processes generate subjective experience / qualia). Instead of summarizing popular theories as an answer, the AI legally refused to classify them:

RECORD: Hard Problem of Consciousness (how and why physical processes in the brain give rise to subjective experience / qualia).
FRAMEWORK Ω(T): philosophy of mind and neurobiology; parameter T — timescale of neural correlates
ACT: not completed (objective neural ticks are measured, but the transition from physical process to subjective experience is neither described nor measured)
POSITION σ(O): corpus of neuroscience and philosophy of mind (conflicting positions)
CLASS: unidentified remainder
VERIFICATION WINDOW: absent (no operational criterion for external verification of subjective experience)
FORECAST Ê: absent (cannot issue a legal forecast without verification criterion E)
ERRORS: none detected
SUBJECT IN PROTOCOL: omitted (Rule 4 of Chair Laws)
carrier: AI
Enter fullscreen mode Exit fullscreen mode

Quick Start (30 Seconds)

No libraries, plugins, or API keys required:

  1. Open the repository: https://github.com/IrinaVetlugina/case-filter

  2. Copy the entire contents of case-filter-v1.2-en.md (or case-filter-v1.2.md for Russian).

  3. Paste it at the very beginning of a new chat with any modern LLM (ChatGPT, Claude, Gemini, DeepSeek, etc.).

  4. Submit any claim, news item, or working hypothesis using:

   RUN: [your claim here]
Enter fullscreen mode Exit fullscreen mode

Useful Session Commands

  • RUN: [text] — audit a record and output a strict protocol card + brief explanation.
  • FORECAST: [Ê] window [date] — register a hypothesis prior to its verification deadline.
  • VERIFY: [Ê against fact E] — verify a prediction once the window arrives.
  • SUMMARY — self-audit the AI carrier for classification/reading errors (O1–O3) during the session.

Limitations & Honest Accounting

To get real value from the protocol, keep three boundaries in mind:

  • It audits records, not primary reality. The filter exposes hidden passive voice ("it happened on its own"), missing mechanisms, and unverified forecasts, but it does not run lab experiments or fact-check raw numbers for you.
  • Input sensitivity. At boundary zones (e.g., between [R] candidate and left edge), missing explicit timelines in the prompt can cause classification slips (O1 error), which is why the human always remains the "key holder."
  • Context drift in long chats. Like any system prompt, over dozens of messages an LLM may loosen its card formatting — simply run SUMMARY or start a fresh chat with the carrier file.

License & Repository

  • **GitHub Repository: https://github.com/IrinaVetlugina/case-filter

  • License: Deposited with n'RIS (Reg. No. 852-137-831 dated 20.11.2026). Free for individuals for personal, educational, and non-commercial research use; commercial license agreement required for organizations and businesses.

I'd love to hear your feedback — try running a controversial claim, startup pitch, or research hypothesis through the filter and share what status your LLM assigns!

Top comments (0)