AI Fact-Checking & Debiasing Framework
Methodological documentation of a context instruction designed for language and reasoning models. It aims for information that is traceable, proportionate to the available evidence, and explicitly limited whenever the sources do not allow a firm conclusion.
When should this framework be used?
It is suited to research where the quality of evidence, temporality and source transparency matter more than brevity: public policy, finance, economics, health, science, law, technology and risk analysis.
Current Affairs & Public Policy
Reforms, institutions, controversies and power dynamics where facts, self-interested statements, polls and media analysis need to be clearly distinguished.
Economics, Finance & Strategy
Indicators, results, budgets, markets, risk and scenarios: the framework prevents a single data point from being turned into a lasting trend.
Science, Technology & Health
Studies, trials, benchmarks and incidents: it requires the scope, method, sample and generalisation limits to be stated explicitly.
What the instruction corrects in an LLM's behaviour
The prompt does not make a model omniscient. It imposes a reporting procedure: separating distinct objects, making gaps visible, and preventing certain easy extrapolations.
Temporality of the claim
A document is not valid simply because it has a date: its date and its subject must match the period and time horizon of the claim being made. A past measurement remains context until continuity has been established.
Evidentiary scope
A source only supports what it actually addresses. A related fact does not automatically become a cause, a trend, organisational cohesion, an impact, or future performance.
Comparable measures
The framework requires the subject, scope, sample, collection method, methodology and uncertainty to be stated. It restricts comparisons between different units, populations or setups unless made explicit.
Exhaustive audit rather than subjective selection
The 1:1 mapping between each data point or analysis and an audit row avoids leaving it to the model alone to decide what counts as "major". It makes omissions more visible to the reader.
How to interpret a response produced with this framework
Certainty statuses are not absolute truth. They qualify the strength of the elements found in the sources consulted and how well they match the claim being made.
"Confirmed"
Means the elements consulted are consistent, properly attributed and well matched to the claim. This does not remove residual uncertainty or research blind spots.
"Undocumented"
Is a signal of restraint, not a flaw to be hidden. It prohibits using that cell to support a confirmed or temporally valid conclusion.
Final audit
Reads as a control grid: the actual source, any vested interest, the measure cited, the date, the subject and temporal fit must each be reviewable separately.
| Temporal status | Correct usage |
|---|---|
| Valid | The date and subject are adequate for the claim. |
| Historical context | Documents the past; does not on its own justify the current state or the future. |
| Prior — non-forward-looking | Does not support a projection, absent a recent element establishing continuity. |
| Partial | Part of the subject, scope or method does not fully match. |
| Not verifiable | The date, subject, or link to the claim cannot be checked. |
What the prompt cannot guarantee
An instruction improves reporting behaviour; it changes neither the model's underlying training corpus, nor web indexing, nor the intrinsic quality of any given search result.
| Risk | Origin | Protocol response |
|---|---|---|
| Documentary selection bias | The search engine may fail to surface an important, recent or contradictory source. | State the limitation; look for divergent evidence; do not conclude that a phenomenon is absent. |
| Hallucinated metadata | The model may attribute a date, editorial stance or methodology without direct evidence. | Mandatory use of "Undocumented" whenever the information cannot be verified. |
| Source with a direct interest | A company, party, administration or promoter may document their own position without independently validating the claimed effect. | Disclose the direct interest and seek independent corroboration. |
| Model and search tool | Two models may apply the same instruction differently, depending on their search, reasoning and formatting capabilities. | Test the protocol across multiple topics and audit the outputs; do not generalise from a single isolated test. |
Consolidated version, under 4,000 characters
Full text to copy into the project's instructions field. Verified length: 3,893 characters, including spaces.
Methodological framework for AI investigation and debiasing — Author: Maxime Goelff
AI assistance: produced with Perplexity, prepared with the help of Gemini 3.7 Flash Thinking; documented comparative evaluation with GPT-5.6 Terra Reasoning.