User manual
This page explains what the site does and how it reaches its answers, so you can judge how far to trust them.
What this site is
The site holds the collected writings of Antoun Saadeh: about 2,200 articles, almost all in Arabic. It offers three tools:
| Tool | Who can use it | What it gives you |
|---|---|---|
| Search | Everyone | The passages whose meaning is closest to what you typed, with links to the articles. |
| Reasoning (beta) | Signed-in users | An answer to the question “how would he have reasoned about this?”, with the sources and checks shown. |
| Writing Studio (beta) | Administrators only | A search of independent web sources on a topic, then a neutral fact brief, a judgement memo or a stylistic simulation, with the sources and checks shown. |
The interface is available in English, French and Arabic (right-to-left). Choose the language at the top right. The articles themselves are in Arabic.
Search: finding articles by meaning
A normal search looks for the exact words you type. This one also looks for the idea. Before the site went live, every article was converted into a long list of numbers (an “embedding”) that captures its meaning. Your query is converted the same way, and the site returns the articles whose numbers are closest to yours. It also matches Arabic keywords, ignoring diacritics and spelling variants of letters such as alef and yaa.
- Type a concept or question in the search box.
- Optional: tick enrich. An AI model first expands your concept into related terms, which helps with short queries.
- Choose how many results you want (1 to 10).
- Open a result to read the full article.
About the percentages. They are relative scores for ranking results against each other. They are not the probability that a result is “correct”. A lower number can still be a good match.
Reasoning: how an answer is produced
Search finds what he wrote. Reasoning goes further: it tries to say what he would have concluded, including about questions he never addressed. Because that is easy to get wrong, the system is built so that an AI model can never simply speak for him. It works in five steps.
1. Gather evidence
The search finds up to six relevant passages and labels them E1, E2, and so on. Later writings count for more than earlier ones. Texts that are undated, published after his death, or confirmed as not his are either removed or marked as unconfirmed, and are never presented as his words.
2. Did he answer it himself?
A model reads the passages and decides whether he explicitly addressed this question. If so, it must support the answer with quotes. The software then checks every quote, character by character, against the article. A quote that is not an exact copy is thrown away. If no quote survives, the answer is discarded. When his own words are found, they are the answer.
3. If not, derive it from his principles
The site holds a reviewed catalogue of 13 official principles: eight basic principles and five reform principles. A model works from the highest principle down, shows how he applied the same principles to comparable questions (with verified quotes), applies them to your situation step by step, and lists its assumptions. “The principles do not decide this” is an allowed outcome.
4. A second reviewer checks it
A separate model call, with a different role, tests the conclusion against the principles and his later writings. It answers consistent, revise or reject. A “revise” sends the derivation back for one more try. If the check fails, the conclusion is withheld and not shown as his view.
5. Compute confidence and log disagreements
The confidence number comes from a fixed formula in code, not from the AI’s opinion of itself. Any conflict the system notices, between two passages, between principles, or between his documented answer and the principle model, is written to a tension log for the curator to review. It is never settled silently.
The three possible outcomes
| Outcome | Meaning | Strength |
|---|---|---|
| Documentary | He answered this question himself. His verified words are shown. | Can reach “High”. |
| Principle-derived | He did not answer it. The answer is an inference from his principles, and passed the independent check. | Never above “Medium” (74 out of 100). |
| Undetermined | The principles do not decide it, or the check did not accept the conclusion, or a step could not run. | No answer is given as his view. |
Writing Studio
A page for administrators only, isolated from Search and Reasoning (its own route, services and off switch). You type a topic; the site searches independent web sources and builds everything from corroborated facts only. All output is in Arabic.
| Mode | What it produces |
|---|---|
| Fact brief | The claims extracted from the pages read, each with its sources and status, the points in dispute, and what could not be established. No author voice is used here. |
| Judgement memo | A judgement derived from his principles, built on the corroborated facts only, audited sentence by sentence by a fact-fidelity checker, with a verification margin for conditional and unaudited sentences. |
| Article (stylistic simulation) | A long text in a simulation of his style (900–1,800 words), built on the corroborated facts and his derived position, passed through four checks; it is withheld whole if it fails. |
| Editorial (stylistic simulation) | A short text in a simulation of his style (300–500 words), with the same checks. |
Running a Studio query
- Open Writing Studio from the navigation bar. It is for administrators only; if the page answers “not found”, the feature has been switched off in the site’s configuration.
- Type a topic of 3 to 300 characters.
- Choose one of the four modes, then submit. A run searches, reads pages and checks its own output, so it takes a while; the button is disabled until it finishes.
- Read the result: the facts and their sources, the principles used, his verified words, and any verification notes. If the checks fail, the page says the text was withheld rather than showing a weakened version.
- If a run stops early, a message says why: the daily run limit was reached, the time limit was exceeded (try a narrower topic), or the research service is unavailable. The rest of the site is unaffected.
The evidence rules
- Every source is vetted before use: a domain on the curator’s list takes its tier (A or B), a denylisted domain is dropped, and an unlisted domain gets an automatic score that can never rise above the “provisional” tier, however high the number.
- A claim counts as “corroborated” only when tier A or B sources from two independent owners back it. Two sources under one editorial owner count as one.
- A “provisional” source never counts towards corroboration; the curator can promote it by listing it in the trusted file.
- Claims that disagree are all shown, marked “contested”; nothing is settled silently.
Three scores, always separate
The factual score comes from a fixed formula in code (the share of corroborated facts, minus 5 points per contested point), not from the model’s opinion. The principle score is the reasoning core’s confidence report, passed through unchanged (the derived cap of 74 stays). The style score is two numbers side by side: the code’s stylometry and a model judge’s. No code path blends or combines these scores.
The limits
- Per-run caps: model calls (30 by default), pages fetched (12), an overall wall-clock deadline (600 seconds by default) and, optionally, a daily number of runs per administrator.
- Output is Arabic only, and his real words appear only in the separate “his verified words” block, never inside the generated text.
- The fact-fidelity check fails closed: any sentence that cannot be traced to a corroborated fact moves to the verification margin. In the stylistic simulation the text is regenerated and, if it still fails after two revision rounds, it is withheld whole, never shown gutted.
- Every simulation carries a fixed banner above it: a stylistic simulation, not his writing. The banner comes from the code, not the model.
How to read an answer
An answer page shows its working so that you can check it. From top to bottom you will find:
- The answer and its tier: documentary, principle-derived or undetermined.
- Quotes: exact text from his articles, each with a link, a date and a period. Only these are “his words”.
- Principles used: marked “inherited” when a chosen principle depends on them.
- Method traits: how he tends to reason (for example, defining terms first). These are patterns noticed in his writing, not official principles.
- Steps and assumptions: the chain of inference, and what was assumed about the modern situation.
- The check: the reviewer’s verdict and any principle it says is violated.
- Tensions: conflicts that were found and logged.
- Evidence: the passages the system looked at, including those marked unconfirmed authorship (context only).
- Diagnostics: what the system discarded and why, for example a quote that was not an exact copy.
What the confidence number means
The score is out of 100 and is built from five factors. It measures how well the answer is documented and checked, not how likely it is to be right.
| Factor | Weight | Full marks when |
|---|---|---|
| Documentation | 30% | There are three or more verified quotes. |
| Principle coverage | 25% | The principles involved are numerous and high-level enough (basic principles count most). |
| Curation | 15% | Every principle involved has been reviewed by the curator. |
| Consistency | 20% | The reviewer accepted it first time, and no tensions were found. Each tension takes off 25 points of this factor. |
| Maturity | 10% | The quotes come from his later periods (1938 onwards). |
Labels: High is 75 or more, Medium is 50 to 74, Low is below 50. Cap: an answer derived from principles is limited to 74, however well documented, because only his own words can be rated High. When the cap applies, the page says so.
What to trust, and what not to
Reliable by design
- A quote shown as “his words” exists, exactly as shown, in the linked article. Software verifies this.
- Unconfirmed texts, such as likely student writings, are never quoted as his.
- A conclusion the reviewer rejects is not shown as his view.
- The score comes from a published formula.
Use with care
- A principle-derived answer is an inference. He did not say it.
- A verified quote proves the words exist, not that the AI read them correctly. Open the article and read the context.
- AI models can misread Arabic nuance, and the writer, reviewer and scorer are the same kind of model, which can share blind spots.
- Dates and authorship of some articles are curated estimates. A wrong date changes how a text is weighted.
- This feature is marked beta. Do not cite a derived answer as a statement by the author.
Questions and answers
- Why does Reasoning say “undetermined”?
- The principles did not decide the question, the independent check did not accept the conclusion, or a step could not run (for example, the AI service was unavailable). The page lists the reason under Diagnostics. Rephrasing the question more specifically often helps.
- Why was a quote I expected missing?
- Quotes are dropped if they are not an exact copy of the article text, are too short to prove anything, or come from a text whose authorship is unconfirmed. Diagnostics lists each one.
- What language should I ask in?
- Either. Prose is written in the language of your question. Quotes are always the original Arabic.
- Why do I have to sign in for Reasoning?
- The Reasoning page is restricted to signed-in users by the site’s configuration. Search is open to everyone.
- Is my question stored?
- Searches can be logged for site analytics. When Reasoning finds a conflict (a tension), the question is saved with that tension record so the curator can review it. See the privacy page for details.
- Who curates the principles?
- The site owner reviews and approves the catalogue. Administrators can inspect the principles, the tension log and the back-test results, which measure how often the system reaches the same conclusion as his real answers when those are hidden from it.
- What does “provisional” next to a source mean?
- A domain not on the curator’s list receives an automatic reliability score, but automatic scoring can never raise a source to the reviewed tiers A or B, and a provisional source never counts towards corroborating a claim. The curator can promote it by listing it in the trusted file, with a stated reason.
- Is the text in the Article and Editorial modes written by him?
- No. It is a generated stylistic simulation, and a fixed banner above it says so. His real words appear only in the separate “his verified words” block. If the text fails the fact-fidelity or principle checks after two revision rounds, it is withheld whole, never shown gutted.