A public disclosure · last updated May 30, 2026
How this site is written.
A public commitment to a production standard. What every essay must contain, what AI does and doesn't do here, and the weekly rhythm that keeps two serious pieces a week sustainable.
The seven standards every published piece must meet.
These are not aspirations. A draft that misses any of the seven does not ship — it goes back to the queue. Readers should be able to call us on it.
- 01
Primary data, not synthesis
Every essay contains at least three things only the author could write: a dated observation from client work, a screenshot or transcript from a specific model on a specific date, or a quoted exchange the author was present for. If a piece can be assembled from public sources alone, it is not published.
- 02
Methodology is disclosed
Empirical claims carry a Methods block: the window, the sample, the prompts, the model versions tested. Where possible, the raw spreadsheet or dataset is linked. The reader is meant to be able to disagree with the method, not just the conclusion.
- 03
References are formal
Working papers carry a numbered references section with author, year, title, venue, and URL. Footnotes mark technical qualifications. The first time a paper, framework, or person is named, the citation is at hand.
- 04
Version, dated, signed
Each piece carries a semantic version (v1.0, v1.1…), a published date, and a last-revised date. Substantive revisions bump the version and append a changelog entry. Errata are encouraged and acknowledged by name.
- 05
Falsifiable, not rhetorical
Working papers include at least one claim specific enough that a reader with the data could disprove it. This is the test of seriousness: if nothing in the piece could be wrong, nothing in the piece is a contribution.
- 06
Scannable summary up front
Every working paper carries a Key Claims block immediately after the abstract — 6 to 8 bullets, each a complete sentence with a specific number, date, or named entity. The discipline forces the writer to state the contribution sharply; the affordance lets the reader screenshot and quote it. Generic platitudes do not survive the bullet form.
- 07
Voice is human, scaffolding is not sacred
The thesis, opinions, lived examples, contrarian moves, and final argument of every piece are written by Gilad Sasson. AI tools are used as research assistants — for reference hunting, structural critique, and prose polishing — never as ghostwriter. Every published piece is rewritten paragraph-by-paragraph in the revision pass, replacing generic competence with specific authority.
The weekly production cadence.
Two pieces a week is the upper edge of what one practitioner can sustain at this length. The way it stays sustainable is by separating the work that needs the author from the work that doesn't — extraction first, drafting in the middle of the week, revision at the end.
Extraction
30 min voice memo on the week's observations — client work, new model behaviors, conversations. Transcribed. This is the raw, uncopyable material every piece is built from.
Selection · outlining
Two topics chosen from the pool. References pulled for each — 3 to 8 primary sources per piece. 200-word outline written: thesis, sub-claims, evidence, counter-position, what the piece would fail at.
Piece I draft
Working paper or field note draft. AI assists with structure scaffolding, reference formatting, transition smoothing. Thesis, opinions, examples, contrarian moves, closing — author only.
Piece II draft
Second piece, same discipline. If one of the two is a working paper, the other is usually a field note to maintain the weekly rhythm.
Revision pass
Both pieces reviewed paragraph-by-paragraph. Three rules: (1) add a specific date, name, or number; (2) cut any sentence that could appear in any SEO blog; (3) insert at least one aside, hedge, or contrarian qualifier per piece.
Publish
Methods, references, and version footer applied. Published in reverse-chronological order on /research. RSS, social, and the home page update automatically.
Tier mix: one field note (~1,500 words) + one working paper (~4,500 words) per week. Working papers are the substantive contribution; field notes are the dispatch from the practice. Both meet the six standards. Neither is filler.
Where AI is and isn't used here.
The honest answer matters more than the rhetorical one. Pretending the writing is untouched by AI tools is neither true of any modern practice nor more credible than transparency. Here is the line:
- Reference hunting — finding the relevant paper, patent, or prior post
- Outline critique — stress-testing structure before the draft
- Citation formatting and bibliography mechanics
- Transition polishing in the second-to-last pass
- Catching factual contradictions inside a long draft
- Grammar and consistency cleanup
- The thesis of any piece
- The opinions and contrarian positions
- The lived examples, dates, named clients, and case specifics
- The methodology disclosures and data tables
- The closing argument of any piece
- The revision pass — every paragraph rewritten by hand
The test: every essay must contain at least three things only the author could have written. That requirement is the part no LLM can fake — and the part that earns the byline.
Patterns that mark prose as AI-generated, and what replaces them.
Search engines and LLMs are increasingly tuned to recognise the surface tics of unedited model output. The revision pass strips these out — not to evade detection, but because they signal lack of care. Below: the list we audit against before publishing.
| Tell — avoid | Instead |
|---|---|
| Tricolons in every paragraph (X, Y, and Z) | Varied — pairs, single emphases, lists of 4 |
| "Not just X but Y" construction | Just Y |
| "It is worth noting that…" | Cut entirely |
| Round numbers (1000, 500) | Real numbers (847, 312) |
| Perfectly balanced ~80-word paragraphs | 30–200 words, varied deliberately |
| Generic transitions (Furthermore, Moreover) | Specific connective tissue |
| Missing dates, names, model versions | Always specific: "GPT-5 Pro, 2026-05-15" |
| "In conclusion…" wrap-ups | End on the strongest claim |
How to call us on it.
A standard that cannot be checked is not a standard. Three concrete commitments:
- Errata are public. Spot a factual error, a citation that doesn't support the claim, a methodology that doesn't replicate? Email algoholic@algoholic.com. The piece is corrected, the version bumps, the changelog notes what changed and who flagged it.
- Methodology data is shareable. Any empirical claim is backed by a dataset, prompt log, or spreadsheet that we will share with serious requesters. Ask.
- The full archive is auditable. The git history of this site is public. Every essay's revision history is visible at github.com/nekudai-google/algoholic.com. Nothing is silently rewritten.
Hold us to it.
This page is itself a versioned document. If the practice drifts from the standard, the standard wins — not the practice. Write.