Oleg Malkov is choosing where to work next; this registry is his criteria, applied in public. This page is the whole method behind the registry.
Levels
Drawn from the registry data: the grading scale and the zero-rendering rule.
Each dimension scores 0 absent, 1 partial or isolated, 2 repeated and concrete; maximum 12.
0
No qualifying public evidence in the reviewed sources. The row shows a dated line for this dimension, no digit.
1
Claimed or partial. Hiring language earns a 1 only when it describes a concrete operating practice.
2
Named and repeated-concrete: named systems with their jobs, bounded numbers, a recurring first-party record. Fleet reach and longitudinal depth are recorded as evidence notes on the claims.
The six dimensions
Named operating systems
Does the company name internal systems and show the job each performs?
Measured production results
Does it distinguish measured production results from illustrative arithmetic?
Adverse evidence
Does it publish failures, attacks, or limitations that make the company look fallible?
Reproducible mechanism
Does it expose enough architecture, sequence, controls, or code for another engineer to challenge or reproduce the mechanism?
Engineer authored cadence
Is this a continuing record by engineers across systems, rather than one campaign page?
Human and societal consequences
Does technical writing address trust, authority, privacy, or social consequences as part of the system design?
Admission
A company is admitted with human-graded evidence on at least 4 of the 6 dimensions and a total of at least 7 of 12.
7/12 with evidence on 4 or more dimensions is the qualification the research flow exercised on eight scored companies against a predeclared falsifier, resolved 2026-08-21. The threshold was calibrated for retrieval findability; the stated default is that the bar stays at 7/12, and a moved bar is recorded here with its reason.
Freshness, per claim
Each claim carries a status: current, stale or withdrawn. A monthly automated sweep will compare claim CONTENT at each URL (not yet built; it activates with the first admitted rows); any change flips the claim to stale and routes it to human re-grading. Changed or vanished evidence downgrades the claim while its history stays visible. Admitted rows get a human re-grade every 90 days, volatile hiring-language claims every 30.
What this cannot see
Public technical writing is selected, edited, and approved by the company; it cannot establish a team's ordinary management quality, psychological safety, compensation fairness, workload, promotion practice, or whether published lessons changed daily behavior. A polished publisher can outscore a stronger, quieter operator.
The grader's own record
The grader's own six grades, and why they are here
The person grading these companies is graded on the same six dimensions, from the same kind of public evidence, under the same gate. One record, two renderings: the method page shows it as the grader's disclosure, the practice page as the worked example.
Oleg Malkov
graded by a human · 2026-08-24
Named operating systems
No qualifying public evidence found in the reviewed sources · reviewed 2026-08-24
Measured production results
No qualifying public evidence found in the reviewed sources · reviewed 2026-08-24
Adverse evidence
No qualifying public evidence found in the reviewed sources · reviewed 2026-08-24
Reproducible mechanism
No qualifying public evidence found in the reviewed sources · reviewed 2026-08-24
Engineer authored cadence
No qualifying public evidence found in the reviewed sources · reviewed 2026-08-24
Human and societal consequences
No qualifying public evidence found in the reviewed sources · reviewed 2026-08-24
A company row cites published claims by other people. This record cites Oleg's own dated work samples by id, in the sample log the practice page renders. One kind of evidence, one home. The admission bar decides which COMPANIES are published. This record is published at whatever it measures, zero included, so the bar is not applied to it. Holding it back until it scored well would be the flattery the whole registry refuses. The record in full, with the exercises behind it, is on the practice page.
Provenance and effort
The aligned-companies research flow record (schema 1, version 5), research run 2026-08-21. The six dimension ids are adopted from it verbatim. Grading budget: 30 to 60 minutes, labelled an estimate; grading runs inside a stated weekly reviewer-hour budget and no row-count forecast exists anywhere.
Corrections
The house builds every listing; no company joined, consented or endorsed one. A company can challenge any claim or ask for removal. Corrections stay on the page with what was wrong. Write to oleg@mlkv.org.
References this design learned from
An AI-first ranking (Monolith Labs) and an employer directory (Aigents) already exist, so this registry claims no novelty. It differs by per-claim evidence a stranger can open, dated review, and a published admission bar.
The evaluated references, with what was adopted and what was rejected
Monolith Labs, Top 50 AI-First Corporations. REJECTED as a model: it ranks compute footprint, custom silicon and CapEx share, with no per-claim evidence URLs and no refresh dates. ADOPTED: the non-duplication caveat it forces. An AI-first list exists, so this registry claims evidence-graded operating practice and no novelty.
Aigents companies directory. REJECTED: a job board and employer showcase with no scoring. It defines the adjacent category this registry must stay distinguishable from.
KonMari consultant directory API. ADAPTED: a per-row graded attribute (certificationLevel, measured on 437 rows) and structured geography filtering. REJECTED: contact-detail publication. The endpoint exposes a personal email on every row, which the third-party privacy rule forbids outright.
Awesome manifesto. ADAPTED: narrow scope, a reason for every inclusion, and public contribution rules scoped to corrections only. REJECTED: pull-request inclusion as the admission mechanism (consent-shaped, incompatible with a registry the house compiles) and the ecosystem's star-count trust signal, stated nowhere in the manifesto and unverifiable.
sumodirjo/engineering-blogs. ADOPTED as the screening universe: it supplied 17 of the 20 expansion origins in the recorded provenance. A correction stays visible: an early design named kilimchoi from memory of which list is famous; the artifact says sumodirjo.
sso.tax (SSO Wall of Shame). ADAPTED: the checkable-fact discipline of grading named companies without their consent on evidence a stranger can open. Polarity inverted: this registry honors, and absence is silence.
B Corp directory. ADAPTED: a published score with the method behind every listing, and dated assessments. REJECTED: certification, which needs the company's consent, payment, and an independent auditor this registry lacks.
AI Incident Database. ADAPTED: evidence beside every claim and a standing correction route. REJECTED as an index shape: incident-keyed and adverse-only.
Have I Been Pwned API. ADAPTED: stable ids, separate added and modified dates, a retirement state, and source attribution. REJECTED: the single row-level verified flag. Verification here belongs to each claim.
Stripe (Minions) and Shopify (River, Aquifer). ADOPTED as rubric calibration exemplars of what a 2 looks like: a named system plus its job, a bounded denominator, human review in the operating flow. Calibration only, no implementation template.