Frozen scope
The question, mechanism map, date, language and source boundaries, inclusion rules, limits, and prospective stopping rule.
The SOTA Repo Method
Find strong mechanisms without pretending the search is complete.
We define the founder decision first, inspect public projects at exact versions, preserve the research trail, and extract the mechanisms worth adopting, adapting, integrating, or building. The result is a dated evidence map—not an automatic recommendation to install a repository.
Plain-English definition
SOTA is a search direction, not a trophy.
SOTA is shorthand for “state of the art.” Here, “SOTA Repo” means a disciplined attempt to find the strongest publicly inspectable mechanisms inside a declared boundary. It does not mean that one repository is permanently the best or that the open web has been exhausted. Every map states its date, sources, limits, unresolved areas, and honest coverage label.
The method
From a founder question to a build decision.
The order matters. We freeze the question before attractive findings can quietly redefine it, and we preserve misses and contradictions alongside wins.
- 01
Freeze the decision
Write down the founder question, cutoff date, languages, allowed sources, inclusion rules, resource limits, and stopping rule before searching.
- 02
Break the job into mechanisms
Split a broad label into the inputs, judgments, controls, outputs, and handoffs a complete workflow would need. Empty cells stay visible.
- 03
Search from several directions
Use terminology, native-language, documentation, maintainer, citation, dependency, license, release, adverse, and gap-focused routes so one ranking system does not define the map.
- 04
Keep the negative space
Record every plausible lead as retained, excluded, duplicated, blocked, or unresolved. “Not found,” “excluded,” and “not applicable” never mean the same thing.
- 05
Inspect the exact manifestation
Anchor important claims to a source file, documentation section, release, license, test, configuration, or original paper at a named version and date.
- 06
Extract mechanisms, not reputations
Ask what each project contributes to the workflow. Stars, citations, activity, and a strong demo help discovery; they do not validate every part of a system.
- 07
Turn the map into a build choice
Identify the strongest components, unsafe defaults, thin areas, missing controls, and the smallest complete capability worth building.
- 08
Stop honestly, then challenge it
Stop only under the predeclared rule. Freeze the bundle and let a fresh different-vendor reviewer recompute the accounting and probe the weak spots.
Evidence packet
The trail must survive the summary.
Search log
Every query or traversal, its visible result boundary, limitations, new leads, duplicates, and effect on the stopping clock.
Screening ledger
Every plausible project, its stable identity, discovery route, disposition, reason, and duplicate or successor relationship.
Evidence ledger
Exact claim locators, versions, licenses, evidence class, confidence, contradictions, and limitations.
Findings
The mechanism landscape, dangerous defaults, thin cells, reuse choices, human gates, build boundary, and remaining uncertainty.
QA + landing record
The reviewed bytes, discrepancies, repairs, coverage judgment, validation result, and reproducible hash of record.
From evidence to architecture
Four verbs prevent accidental wholesale adoption.
Adopt
Use a narrow component behind explicit tests.
Adapt
Reuse an inspectable pattern without inheriting the whole system.
Integrate
Connect a replaceable tool through a bounded adapter and permission gate.
Build
Create the shared layer that no inspected project supplies safely.
Why not copy a repository wholesale?
No single project owns the complete founder workflow.
A whole project can import hidden assumptions, unsafe defaults, unnecessary infrastructure, mixed-license surfaces, brittle integrations, or authority the founder never intended to delegate.
A useful idea is not automatically a suitable dependency.
Mechanism-level synthesis keeps strong parts replaceable. A project can teach us an excellent approval pattern while still being the wrong runtime, license, interface, or operating boundary for Founder Hub.
Independent QA
The producer does not grade its own map.
- 01
Freeze exact bytes
The reviewer receives the completed bundle and does not edit it.
- 02
Read ledgers before conclusions
Scope, searches, screening, and evidence are checked before the narrative earns trust.
- 03
Recompute the accounting
Counts, identities, mechanism cells, yields, lineage, and the stop decision are independently replayed.
- 04
Probe thin and decisive areas
The reviewer runs bounded searches where a miss would change the recommendation.
- 05
Record a verdict or repair
Acceptance, discrepancies, required repairs, or a lower coverage label remain durable.
- 06
Regrade changed bytes
Any repair creates a new hash and a new gate. A verdict cannot follow a changed artifact.
The stopping rule
Stop when the frontier goes quiet—not when the web is “finished.”
A working plateau requires every mechanism cell to have a disposition, every planned route to run or show its limitation, no high-priority lead to remain open, and two different final passes to find no new mechanism type and only the small yield allowed in the frozen rule. New mechanisms reset the clock.
If useful projects are still appearing when the resource ceiling arrives, the honest label is high-coverage, effort-bounded, not saturated.
What this method does not claim
Strong research still has a boundary.
No completeness claim
It does not prove exhaustiveness, saturation, true recall, or a permanent “best repository.”
No runtime guarantee
Documentation, tests, and configuration do not prove live quality, security, privacy, or production fitness.
No legal conclusion
License and rights observations are evidence for review, not legal advice or compliance certification.
No business-impact proof
The map does not establish customer demand, conversion lift, decision accuracy, or venture outcomes.
No installability transfer
Inspecting a project does not make it an installable or accepted Founder Hub capability.
No infallible QA
Independent review is bounded and can still miss relevant projects, evidence, or defects.
Current method status
Seven maps completed; the method itself remains a supervised draft.
The current maps used protocol v0.5.2. Their landing gates validate the individual artifacts and the honesty of their coverage labels; they do not ratify the method or turn a research bundle into a runnable capability.
From map to registry record
Evidence can justify a capability boundary. It is not the capability itself.
A record stays Planned until implementation, adversarial testing, repairability, and independent acceptance all exist together.