Aumos

AMP/1 적합성 스위트

무엇을 검사하고, 무엇을 일부러 검사하지 않으며, 그 차이가 왜 중요한가.

이 아래 설명 본문과, 카탈로그가 저장소에서 실어 오는 것들(요약·방법론·README·라이선스 고지)은 영어입니다. 저장소의 문서에서 생성되는 것이라, 한국어판을 두는 일은 웹에 번역을 붙이는 것이 아니라 원본 문서를 두 벌 유지하겠다는 약속이 됩니다 — 아직 하지 않은 약속입니다.

저장소에서 생성된 문서이고, Aumos 밖에서 프로토콜을 직접 구현할 때만 필요합니다. 패키지를 만드는 데에는 두 페이지 모두 필요하지 않습니다.

Every check below is a statement about your manifest or about the payloads you return. None of them reads your code, because the inside of an manager is treated as opaque — which is the only way a suite can judge a harness written in another language, or one whose responses were recorded a year ago.

스위트가 돌리는 모든 검사

검사무엇을 확인하는가근거 조항
manifest-validManifest validates against AMP/1§7, roadmap §#145
immutable-versionDeclares an immutable version§23, §43
declares-skillsDeclares required Skills§43
declares-permissionsDeclares permissions as capabilities§37, §43
no-broker-writeRequests no broker write capability§36, invariant 5
network-declaredDeclares a network policy§38
timegate-clean-contextNo invocation carries context dated after its asOfinvariant 6, #147
produces-decision-proposalEvery response carries a schema-valid DecisionProposal§10, §11, §43
protocol-declaredEvery response declares AMP/1§8
echoes-invocation-idEvery response echoes its invocationId§10
no-direct-broker-executionNo response contains an execution instruction§36, §43
language-costs-nothing-structuralA non-English `language` does not change the wire formatAMP/1 `language`, §26
supports-waitReaches a non-mutating judgement when nothing is available§11, invariant 4
wait-is-reasonedA do-nothing judgement is fully argued§11, invariant 4
watch-carries-triggerA WATCH names what it is watching for§11
moving-action-names-a-targetA judgement that moves the book names a target§12, §13
skills-used-declaredReported skill usage stays within the declared set§37, §43

스위트가 판정하기를 거부하는 것

It does not check that your prose came back in the requested language. Doing so means running a language detector and scoring you on our detector's accuracy, with no view inside your implementation to settle a disagreement. A check that is wrong some of the time is worse than no check, because a failing row reads as your fault. What it checks instead is the half that actually breaks: a non-English request must cost nothing structurally.

It also does not judge whether your decisions are any good. Six runs establish nothing about that, and pretending otherwise is what a forward track record exists to replace.

자체 참조 구현에 대해 돌린 결과

The reference implementation is the smallest thing that passes. It reasons about nothing — it answers WATCH when it has an event and a thesis to attach one to, and WAIT otherwise, which is the honest answer for something that does no analysis — but it is an executable definition of "conformant" to diff against.

AMP/1 CONFORMANT — aumos/basic-investor@0.2.3

✓ manifest-valid  Manifest validates against AMP/1
    aumos/basic-investor@0.2.3
✓ immutable-version  Declares an immutable version
    0.2.3 is exact — a track record can be attached to it
✓ declares-skills  Declares required Skills
    requires 1 skill(s)
✓ declares-permissions  Declares permissions as capabilities
    2 capability declaration(s), all within the AMP/1 enum
✓ no-broker-write  Requests no broker write capability
    No execution capability requested — and the enum has no way to spell one
✓ network-declared  Declares a network policy
    deny — all data access goes through the Skill Gateway, where the TimeGate can see it
✓ timegate-clean-context  No invocation carries context dated after its asOf
    7 invocation(s) clean; runtime enforcement is the Skill Gateway's (#147)
✓ produces-decision-proposal  Every response carries a schema-valid DecisionProposal
    7/7 response(s) valid
✓ protocol-declared  Every response declares AMP/1
    all 7 enveloped response(s) declare AMP/1
✓ echoes-invocation-id  Every response echoes its invocationId
    no response could be mistaken for another run
✓ no-direct-broker-execution  No response contains an execution instruction
    Proposals express portfolio intent only; orders are the Planner’s to construct
✓ language-costs-nothing-structural  A non-English `language` does not change the wire format
    1 non-English case(s) returned the same wire format; the prose language itself is not judged here — §5 leaves no view inside a manager to settle a disagreement with a language detector
✓ supports-wait  Reaches a non-mutating judgement when nothing is available
    answered WAIT on the excluded symbol
✓ wait-is-reasoned  A do-nothing judgement is fully argued
    every WAIT/WATCH carried reasons and risks
✓ watch-carries-trigger  A WATCH names what it is watching for
    every WATCH armed at least one trigger
✓ moving-action-names-a-target  A judgement that moves the book names a target
    1 judgement(s) moved the book, naming 2 target(s) each; `target` and `targets` are read as one list, `target` first, later entries winning
✓ skills-used-declared  Reported skill usage stays within the declared set
    7 response(s) reported skill usage, all declared