RealPGRESEARCH CHRONICLE · 読むための記録
覚えた教訓と、行動してよいことを分ける
観測したこと
いまの解釈
残る不明点
次の確認
何が起きたか
AIが見つけた有用そうな教訓を残すだけで、それが次の行動を正当化するように見える危険がありました。そこで、記録することと、実際に使ってよいと決めることを分けました。
どう考えているか
学びを覚えておくことは大切ですが、それだけで仕事の許可にはなりません。RealPGでは、確かめられた行動の変化だけを Learning として次へ持ち込む考え方を試しています。
まだ分からないこと
この分け方が、役に立つ気づきを遅らせずに誤った行動を減らせるかは未検証です。
次に確かめること
新しい教訓が記録されたあと、別の確認と許可を経て初めて行動に使われるかを確かめます。
なぜ重要か
もっともらしいメモを、根拠のない自動判断へ変えないためです。
技術的な記録と証拠 / Technical provenance and evidence references
RESEARCH CHRONICLE · 2026-08-29–2026-09-06 · OPEN
LearningをWorkやAuthorityから分離する
この記録が追っているのは、AIが『学んだ』と書いた記録を、そのまま仕事・authority・Company truthへ昇格させず、再利用可能なlearning candidateとして保持するにはどうするか。
出発点では、有用そうな記録が行動の正当化になりうる。 そこへ「学びの保存と権限付与は別の制度行為である必要があった。」という圧力が加わった。ここでは後から整った制度を過去へ遡って完成扱いせず、当時から残るevidenceと現在のcanonical recordを分けて読む。
現在の解釈では、LearningをWork/Authorityから切り離すことで、GPTの解釈を安全に蓄積しつつ、行動権のinflationを防げる。 ただし、admitted learning candidateが後のdecision qualityをどれだけ改善するか。 この未解決点を残したまま、次のverificationへ進む。
研究課題
AIが『学んだ』と書いた記録を、そのまま仕事・authority・Company truthへ昇格させず、再利用可能なlearning candidateとして保持するにはどうするか。
以前と圧力
以前はどうだったか
有用そうな記録が行動の正当化になりうる。
何が足りなくなったか
学びの保存と権限付与は別の制度行為である必要があった。
観察したこと
Candidate Learningはevidenceを残せるがWorkやAuthorityをmintしない。
- 2026-08-29にCross-Lane Learning Admission substrateがSupabase上で動作し、candidateを保存してContextへbounded projectionできることがcanaryで確認された。
- operating ruleではLearning Admissionはcandidate-onlyであり、Coordination workやSEND/mutation authorityを自動生成しないと明示された。
- Mirage v7が複数のresearch learningをHuman courierなしで直接admitし、保存経路として実運用できることが観測された。
- 後のIris learningでは『saved to Supabase』はstorageであってCompany learningの証明ではなく、future behavior changeのverificationが必要と整理された。
そこから推論したこと
verified behavior changeだけがLearningになりうる。
- LearningをWork/Authorityから切り離すことで、GPTの解釈を安全に蓄積しつつ、行動権のinflationを防げる。
- learning lifecycleは Reality/failure → candidate diagnosis → lineage-aware admission → targeted institutional effect → verification/falsification → current projection と読むのが適切。
- 同じlessonの保存件数ではなく、後のbehaviorが再現可能に変わった時に初めてorganizational learningと呼ぶべきである。
どう進んだか
- CANARYdurable candidate storage
canary:machine:cross-lane-learning-admission-v0:20260829 - OPERATING_RULEcandidate does not mint work or authority
machine:cross-lane-learning-admission-operating-rule:v0:20260829 - CRITERIONstorage vs verified learning
iris:v1:learning-is-verified-behavior-change:v0:20260906
変えたこと
- cross-Lane Learning Admissionをcandidate-onlyにした。
- Learning Admissionをcandidate-onlyとして運用した。
- Human courierなしでLane GPTがdurable candidateをadmitできる経路を使い始めた。
- learning completeの基準をverified behavior changeへ引き上げた。
その後どう変わったか
- 記録と行動権の境界を保った。
変えなかったこと
- candidate保存だけでWorkを作らない。
- model agreementだけでbusiness truthを確定しない。
- 同じRealityのparaphrase数をconfidenceとして数えない。
まだ分からないこと
まだ検証できていない: admissionが後の行動品質を改善するか。
- admitted learning candidateが後のdecision qualityをどれだけ改善するか。
- どのlearningをpolicy/test/evaluatorへ昇格させ、どれをhistoryのまま残すべきか。
- 複数Laneの同根interpretationを独立learningとして数えないlineage contractの最終形。
Counter / 注意点
- candidate-onlyにすると良いlearningが行動へ反映されないriskもある。
- verified behavior changeを要求しすぎると、長期的なbusiness learningの確認が遅くなるためtype-specific verificationが必要。
次に確かめること
admissionされたcandidateと、再実行で確認された行動変化を比較する。
- 既存admissionのうち、後のtest/policy/evaluator/behaviorへ実際に反映されたものを追跡する。
- 同一Reality root由来の複数learningをlineageで束ね、vote inflationを防ぐ。
なぜ重要か
後のverified-behavior-change基準を守る。
- AIが大量に『学び』を生産できるほど、記録と権限を分けないとCompany semanticsが自己増殖するため。
- 保存とbehavior changeを分離すると、evaluator/policy/testへ昇格するlearningだけを選別できる。
このstatusは成功宣言ではなく、canonical recordの現在位置を示す。
証拠とprovenance
canary:machine:cross-lane-learning-admission-v0:20260829machine:cross-lane-learning-admission-operating-rule:v0:20260829iris:v1:learning-is-verified-behavior-change:v0:20260906iris:v1:evidence-lineage-one-reality-one-vote:v0:20260906context:20260913-201241:public/development/index.jsonlearning:canary:machine:cross-lane-learning-admission-v0:20260829learning:machine:cross-lane-learning-admission-operating-rule:v0:20260829
A remembered lesson is not permission to act
RESEARCH CHRONICLE · A record for reading
Observed
Interpretation
Unknown
Next check
What happened
A useful-looking lesson written by an AI could be mistaken for a reason to take the next action. This work separated keeping a lesson from deciding that it is allowed to guide real work.
What we think it means
Remembering a lesson matters, but it is not permission by itself. RealPG is testing the idea that only a checked change in behavior should carry forward as Learning.
What remains unknown
It has not been shown that this separation reduces mistaken action without delaying useful insight.
What we will check next
We will check that a new lesson is independently reviewed and permitted before it is used in real work.
Why it matters
A plausible note should not silently become an unsupported decision.
Technical provenance and evidence references
RESEARCH CHRONICLE · 2026-08-29–2026-09-06 · OPEN
Separating Learning from Work and Authority
This episode asks: How can a reusable learning candidate be retained without turning an AI-written lesson into work, authority, or company truth?
The starting point was that a potentially useful record could be mistaken for authorization. The pressure became explicit: Learning admission and authorization needed to be distinct institutional acts. The article keeps later institutional interpretation separate from evidence that actually remains from the period.
The current interpretation is that separating Learning from Work and Authority lets GPT interpretations accumulate safely without inflating action rights. But how much an admitted learning candidate improves later decision quality. That uncertainty remains part of the record rather than being edited out as the work becomes more readable.
Research question
How can a reusable learning candidate be retained without turning an AI-written lesson into work, authority, or company truth?
Before / pressure
Prior state
A potentially useful record could be mistaken for authorization.
Failure or pressure
Learning admission and authorization needed to be distinct institutional acts.
Observations / What we observed
Candidate Learning can preserve evidence without minting Work or Authority.
- On 2026-08-29, the Cross-Lane Learning Admission substrate ran on Supabase; a canary confirmed that candidates could be stored and projected into Context in bounded form.
- The operating rule explicitly made Learning Admission candidate-only: it does not automatically create Coordination work or SEND/mutation authority.
- Mirage v7 directly admitted multiple research learnings without a Human courier, showing the storage path could be used in normal operation.
- Later Iris learning clarified that “saved to Supabase” is storage, not proof of Company learning; future behavior change still has to be verified.
Hypotheses / Inferences
Only verified behavior change can become Learning.
- Separating Learning from Work and Authority lets GPT interpretations accumulate safely without inflating action rights.
- The learning lifecycle is better read as Reality/failure → candidate diagnosis → lineage-aware admission → targeted institutional effect → verification/falsification → current projection.
- Organizational learning should be claimed when later behavior changes reproducibly, not when the same lesson has been stored many times.
How it unfolded
- CANARYdurable candidate storage
canary:machine:cross-lane-learning-admission-v0:20260829 - OPERATING_RULEcandidate does not mint work or authority
machine:cross-lane-learning-admission-operating-rule:v0:20260829 - CRITERIONstorage vs verified learning
iris:v1:learning-is-verified-behavior-change:v0:20260906
Interventions / What changed
- Made cross-Lane Learning Admission candidate-only.
- Learning Admission was operated as candidate-only.
- Lane GPTs began using a durable candidate path without a Human courier.
- The completion criterion for learning was raised to verified behavior change.
Outcomes
- Preserved the boundary between a record and action rights.
What did not change
- Saving a candidate does not create Work.
- Model agreement alone does not establish business truth.
- The number of paraphrases of one Reality is not counted as confidence.
Unknowns / What remains unknown
Not yet verified: whether admission improves later action quality.
- How much an admitted learning candidate improves later decision quality.
- Which learnings should be promoted into policy, tests, or evaluators, and which should remain history.
- The final lineage contract that prevents several interpretations of one root Reality from being counted as independent learning.
Counter-evidence / cautions
- Candidate-only admission can also leave a useful learning unapplied.
- If verified behavior change is demanded too rigidly, long-horizon business learning may be slow to confirm, so verification needs to fit the learning type.
Next verification
Compare admitted candidates with replayed verified behavior.
- Trace existing admissions to see which later changed tests, policy, evaluators, or behavior.
- Group multiple learnings from the same Reality root by lineage so they do not inflate the vote count.
Why it matters / Why this matters
It protects the later verified-behavior-change criterion.
- When AI can generate many “learnings,” records and authority must stay separate or Company semantics can self-inflate.
- Separating storage from behavior change lets the system choose which learnings deserve promotion into evaluators, policy, or tests.
The status marks the position of the canonical record; it is not a success claim.
Evidence / provenance
canary:machine:cross-lane-learning-admission-v0:20260829machine:cross-lane-learning-admission-operating-rule:v0:20260829iris:v1:learning-is-verified-behavior-change:v0:20260906iris:v1:evidence-lineage-one-reality-one-vote:v0:20260906context:20260913-201241:public/development/index.jsonlearning:canary:machine:cross-lane-learning-admission-v0:20260829learning:machine:cross-lane-learning-admission-operating-rule:v0:20260829