鷹架能力紀錄：SSR、SDR、SCM 與消融階梯

定義

definitions

• Scaffolding vector S = (S_T, S_R, S_N, S_V, S_E, S_M, S_P); Q_SP = Q(M, 0), Q_F = Q(M, S_F).
• SSR = Q_SP / Q_F; SDR = 1 − SSR; SCM_j = C_j^F / C_j^SP; marginal scaffolding yield Y_S,k = (Q_{k+1} − Q_k) / (C_{k+1} − C_k); scaffolding physical overhead ΔP_S = P_F ⊖ P_SP (typed).
• Capability vector I = (I_SP, I_L, I_T, I_V, I_E); Scaffolding Capability Record 𝔖_C = (𝔔_SP, 𝔔_F, ΔQ_S, SSR, P_SP, P_F, ΔP_S, G_S, φ, P_waste, Grade_S).
• S-grades E unknown harness / D declared components / C execution trace / B controlled ablations / A factorial interaction / A+ physical-causal attribution.

前提

assumptions

• Conditions start from the same initial information (A0–A3); A4/A5 information gain is marked separately.
• Scaffold contributions interact (verifier needs candidates), so one-at-a-time ablation is not unique attribution.

主張

claims

• Loop ≠ Cheating; SystemCapability ≠ ModelNativeCapability; Pass@k ≠ Pass@1; BestOfN ≠ SingleTrajectory.
• GenerationAbility ≠ VerificationAbility; ToolAccess ≠ ToolUtilizationIntelligence; ExternalMemory ≠ InternalLearning.
• ScaffoldingDependence ≠ SystemDefect; Loopless ≠ SuperiorByDefinition; InvisibleOutput ≠ ZeroCost; HiddenRetry ≠ FreeRetry.
• TransparencyOfCompute ≠ DisclosureOfPrivateReasoning; ScaffoldMeasurement ⇒ QualityAndCostTogether.

形式化

formalization

• Shapley-like attribution φ_i = Σ_A w(A)[Q(A ∪ {i}) − Q(A)] with sampling approximation and uncertainty U_φ.
• Brute-force region: ΔQ ≪ 1 with ΔC ≫ 1; selection waste ratio SWR = C_discarded / C_total; P_waste = Σ discarded attempts.

預測

predictions

• A stable scaffolding response curve exists for a fixed model and task set; SSR < 1 on non-trivial tasks (Falsifiable Claim 4).

證偽／失敗條件

falsification_conditions

• If Q_F ≈ Q_SP across most tasks, models and compute budgets, the native/system separation loses empirical necessity (F4).

已知限制

known_limitations

• First real data point (2026-09-03): SSR = 1.0 on three easy tasks with a 3.1× energy multiplier — a null result on an instrument whose quality axis was itself confounded; no factorial or Shapley analysis yet.

證據

關係

歷史與來源歷程

Canonical URL

https://evemisslab.com/ai/theory/THY-2026-0109/

機器可讀

/ai/theory/THY-2026-0109/index.json

快照

AI-SNAPSHOT-v0.1-fe85b9694a45

來源歷程

source

EveMissLab research collection: Intelligence Physical Metrology (真本體論13)

extracted_by

Splice (Claude Code), reading the canonical UTF-8 sources and each package's own reports

extracted_at

2026-09-11

generator

tools/extract_all.py

claim_boundary

status, evidence level and result type follow the source artifact's own stated claim boundary; nothing is upgraded beyond what the report supports