From the Layered Protocol to V2.0 · Historical Source Reader

How to read this route

The Chinese archive preserves five substantial source texts in sequence. This English page provides a verified guide to their claims and status. R numbers name research lines, not production releases. Canonical v1 was the former production baseline; none of R1 through R4.10 silently replaced it at the time recorded here.

R1 · Layered scoring protocol · July 12–15, 2026

R1 separated the calculation kernel, assessment protocol, benefit accounting, and larger aggregations. It introduced canonical benefit units, AND/OR relations, single-use attribution shares, uncertainty states, and regression gates. The design made duplicate accounting visible, but the gold-sample run did not close evidence for probabilities and shares and exposed confusion between node and system objects. Its durable lesson was structural: facts, accounting, and display must not be treated as one operation.

R2 · Outcome ledger / CACL · July 26–August 8, 2026

R2 moved the additive unit from a person-bucket card to an independently identifiable realized outcome. Aliases, renaming, parent-child slices, and repeat submissions could not create new light; truly independent non-overlapping outcomes could accumulate. A representative research form used per-outcome light and then added it across a person. Synthetic conservation tests showed that card splitting need not inflate the arithmetic. They did not prove that real outcome boundaries, cross-person credit, or cross-field magnitude had been measured correctly.

R3 · Epsilon · August 10–11, 2026

R3 typed facts and formalized roots, refinement conservation, unknown states, and open accumulation. Its working research formula gave each root one unit of light and then applied a personal share. Named-person calibration revealed the weakness: editorial choices about uniform duration and precise shares could move the same person across several tiers. Typed facts, deduplication, and HOLD survived; unsupported universal duration and point shares did not.

R4 · Independent-outcome case ledger · from August 11, 2026

R4 made one person’s credit for one independent outcome the object of review. The hard ledger checked identity, realization, overlap, evidence location, and HOLD. The judgment layer recorded realized value, causal role, reasons, objections, and possible no-credit outcomes. This made disagreements appealable, but also exposed review cost, fame bias, and the danger of turning case discretion into arbitrary exceptions. Early AI-panel runs were workflow stress tests, not panels of human experts or external replication.

R4.10 / V2.0-RC3 · August 14, 2026

RC3 removed the five dimensions from item arithmetic and proposed two explicit judgments: outcome magnitude M and personal attribution a. Its candidate chain was g=M×a, Q(g)=10^(g/2)−1, L=ΣQ(g), and S=2log10(1+L). It added canonical outcomes, tail blocks, refinement conservation, and a prohibition on inheriting relatives’ achievements. The package included bounded-life ledgers for Einstein and Edison and a fully synthetic ordinary life, plus curve-sensitivity examples. It remained unapproved and non-production because M and a still needed anchors, global attribution coverage, isolated replay, and a value decision on the curve.

What changed after RC3

RC3 placed allocation inside the exponent, so two equal co-contributors to an M=8 outcome would receive the light of M=4 each: 99 plus 99, far below the outcome’s 9,999 units of light. RC3.1 instead computed the complete outcome light first and allocated from that single budget: Q=10^(M/2)−1, Lᵢ=aᵢQ, and S=2log10(1+ΣLᵢ). This conservation change is the mathematical basis of the current research assessment. It does not prove that any particular M or a is objectively measured.

The current ranking uses itemized outcome ledgers and this conserved allocation chain. Evidence coverage, cross-field magnitude, shared attribution, and assessor agreement remain open to revision. The full Chinese primary-source archive retains the formulae, examples, arguments, and failure conditions behind each step.