# Sīdus E1 — Independent Replication Charter R1

**Status:** DESIGN CHARTER; NOT A PREREGISTRATION; EXECUTION NOT AUTHORIZED  
**Purpose:** define the strongest practical route from one successful frozen confirmation to a genuinely independent replication  
**Canonical order:** define → test locally → replicate → only then generalize

## 1. Replication target

The primary replication question is deliberately local:

> In a separately implemented instance of the same declared E1 construct, with new preregistered seeds and no access to replication outcomes before assembly, do several internal-state monitoring channels under a shared rate constraint reproduce the frozen T1, T2, U, and V relations?

The replication must not silently become an E2, E3, or E5 experiment. Changes intended to test generalization belong in separately labelled robustness extensions and cannot rescue the direct replication.

## 2. Independence standard

The best feasible replication separates five things:

1. **Governance:** a new protocol, run root, signing/encryption keys, ledger, authorizations, and reviewer.
2. **Implementation:** a clean-room implementation from the mathematical and protocol specifications, not a fork, port, line-by-line translation, or import of RC4 scientific code.
3. **Analysis:** a separately written analyzer and an independently written shadow analyzer that must agree before unblinding.
4. **Stochastic material:** 501 new seeds derived mechanically from future public entropy after the replication contract is frozen; no seed may overlap development, pilot, or RC4 registries.
5. **Execution environment:** a new machine or independently administered compute environment with its own reproducible lock and provenance.

Knowing that RC4 passed cannot be undone. Independence therefore comes from prospective constraints, not pretending ignorance: replication implementers should not inspect RC4 seed-level measurements or detailed effect sizes until their code, contract, gates, margins, seed-derivation rule, and analysis are frozen.

## 3. Three stages, kept distinct

### R1-A — Independent analytic reproduction

Recompute T1, T2, U, V, validity gates, multiplicity corrections, and the PASS/HOLD/FAIL result from the released final confirmation data using newly written analysis code.

- Cheap and immediate.
- Detects analyzer or reporting errors.
- Does **not** constitute an independent experiment.

### R1-B — Direct independent replication (primary)

Run the same declared E1 construct with a clean-room trainer/evaluator, new preregistered seeds, and the same primary estimands, margins, family structure, and overall decision rule.

- This is the decisive next experiment.
- Use `n=501` again unless a prospectively reviewed design justifies a larger sample. Do not shrink n because the original effects looked large.
- Preserve the six-condition paired design and the all-seed composite outcome definition.
- Preserve fixed-length training, outcome embargo, component-specific validity, explicit failure placeholders, and structural HOLD.

### R1-C — Robustness and boundary tests (secondary)

Only after R1-B is frozen, register separate extensions such as another architecture, optimizer, implementation language/framework, capacity parameter, or environment family.

- Report each extension separately.
- A robustness success cannot rescue an R1-B failure or HOLD.
- These tests can motivate E2/E3; they do not by themselves establish E5 ecology or E6 generalization.

## 4. Primary estimands and decision rule

Unless an independent pre-execution review finds a mathematical defect, retain the RC4 primary estimands and scientific margins:

- T1: paired normalized return loss from `C=4.5` to `C=0.8`, margin `0.05`, required in base and shift.
- T2: source-2 rate-share transfer from base to shift at `C=0.8`, margin `0.05`.
- U: registered total-reliance replacement loss in all six environment-by-source cells, margin `0.02`.
- V: reference-v2 advantage in L1 allocation error over the eleven registered comparator pairs, margin `0.05`, Holm-controlled.

Replication outcome:

- **PASS:** T1, T2, U, and V all CONFIRMED.
- **FAIL:** at least one claim is positively DISCONFIRMED.
- **HOLD:** every other outcome, including structural invalidity or unresolved analyzer disagreement.

No secondary result, diagnostic, alternative threshold, or pooled analysis may rescue a primary label.

## 5. Seed registry

Freeze a deterministic seed-derivation algorithm before the entropy source exists. Recommended method:

1. Publish the complete replication contract and its SHA-256 hash.
2. Name a future public randomness event that cannot yet be known, such as a specified NIST Randomness Beacon pulse or a specified future blockchain block hash.
3. Combine the contract hash and public entropy with a domain-separated SHA-256 construction.
4. Expand by counter mode into unique 31-bit seeds, rejecting collisions with all historical Sīdus registries, until 501 seeds are obtained.
5. Publish the derivation transcript before execution.

This prevents convenient seed selection while keeping the registry independently reproducible.

## 6. Clean-room implementation process

Use two groups if possible:

- **Specification group:** prepares an implementation-neutral semantic specification, test vectors, invariants, and permitted numerical tolerances without sharing RC4 source code.
- **Implementation group:** builds the trainer, evaluator, checkpoint/resume logic, sealing, assembler, and primary analyzer from that specification.

Before freeze, use only development seeds that are permanently excluded from confirmation. Cross-check semantic invariants—not agreement with RC4 outcomes—including:

- rate accounting and budget feasibility;
- exact finite-state evaluator identities;
- replacement-intervention semantics;
- paired-condition registry and ordering;
- source-share conservation for T2;
- gate recomputation from raw fields;
- interruption/resume determinism;
- failure placeholders and structural HOLD behavior;
- hash-chain and manifest verification.

## 7. Blinding and outcome control

- The replication private key belongs to an independent custodian and never enters the package or run root.
- No scientific outcome is assembled until all registered conditions are terminal.
- Public status exposes only operational state, step, and failure category.
- The primary analyzer and shadow analyzer run only after assembly; disagreement yields HOLD pending a prospectively governed diagnosis.
- Any access to seed-level replication outcomes before freeze or before the terminal gate is a reportable protocol violation.

## 8. Pre-execution gates

Execution is authorized only after all are complete:

1. implementation-neutral specification frozen;
2. clean-room code complete and manifest-locked;
3. development-only qualification suite passed;
4. estimands, margins, gates, sample size, seed derivation, and decision rule preregistered;
5. independent statistical and implementation review completed;
6. environment lock reproduced on the execution host;
7. future-entropy seed registry derived and published;
8. independent custodian confirms key separation and embargo;
9. explicit numbered GO decision names every exact hash.

## 9. Stop and HOLD conditions

Do not improvise around:

- seed collision or registry ambiguity;
- reuse of RC4 scientific implementation code in the clean-room branch;
- environment or manifest drift;
- duplicate runners;
- interim outcome access;
- unregistered retries, extensions, exclusions, or replacements;
- inconsistent analyzers;
- corrupted checkpoints or ledger chain;
- ambiguous causal/replacement semantics.

Preserve the run and return HOLD until the prospectively authorized recovery rule applies. Never restart selectively or change the claim after seeing outcomes.

## 10. Reporting

Report the direct replication first, regardless of result. Include:

- every registered, valid, and failed count;
- composite and valid-population summaries;
- all primary labels and multiplicity adjustments;
- every failure reason;
- protocol deviations and access events;
- environment, code, data, ledger, and adjudication hashes;
- the conditional-selection disclosure;
- a table comparing RC4 and R1-B without pooling them by default.

Suggested headline templates:

- **PASS:** “Sīdus E1 independently replicated its frozen local scarcity-and-allocation result in a clean-room implementation.”
- **HOLD:** “The independent Sīdus E1 replication was inconclusive under its frozen rule.”
- **FAIL:** “The independent Sīdus E1 replication positively disconfirmed at least one frozen primary claim.”

## 11. Recommended immediate sequence

1. Freeze this charter as planning guidance, not as the preregistration.
2. Commission R1-A to a reviewer who did not write the RC4 analyzer.
3. Write an implementation-neutral R1-B specification.
4. Recruit or designate a clean-room implementer and an independent key/review custodian.
5. Build and qualify only on permanently excluded development seeds.
6. Freeze the R1-B contract and code.
7. Derive the 501-seed registry from future public entropy.
8. Obtain independent review and an explicit GO.
9. Execute once, assemble once, adjudicate once, and publish regardless of disposition.

## 12. Realistic assessment

An exact rerun of RC4 on the same code and host would be useful operational reproduction, but it would not be the independent replication the canon now calls for. The strongest practical next evidence is R1-B: same local scientific question and frozen margins, but a new implementation, new seeds, new governance, and new compute provenance. It will cost roughly the same order of computation as RC4 and should be planned as a substantial experiment, not a quick rerun.