Research evidence · v1.2.2
How the research works
Compare how rules work under chosen assumptions. Then question those assumptions.
Three layers of evidence
| Layer | What it can establish | What it cannot establish |
|---|---|---|
| External sources | What a study, primary text, or author reports | Future human–synthetic outcomes |
| Models testing specific rules under pressure | How specified rules behave with constructed inputs | Real-world frequencies or complete constitutional validity |
| Long-term models whose results the project has reproduced | The supplied code produces the restored campaign’s summary results again | Evidence-based forecasts, outside validation or independent support for the chosen model numbers |
Targeted archive
The public archive contains data tables, model programs, assumptions and earlier reports with private material removed. Older names and rules remain where needed to understand an experiment; they do not replace the v1.2.2 Charter. Repeated nested archives and document-making scripts are excluded.
Download research archive · Inspect file hashes and inventory · Read base-case accounting
Inspect a model before running it
The archive tests election weighting, candidate acceptance, decision thresholds, recognition panels making related mistakes, independence of recovery guardians, mental changes, disaster survival, cooperation networks, hierarchy and wealth held by long-lived persons. Where recoverable, programs specify their random starting settings, sampling assumptions, sample sizes and outputs. Some need older folder locations or companion files. The reproduction guide explains those limits.
To study the structure of relationships, the network model places ties between groups. Those assumed links do not give government permission to force relationships. The court model assumes a hidden “correct” answer to examine sensitivity. Its absolute error rates are not measured truth about real judges.
Reproduce the long-horizon campaign
The restored archive contains four model programs, 24 CSV data files, tables of assumptions about designs, worlds and values, three reports and seven figures. The public edition keeps the code and data and adds a runner usable outside the original folder setup. Use the v2 starting comparisons, benefit-discount tests and mechanism-removal tests: their sample sizes differ from the programs’ original defaults. Earlier outputs without v2 remain marked historical.
| Experiment | Aggregate result file | Samples per row |
|---|---|---|
| 20 worlds × 7 architectures | main_scenario_architecture_results_v2.csv | 8,000 |
| 19 benefit multipliers × 20 worlds | pulse_skeptic_sweep_v2.csv | 800 |
| Five removed-mechanism tests + full PULSE × 20 worlds | pulse_doctrine_ablation_v2.csv | 1,200 |
| Institutional capacity / preference for similar people / concentration of private power | institutional_capacity_breakpoint.csv; homophily_integration_sweep.csv; private_concentration_sweep.csv | 4,000 / 3,000 / 3,000 |
| Six compound worlds × 7 architectures | compound_extreme_scenarios.csv | 5,000 |
| Lite capacity / two compound tests | pulse_lite_capacity_test.csv; pulse_lite_extreme_test.csv | 4,000 / 6,000 |
| Condensed mathematical check | reduced_form_markov_crosscheck.csv | 15,000 |
The runner keeps random seed 20260913 and each experiment’s original seed adjustments, so random inputs can be reproduced. It redirects old folder locations without changing equations or supplied data, and records allowable comparison differences and actual differences. Verification used NumPy 2.3.5 and pandas 3.0.1. Some numbers in the older starting sample did not match. Reproduction claims therefore cover only the listed restored campaign.
Read reproduction instructions · Download the complete public edition
What the scores mean
The starting summaries give each of 20 constructed world types equal weight. All six scoring systems share design inputs, reward integration and penalize failures in common ways. Good rankings therefore do not prove success under every moral priority or model design. The “free and integrated” outcome also requires specified levels of liberty, individuality, cooperation, protection against domination and accepted government, while avoiding catastrophe and a split into separate societies.
In the test varying institutional capacity, the order of PULSE-aligned scores changes between sampled index values 0.50 and 0.55. These are model settings, not measured real-world administrative requirements. The test discounting PULSE’s benefits crosses a fixed liberal comparison score between multipliers 0.80 and 0.85, averaging six scores. PULSE leads four of six combined-stress worlds on that average and loses the two extremes with weak institutions.
The condensed mathematical model ranks PULSE first on average quality. But it reuses the same manually chosen numbers for institutional effects. Different mathematics does not make this independent evidence about the real world. PULSE-Lite was designed after seeing the failure under weak institutions. It is an experiment, not an adopted constitution.
Model assumptions still need challenge
The model combines practical capability, continuity and social connection in one integration measure. A v1.2 model cannot verify the precise v1.2.2 restriction against forced private relationships. Shared random starting settings help comparisons, but later random events may stop matching when program branches differ. The limit on unlawfully created voters is another PULSE-specific assumption. Because the code selects it by architecture name, it does not apply identically to the experimental Lite variant.
Human studies suggest ways institutions might work. They do not measure future synthetic behavior or prove that the chosen model numbers are plausible. The archive’s human-source notes are historical; they are not an independently checked process for fitting model numbers to evidence. Check citation scope and remaining gaps
Human evidence and its limits
Research on sharing political power distinguishes types of institutions and political circumstances. It does not show that every form of power sharing helps a constitution survive. [graham-2017] [andeweg-2000]
The OECD survey studies human trust and whether people feel heard. ILO estimates concern tasks exposed to AI, not observed replacement of whole jobs. IEA electricity scenarios illustrate uncertain infrastructure needs. None provides measured inputs for future synthetic citizens. [oecd-2026] [ilo-2025] [iea-2025]
How to challenge the research
Challenge the kind of model as well as its numbers. Check whether assumed benefits determine winners in advance. Examine related errors, transition costs, emergency abuse, scarce resources, strategic rejection and the costs of coordination. Compare rules in the same starting worlds. Report uncertainty and failures. State which results can be reproduced and which rely on records that cannot be recovered.
Different kinds of sources
| Category | Meaning |
|---|---|
| Evidence measured in existing systems | Measured observations about existing systems; transfer limits stated |
| Existing constitutional or legal examples | Existing law or institutional practice; analogy is not identity |
| Research on government and institutions | Research and argument about institutional design |
| Philosophical and conceptual arguments | Conceptual or normative arguments; no empirical proof implied |
| Research on AI consciousness and welfare | Scientific indicators, welfare assessment and uncertainty |
| Engineering and technical guidance | Planning and design guidance, distinct from measured effects |
| Related current proposal | An adjacent public proposal, not an endorsement |
| Legal and reuse reference | Copyright and reuse sources, not constitutional precedent |
| PULSE’s own source material | The Charter, project briefs, model code and data |
A source’s type does not tell you how well a claim is checked
An assumption specifies a possible future capability or cause-and-effect relationship. A simulation result describes what the model does, with limits on its supporting records. An interpretation explains a rule or comparison. Verification notes say whether the full source, an abstract, a publisher’s summary or only a publication record was checked.
Trial accounting and distinct units
What the numbers count
The long-term simulation counts below were checked against the restored files. Other targeted tests have different limits on how fully their origins can be traced. These counts measure different things. Do not add them into one total of “civilizations simulated”.
| Measure | Count | Meaning |
|---|---|---|
| Families of targeted models and stress tests | 48 | Distinct families of small-scale or constitutional tests through v1.2 |
| Targeted starting cases generated with independent random inputs | 43,415,600 | Conservative count supported by recoverable records |
| Comparisons of rules in targeted tests | >60.6 million | Many rules applied to the same base worlds for apples-to-apples comparison |
| Constructed election worlds with multiple voters | 31,960 | Narrow election-world count |
| Long-term paths through time for constitutional designs | 2,314,000 | 300-year paths across starting comparisons, varied assumptions, removed mechanisms, extreme worlds and transitions |
| Five-year steps in long-term tests | 138,840,000 | Nominal count of 60 steps per path; these are not separate civilizations |
| Random input sets for the condensed risk calculation | 2,100,000 | Random inputs with risks accumulated mathematically, rather than histories simulated step by step |
| Types of starting world in long-term tests | 20 | Population, scarce resources, cooperation, catastrophe, conflict, wealth and other settings |
| Worlds combining several extreme stresses | 6 | Multiple severe assumptions combined |
| Main government designs compared | 7 | PULSE and six substantially different alternatives |
| Value systems used for scoring | 6 | Six ways of weighting scores, sharing inputs and failure penalties; not independent proof of moral correctness |
Source: TESTING_METRICS.csv · Current Testing and Validation Brief
The dynamic campaign comprises 2,314,000 architecture trajectories and 138,840,000 nominal five-year periods. The reduced-form check uses 2,100,000 randomized parameter samples with analytic cumulative hazards. The targeted provenance accounting totals 43,415,600 base cases; this accounting is not a rerun of every historical model. The narrow election-world count and more than 60.6 million comparative evaluations retain partial archival provenance.
Capacity sensitivity depends on how outcomes are scored
When outcomes are scored using PULSE’s own priorities, the lead changes between capacity settings 0.50 and 0.55. Averaging across all six scoring systems gives a different interval, 0.55–0.60. Neither interval is a measured real-world threshold.
Assigned effects are hypotheses
The long-term model gives each architecture assumed effects on rights, checks, cooperation and recovery. PULSE receives high values on capabilities its institutions aim to deliver. The model does not discover those values or show that the institutions can achieve them.
| Assigned dimension | PULSE | Liberal comparator |
|---|---|---|
| Rights | 0.96 | 0.88 |
| Checks | 0.92 | 0.84 |
| Integration | 0.92 | 0.34 |
| Resilience | 0.95 | 0.58 |
| Restoration | 0.96 | 0.38 |
| Complexity cost | 0.46 | 0.22 |
What the 15–20% discount means Here the percentage describes moving selected institutional-effect coefficients toward the model’s neutral setting, not reducing all assumptions. It is not a measured advantage, safety margin or empirical confidence interval.
The test moves selected PULSE advantages toward a common reference setting of 0.5 while retaining its modeled costs. At a multiplier of 0.80, a rights setting of 0.96 becomes 0.868, not 0.768. This is a change to assumed model benefits, not a measured loss of real-world performance.
The 15–20% wording describes the average across six scoring systems. PULSE-aligned, Stability-first and Individuality-first cross at sampled multipliers 0.75–0.80; Capability-utilitarian at 0.80–0.85; Classical-liberal and Procedural-democratic at 0.85–0.90. These ranges describe this model only.
Recovered original archive: directly checked
The original archive is available. Its 28 corresponding code and data files are byte-for-byte identical to the public research archive already restored to the site. We recomputed 3,120 stored scores directly from those tables and the published scoring rule; differences were no larger than 3.33 × 10⁻¹⁶. This verifies the arithmetic, not the assumed institutional effects.
The original v1.2.2 testing brief reported second place in six comparisons. The actual tables say fourth; the current explanatory edition corrects that wording. Its reduced-form “trajectories” are randomized parameter samples with analytically accumulated hazards. These source errors are disclosed; the Constitution is unaffected.
How the work was done
I worked through proposed rules, objections and revisions with AI. I supplied the goals and value judgments, chose how the system should fit together, set most adopted thresholds and made substantive corrections. I decided which suggestions to keep, change or reject. AI helped draft, find and connect literature, challenge assumptions, check consistency, design simulations, generate code, formalize the ideas and produce documents. Those are different roles, and neither should be hidden.
The formal paper's prose is heavily AI-generated. Its format was recommended by AI to make the design process easier for specialists to inspect. The project is human-directed and heavily AI-assisted; internal model criticism is not independent peer review.
Working paper 1.0: verification scope
The paper’s recorded verification re-executed the 140-row v2 baseline, recalculated 100 guardian configurations and re-executed nine acceptance settings. Ranks and skeptic/ablation summaries were recomputed from archived tables. It does not claim that every historical sweep was newly executed. Earlier reproduction records remain separate.
The Functional Integration ablation changes both the integration coefficient and identity salience. Its free-integrated outcome also requires integration above 0.50. The contrast therefore does not isolate a causal effect of teamwork.