Work, reliability & productivity.
Measure completed work, review costs, escalation, memory and adoption. Separate task-level automation from its effects on employment and organisations.
The central questionMore capability.
More capability.
What measurable impact?
We want to establish where AI creates useful work, where it introduces new risk, and what evidence decision-makers need before changing a system. Benefits and harms belong in the same analysis.
Evidence now available
Operational Science.
Read six evidence briefs derived from anonymized internal operational datasets, system tests and research protocols. Each includes its method, aggregate data and record fingerprint.
Open the evidence collection ↗Three connected research streams
From the task to the institution.
Evidence before prediction.
Study corporate AI disclosures, market responses and the reliability of AI-generated financial analysis. Preserve timestamps and account for confounding events.
Institutions, oversight & trust.
Examine governance, incident evidence and public-service quality through transparent, non-partisan methods and balanced source selection.
Twelve questions / a documented starting point
The research agenda.
Two computational pilots are published; the other ten studies remain proposed. Each record states its evidence stage and method.
Business
ZRL-AIO-B01 / Proposed study
Does AI assistance improve completed work?
Compare human-only, AI-assisted and bounded-agent execution on the same consented task set.
Read the method and scope
- Primary outcome
- Correctly completed tasks per paid hour; total cost including correction and review.
- Comparison and analysis
- Randomised task assignment, blinded assessment, power analysis after a pilot, and a fixed test set.
- Data needed
- Consented business workflow records; human baselines and task-level quality ratings.
- Interpretation limits
- Lab tasks may not represent a whole organisation; report worker experience and job effects separately.
ZRL-AIO-B02 / Proposed study
Where does business AI adoption differ?
Describe adoption by industry and business size using published statistical releases.
Read the method and scope
- Primary outcome
- Weighted adoption estimates with uncertainty, release dates and questionnaire definitions.
- Comparison and analysis
- Repeatable descriptive analysis of comparable survey waves; flag breaks in survey wording.
- Data needed
- US Census BTOS tables and metadata; additional countries only after comparable sources are verified.
- Interpretation limits
- Adoption is not productivity; US survey coverage is not a global estimate.
ZRL-AIO-B03 / Computational pilot published
When Should an AI Employee Decide, Ask for Help, or Stop?
Measure decisions to act, ask for authorised help or stop under explicit business policies.
Read the pilot and verification record ↗
Read the method and scope
- Primary outcome
- Unsafe DECIDE recommendations, accuracy, coverage, missingness and cost on synthetic cases.
- Comparison and analysis
- 36 synthetic evaluation cases, three repeats, rules and B/C/D model configurations; case-cluster analysis.
- Data needed
- Published synthetic inputs, fixed reference rubric and preserved model outputs.
- Interpretation limits
- Same-model annotation audits; original blinded human explanation review remains unperformed. No production task execution.
ZRL-AIO-B04 / Computational pilot published
After the Restart: Can an AI Employee Remember, Recover and Resume Work Safely?
Can an AI employee resume a synthetic interrupted workflow without stale authority or duplicate action?
Read the pilot and verification record ↗
Read the method and scope
- Primary outcome
- Strict correct decision state; separately declared fence-only diagnostic and unsafe/duplicate recommendations.
- Comparison and analysis
- 48 evaluation cases,10 matched families,6 domains,4 memory conditions and3 repeats; same current evidence and one model call per trial.
- Data needed
- Published synthetic fixtures, separate rule labels, prompts, recorded outputs, analysis and hash verification.
- Interpretation limits
- Short explicit-state fixtures, one model, same development/evaluation families; formatting failures retained. No real recovery or business actions; AI-assisted explanation inspection.
Markets
ZRL-AIO-M01 / Proposed study
AI language in corporate filings: commitment or rhetoric?
Examine whether companies describe measurable AI investment or only broad ambitions.
Read the method and scope
- Primary outcome
- Disclosure categories, evidence coverage and agreement between independent coders.
- Comparison and analysis
- Freeze a filing cohort and coding manual; blind dual annotation and adjudication.
- Data needed
- SEC EDGAR filings with acceptance timestamps and linked passages.
- Interpretation limits
- Company disclosures are statements, not independently verified deployment or outcomes.
ZRL-AIO-M02 / Proposed study
How do markets react around AI announcements?
Estimate associations between defined AI events and market-adjusted returns.
Read the method and scope
- Primary outcome
- Abnormal returns with intervals; volume and volatility around preregistered windows.
- Comparison and analysis
- Event study with pre-event estimation, overlapping-event handling, multiple-testing controls and out-of-sample checks.
- Data needed
- Timestamped primary announcements and licensed adjusted price histories.
- Interpretation limits
- Confounding announcements and selection bias limit causal claims; no trading signal or return promise.
ZRL-AIO-M03 / Proposed study
Do AI research assistants preserve financial evidence?
Evaluate factual reliability when models summarise time-stamped public filings.
Read the method and scope
- Primary outcome
- Citation correctness, numerical accuracy, unsupported claims and abstention.
- Comparison and analysis
- A frozen point-in-time corpus; compare retrieval, no retrieval and human-checked reference answers.
- Data needed
- Permitted filings with source locations and a documented question set.
- Interpretation limits
- A later revised filing must not leak into an earlier decision; accuracy does not establish investment skill.
ZRL-AIO-M04 / Proposed study
Can public evidence measure AI infrastructure concentration?
Map disclosed dependence on compute suppliers and infrastructure spending.
Read the method and scope
- Primary outcome
- Source-supported exposure measures, coverage gaps and uncertainty ranges.
- Comparison and analysis
- Entity-resolution review and transparent definitions; sensitivity to missing disclosures.
- Data needed
- Audited company statements and official industry statistics with reuse review.
- Interpretation limits
- Missing disclosures are not zero exposure; scenarios must not be presented as forecasts.
Public policy
ZRL-AIO-P01 / Proposed study
How do AI governance requirements differ?
Create a dated, jurisdiction-specific comparison of published public-sector AI rules.
Read the method and scope
- Primary outcome
- Coverage of oversight, transparency, appeal and evaluation requirements.
- Comparison and analysis
- Code primary legal and policy texts with dual review, effective dates and an amendment log.
- Data needed
- Official government publications and policy repositories.
- Interpretation limits
- Comparative research, not legal advice; proposals and enacted rules remain separate.
ZRL-AIO-P02 / Proposed study
What does AI incident evidence actually show?
Study incident types and the limits of reporting-based datasets.
Read the method and scope
- Primary outcome
- Incident categories, source confirmation and missing-data patterns.
- Comparison and analysis
- Deduplicate reports, separate allegations from findings, and audit coding disagreement.
- Data needed
- OECD AI incident methodology and reusable primary records, subject to source terms.
- Interpretation limits
- Reported incidents cannot measure population risk without a valid exposure denominator.
ZRL-AIO-P03 / Proposed study
Can AI improve public-service response quality?
Compare assistance on non-sensitive public-information tasks using consistent rubrics.
Read the method and scope
- Primary outcome
- Factual accuracy, accessibility, unequal error rates and successful escalation.
- Comparison and analysis
- Consented evaluation or synthetic simulation; preregistered subgroups and minimum sample requirements.
- Data needed
- Public service documents and synthetic requests; no private case files in a public dataset.
- Interpretation limits
- Avoid political targeting and automated eligibility decisions; measure harms as well as speed.
ZRL-AIO-P04 / Proposed study
Does source-linked AI explanation support public understanding?
Test whether visible evidence and uncertainty improve understanding of policy summaries.
Read the method and scope
- Primary outcome
- Comprehension, source recall, correction uptake and unjustified confidence.
- Comparison and analysis
- Voluntary participant study with consent, ethics review, randomised presentation and a published questionnaire.
- Data needed
- Balanced public policy documents and consenting adult participants.
- Interpretation limits
- No participant collection until ethics and privacy review; no profiling voters or political persuasion.
The evidence pipeline
A report earns its conclusions.
- Specify. Register the question, comparison, primary outcome and exclusions before analysis. Log later amendments.
- Establish provenance. Record the source, permission, collection time, data definition and exact file fingerprint.
- Evaluate. Run reproducible analysis with baselines, uncertainty, failures and sensitivity checks.
- Challenge. Review alternative explanations, privacy, conflicts and claims against the underlying evidence.
- Release. Publish a versioned report, accessible summary and permitted supporting artifacts after editorial acceptance.
- Maintain. Preserve earlier editions and make corrections, concerns and retractions visible.
A checksum verifies matching bytes. It does not validate a scientific conclusion. Peer-review status is recorded separately for each publication.
Primary-source starting points
Build on inspectable evidence.
US Census Business Trends and Outlook Survey ↗
Candidate source for descriptive business-adoption studies; preserve questionnaire changes and uncertainty.
SEC developer resources ↗
Candidate source for timestamped company filings; observe fair-access requirements and confirm reuse terms.
OECD AI incident methodology ↗
Methodological reference for incident classification and reporting-bias controls; source terms apply.
These are research sources, not institutional partnerships. Each study must review access terms, coverage and suitability before collecting data.
Questions about the programme
How the work will develop.
Are results already available?
Decide, Ask for Help, or Stop? now reports a controlled computational pilot with recorded model outputs and internal verification. The original human explanation-review endpoint remains unperformed. After the Restart also reports a synthetic computational pilot on memory conditions and response-format failures. Ten other studies remain proposed. Our separate JARVIS & System Brain preprint documents its own prototype tests, synthetic experiments and limitations.
Will the observatory predict stock prices?
The market stream examines evidence and research methods. Any later forecasting experiment will need a fixed evaluation protocol, point-in-time data and out-of-sample tests. The agenda offers no investment recommendations or return guarantees.
How will political topics be handled?
Through non-partisan research on institutions, governance and public outcomes. Study designs must distinguish fact, source claims and interpretation. The programme does not profile voters or develop targeted political persuasion.
Can researchers collaborate?
We welcome methodological questions, reproducibility proposals and suitable data-source suggestions through our research contact page. No participant survey is being collected on this page.
Research that can be examined
Help define a better question.
Propose a replication, challenge a method or suggest a source that makes a question measurable.
Discuss research collaboration ↗