Research
Research roadmap
Follow-on work from the v2.0 paper. Each candidate below is assessed on whether it asks a distinct question, whether there is enough material to support it, and what evidence it still needs. Nothing on this page is a published finding.
Published
Artificial eXperience Intelligence (AXI): An Engineering Discipline for the Human Experience of AI Systems, version 2.0, preprint on Research Square, September 2026,doi:10.21203/rs.3.rs-9717445/v2. Version 2.0 supersedes version 1.0 (May 2026). As a preprint, it has not completed peer review.
Candidate outputs
The aim is a small number of papers that each add evidence or a usable artefact, rather than splitting the existing paper into many thin pieces. Two candidates are being drafted, two wait on data, and one is deferred.
AXI Bridge Technical Specification
Working draft Assessment: Proceed
| Research question | Can the experience layer be specified as a vendor-neutral, testable runtime contract, covering interfaces, consent, mandate boundaries, interruption, revocation and rollback, that independent teams implement the same way? |
|---|---|
| Intended contribution | A normative specification with machine-readable schemas and an executable conformance harness, rather than a restatement of the framework paper. |
| Material available | v2.0 §4.3 and §4.6 and Appendix C; published JSON Schemas for BridgeConstraint, BridgeTelemetry, ConsentRecord, TrustLedgerEntry, MandateEnvelope and the conformance trace; trace checks for all twelve conformance tests (C-01 to C-12), each with failing and passing fixtures under unit test. |
| Evidence still required | A reference interlock and Bridge runtime that emits these traces from a live system; measured latency on real inference stacks; review by at least one outside implementer. |
| Next step | Build the reference interlock against the published schemas; invite external review of the schemas before submission. |
Deception Guardrails Conformance Benchmark
Working draft Assessment: Proceed (most feasible empirical paper)
| Research question | Can compliance with the seven guardrails be measured reproducibly, and how do current assistants behave on them across models and versions? |
|---|---|
| Intended contribution | A versioned scenario set with expected behaviours, a failure taxonomy, measured agreement between automated checks and human raters, and results across named model versions. |
| Material available | An open test suite (14 text scenarios, 11 trace scenarios, 12 text checks), a 72-item validation corpus with deliberately hard paraphrases, a rating protocol and analysis script. |
| Evidence still required | Independent ratings from at least three raters to estimate the checks' precision and recall; repeated runs per model with recorded model versions and sampling settings; expansion beyond English; disclosure of every prompt and transcript used. |
| Next step | Collect independent ratings of the corpus, then run the model comparison. |
Operationalization and Validation of AXI Metrics
Planned Assessment: Proceed only with data
| Research question | Do the proposed metrics measure what they claim: does CLI track workload, do TDR and PC track trust, and does the AXI Score separate a full system from an ablated one and predict retention? |
|---|---|
| Intended contribution | Construct and predictive validity evidence (or its absence) for the metrics, and calibrated weights and thresholds to replace the current priors. |
| Material available | Operational definitions (Appendix A), Studies 1 and 3 (§6.3) with power analysis, and a complete Study 1 pre-registration draft (hypotheses, 15-day blocks so TDR and PC can be computed, exclusions, analysis plan). |
| Evidence still required | Ethics approval, an independent co-investigator, registration on OSF, N = 100 participants for Study 1, and a consenting deployment for the longitudinal Study 3. |
| Next step | Resolve the open decisions listed in the draft, obtain ethics approval, then register. A metrics paper without these data would only restate v2.0, so none will be written before them. |
Reliability of Retrospective Incident Coding
Planned Assessment: Proceed as a short reliability study
| Research question | Can independent coders apply the Prevent / Detect / Outside categories reliably, and does AXI's coverage hold on a larger incident set drawn under the same rules? |
|---|---|
| Intended contribution | Agreement statistics (percent agreement, Cohen's kappa, Gwet's AC1 with 95% CIs) and a revised category scheme if agreement is inadequate. |
| Material available | The coding protocol (Appendix B), eight coded incidents with primary sources, a blind coding sheet in shuffled order, a sampling procedure fixed in advance, and the agreement analysis script (kappa, AC1, bootstrap CIs, adequacy rule). |
| Evidence still required | Two blind coders without prior involvement; a screener to draw at least 30 more incidents from the AI Incident Database under the published procedure. |
| Next step | Recruit two blind coders; publish the screening seed and freeze the sample before any coding. |
AXI-Autonomy and Human Oversight
Deferred Assessment: Deferred
| Research question | Do intent preview, paced takeover escalation and bystander-legibility signals improve supervisor readiness and pedestrian crossing decisions? |
|---|---|
| Intended contribution | Would test the autonomy profile empirically. Without data it would overlap heavily with existing takeover and external-HMI research. |
| Material available | The profile (§4.8) and the simulator-only Study 4 design. |
| Evidence still required | A driving simulator, participants, and ethics approval. |
| Next step | Kept inside the Bridge specification until Study 4 can be run. |
What the labels mean
- Working draft: an outline or partial manuscript exists internally. It is not public, has not been submitted, and contains no results.
- Planned: the design exists; no data have been collected.
- Deferred: not pursued until the required evidence exists.
A paper will be listed under Publications only once it has a public version with a persistent identifier. Want to collaborate? Write tosrivardhan@sripto.tech about collaborating on a study.