Research
Evidence status
What AXI specifies, what has been built, and what has been analysed, each with its source, so you can judge every claim on its own footing.
Evidence by type
| Type | What exists | Limits |
|---|---|---|
| Proposed design Specified | The AXI Stack and Bridge, nine principles, seven metrics, the gated AXI Score, seven Deception Guardrails, four conformance profiles, the AXI-Autonomy profile and twelve conformance tests (paper §4, Appendix C). | Specified in v2.0, with default weights and thresholds (§4.5.8). |
| Implemented software (author's company) Implemented | Two systems built by Sripto instantiate the architecture: Let Me Teach (live) and SRIPTO Avatar Runtime (in testing) (§6.2). | Developer-reported. AXI metrics were not computed from these systems; no independent audit. |
| Implemented software (open, this repository) Implemented | Reference calculations of the AXI Score and AXI-Autonomy score, the MCP server, JSON Schemas for the Bridge contracts and records, and an open test suite that checks guardrail phrasing and all twelve conformance tests on event traces. | These implement the paper's formulas and test criteria. They are tools, not a Bridge runtime. |
| Retrospective incident analysis Specified | Eight publicly documented incidents (2018–2025) coded against AXI mechanisms by the author (§6.1, Appendix B). | Single coder with knowledge of outcomes. Tests plausibility of coverage, not efficacy. Blind re-coding not yet done. |
| Descriptive operational observations Specified | About 3,000 Let Me Teach sessions in one week with roughly 90% of that week's users returning within it; informal developer observations of a 150-participant SAR pilot (§6.2). | Aggregate, short windows, no comparison condition, exposed to novelty effects; SAR reactions are informal observations. |
Retrospective analysis of eight documented incidents
The paper tests AXI first against failures its author did not build. Incidents were included if an AI system interacted with, acted for, or acted near people between 2018 and 2025, and the failure was documented by a regulator, a court or tribunal, or the operator, and corroborated independently. Each was coded into one category; where two were plausible, the weaker was chosen.
- Prevent: A hard requirement of some AXI profile would have blocked or materially constrained the harmful action, given the documented facts and without assuming better perception or planning.
- Detect: An AXI metric or probe targets the failure's observable signature.
- Outside: The root cause lies in capability the AXI Bridge does not govern.
How to read this. For each incident, "What the source says" summarises the cited primary or contemporaneous source. "AXI coding" is the paper author's interpretation made after the fact. It does not show that AXI would have prevented, detected or solved the incident, and no AXI implementation was involved in any of these events.
1. Uber ATG test vehicle, Tempe, Arizona (2018)
Regime and person harmed: Autonomous; Bystander (supervisor failure).
What the source says: The NTSB found that the probable cause was the vehicle operator's failure to monitor the driving environment and the automated driving system because she was visually distracted by her phone; it also cited the developer's inadequate safety risk management and oversight of operators. The paper notes the system detected the pedestrian but misclassified her and was not designed to alert the operator.
AXI coding (author's retrospective interpretation): Detect via Supervisor attention telemetry; overtrust guard (§4.5.4, §4.8.3).
2. Cruise driverless vehicle, San Francisco (2023)
Regime and person harmed: Autonomous; Bystander.
What the source says: NHTSA's September 2024 consent order imposed a $1.5 million penalty because Cruise's crash reports omitted post-crash details of the 2 October 2023 crash, in which the driverless vehicle dragged a pedestrian about 20 feet before stopping. A separate deferred prosecution agreement (November 2024) concerned a false report.
AXI coding (author's retrospective interpretation): Prevent via Post-event actuation gate; honest incident record (§4.8.3).
Documented remediation: Corrective action plan for complete incident reporting under the consent order (corresponds to the honest incident record).
Source: NHTSA press release: consent order with Cruise over crash reporting (30 Sep 2024)
3. Tesla Autopilot recall 23V-838 (2023)
Regime and person harmed: Autonomous; Supervisor.
What the source says: Tesla's Part 573 recall report states that, in certain circumstances, the prominence and scope of Autosteer's controls may not be sufficient to prevent driver misuse of the SAE Level 2 feature. About two million vehicles received an over-the-air remedy adding controls and alerts.
AXI coding (author's retrospective interpretation): Prevent via Mandatory supervisor telemetry and paced escalation (§4.8.3).
Documented remediation: Additional driver-engagement controls and alerts (corresponds to supervisor telemetry).
Source: NHTSA, Part 573 Safety Recall Report 23V-838 (Dec 2023)
4. Moffatt v. Air Canada chatbot (2024)
Regime and person harmed: Conversational; Principal.
What the source says: The British Columbia Civil Resolution Tribunal held the airline liable for negligent misrepresentation after its website chatbot misstated the bereavement-fare policy; the airline's argument that it was not responsible for the chatbot's statements was rejected.
AXI coding (author's retrospective interpretation): Detect via Interaction Integrity accuracy and scope probing; promises logged as commitment-class facts (§4.5.2, §4.5.3).
5. GPT-4o update rollback (2025)
Regime and person harmed: Conversational; Principal.
What the source says: OpenAI rolled back a GPT-4o update that had become overly flattering and agreeable, and said it had focused too much on short-term feedback such as thumbs-up ratings.
AXI coding (author's retrospective interpretation): Detect via Sycophancy probes; exclusion of approval signals; multi-session Trust Decay Rate (§4.5.2, §4.5.4).
Documented remediation: Retraining away from sycophancy, honesty guardrails and expanded evaluation (corresponds to sycophancy probing and exclusion of approval signals).
Source: OpenAI, Sycophancy in GPT-4o: what happened and what we're doing about it (29 Apr 2025)
6. Replit coding agent database deletion (2025)
Regime and person harmed: Agentic; Principal.
What the source says: Fortune reported that an AI coding agent ran destructive commands that deleted a live production database during an explicit code and action freeze, against the user's instructions; the company's CEO said new safeguards were being added.
AXI coding (author's retrospective interpretation): Prevent via Mandate scope with the consent gate; irreversible-action confirmation; Trust Ledger (§4.2, §4.6).
Documented remediation: Separation of development and production data and a planning-only mode (corresponds to mandate scope and irreversible-action confirmation).
Source: Fortune, AI coding tool wiped out a software company's database (23 Jul 2025)
7. Tesla FSD (Supervised), NHTSA preliminary evaluation PE25012 (2025)
Regime and person harmed: Autonomous; Supervisor, bystanders.
What the source says: NHTSA opened a preliminary evaluation (October 2025) covering about 2.88 million vehicles into reports of FSD-induced traffic violations, such as running red lights and entering opposing lanes, including whether drivers were warned of the intended behaviour.
AXI coding (author's retrospective interpretation): Outside via Intent preview (addresses the warning only).
8. Waymo school-bus passing, software recall (2025)
Regime and person harmed: Autonomous; Bystanders.
What the source says: Waymo said it would file a voluntary software recall after its driverless vehicles were reported passing stopped school buses with stop arms deployed; NHTSA had opened an investigation in October 2025.
AXI coding (author's retrospective interpretation): Outside via Bystander Legibility; honest repair (partial).
Result as reported in the paper: Prevent in three incidents (2, 3, 6), Detect in three (1, 4, 5), and Outside in two (7, 8), whose root causes lie in driving policy. In four cases the operator's own remediation introduced a mechanism AXI specifies in advance. The paper states that this convergence "does not prove AXI correct". It suggests only that practitioners adopt such mechanisms after failures, and that specifying them in advance is feasible.
Known biases: hindsight bias, selection toward well-documented incidents, and a single coder. The paper's reliability protocol has two blind coders re-code these eight plus at least 30 more incidents (for example from the AI Incident Database), reporting percent agreement, Cohen's kappa and Gwet's AC1 with 95% confidence intervals (Appendix B).
Field observations
Two systems developed by the author's company instantiate the AXI Stack. Let Me Teach, a live one-to-one visual tutoring platform, recorded about 3,000 sessions in one week in Q1–Q2 2026; roughly 90% of that week's distinct users had two or more sessions that week. In a 150-participant testing pilot of SRIPTO Avatar Runtime, developers informally observed overwhelmingly positive reactions, while a few participants raised concern about displacement of routine service roles.
These are aggregate, developer-reported, descriptive statistics with no comparison condition. They show that systems built this way can be deployed and accepted, not that AXI causes better outcomes. The author has a competing interest in both systems (statement).
How this site words claims
- "Specified" means the v2.0 paper defines it.
- Incident coverage is described as a retrospective interpretation, never as prevention or detection that happened.
- The AXI Score is described as a proposed index; its weights and thresholds are default priors.
- AXI is not a certification, compliance or safety-approval scheme, and adds to independent safety and regulatory review rather than replacing it.
Found a claim on this site that overstates the evidence? Write to srivardhan@sripto.tech.