01 · Where are we now?
“Real time” is a deadline,
not a vibe.
Five timing classes prevent a fast benchmark from being confused with a safe control loop. Case rows show what closes, what stays bounded, and what remains open.
- Hard real timeDeadline miss is system failure.
- Operational real timeProcess window; misses need containment.
- Interactive real timeHuman exchange keeps usable cadence.
- Near real timeResult arrives inside encounter or shift.
- Offline intelligenceWorkflow can pause before output.
Showing all six cases.
| Bounded case | Sense | Interpret | Predict | Decide | Act | Observe | Timing | Maturity: D / O / R / A / S |
|---|---|---|---|---|---|---|---|---|
| Sources: California DMV · NHTSA | ● | ● | ● | ● | ● | ◐ | operational-real-time | DORASdemonstrated:supported operational:supported reliable:unknown approved:supported scaled:unknown |
| Sources: FDA · pivotal trial · field study | ◐ | ● | ● | ● | ◐ | ◐ | near-real-time | DORASdemonstrated:supported operational:supported reliable:unknown approved:supported scaled:unknown |
| Source: MLCommons MLPerf v5.0 | ◐ | ◐ | ● | ◐ | ◐ | ○ | interactive-real-time | DORASdemonstrated:supported operational:unknown reliable:unknown approved:not-applicable scaled:unknown |
| Source: ICCV 2025 paper | ◐ | ◐ | ● | ◐ | ● | ◐ | interactive-real-time | DORASdemonstrated:supported operational:unknown reliable:unknown approved:not-applicable scaled:unknown |
| Source: MLPerf Inference definitions | ◐ | ● | ● | ○ | ○ | ○ | operational-real-time | DORASdemonstrated:supported operational:unknown reliable:unknown approved:not-applicable scaled:unknown |
| Source: Amazon Science / IROS workshop | ● | ● | ● | ● | ● | ◐ | operational-real-time | DORASdemonstrated:supported operational:supported reliable:unknown approved:not-applicable scaled:supported |
Complete case passports
Geofenced autonomous driving
- Boundary
- Dynamic driving task inside a permit-defined operational design domain; not unrestricted driving.
- Scope
- California public roads; permitted manufacturers and vehicles
- Period
- Permit status through May 2026; federal crash-reporting framework amended through 2025
- Deadline
- Vehicle-control cycle before road state invalidates the action; exact deadline is system-specific
- Timing class
- operational-real-time
- Intended use
- Navigate, control speed, and respond to road users within approved location, weather, speed, and operating constraints
- Unit / denominator
- Permit categories and reportable incidents; NHTSA reports are not normalized by miles, fleet size, or ODD exposure
- Human fallback
- Remote assistance, minimal-risk stop, passenger or fleet support, and emergency services; exact design is operator-specific.
- Failure boundary
- Outside approved geography, weather, speed, road type, sensor visibility, or system capability; rare interactions and incident recovery.
- Evidence type
- observation
- Interpretation
- Public-road operation demonstrates a bounded closed loop and regulatory authorization in named domains.
- Caveat
- Permits are not proof of comparative safety. Raw report counts are not incident counts or comparative safety rates; reports can be duplicated and lack exposure denominators.
Autonomous retinal screening
- Boundary
- Detect more-than-mild diabetic retinopathy from specified retinal images in eligible adults with diabetes; not treatment or a general eye examination.
- Scope
- United States; indicated primary-care patients with diabetes and no prior diabetic retinopathy diagnosis
- Period
- FDA De Novo decision 11 April 2018; pivotal trial published 2018; Johns Hopkins deployment comparison 2019–2021
- Deadline
- Near-real-time result during a primary-care encounter; acquisition and quality retries remain part of workflow
- Timing class
- near-real-time
- Intended use
- Return an autonomous screening result or image-quality failure and route indicated patients to eye care
- Unit / denominator
- 900 enrolled pivotal-trial participants; deployment study included more than 17,000 adult patients with diabetes
- Human fallback
- Repeat image acquisition, clinician review, referral to an eye-care professional, and usual care when no result is returned.
- Failure boundary
- Ineligible patient, inadequate image quality, disease outside intended use, missed disease, false referral, or failed downstream follow-up.
- Evidence type
- observation
- Interpretation
- A narrow medical loop has authorization, prospective diagnostic validation, and documented health-system operation.
- Caveat
- Authorization and diagnostic accuracy do not prove completed referral, treatment, long-run outcomes, generalization, or scale.
Interactive LLM response
- Boundary
- Generate and stream text for a submitted prompt; not verify truth, complete external work, or own consequences.
- Scope
- Benchmark systems and datasets; not a population estimate
- Period
- MLPerf Inference v5.0 results published April 2025
- Deadline
- Interactive benchmark constrains time to first token and time per output token; application tolerance varies
- Timing class
- interactive-real-time
- Intended use
- Preserve conversational response cadence under benchmarked server load
- Unit / denominator
- Queries and generated tokens under MLPerf workload, model, accuracy, latency, and submission rules
- Human fallback
- User correction, source inspection, retry, tool confirmation, or escalation to a responsible person.
- Failure boundary
- Unsupported request, hallucination, unsafe output, excessive latency, context loss, tool failure, or unobserved downstream effect.
- Evidence type
- observation
- Interpretation
- Latency-constrained generation is measurable jointly with benchmark quality requirements.
- Caveat
- A compliant benchmark run does not establish factuality, field reliability, task completion, approval, or product-wide service quality.
Streaming image generation
- Boundary
- Continuously update generated imagery from changing input on specified GPU pipelines; not general visual truth or production reliability.
- Scope
- Research hardware and benchmark inputs; not a human population
- Period
- Peer-reviewed ICCV publication, 2025
- Deadline
- Frame or update cadence fast enough for documented interactive generation on tested hardware
- Timing class
- interactive-real-time
- Intended use
- Provide continuous visual feedback while a user changes prompts or input streams
- Unit / denominator
- Frames or images per unit time on reported model, pipeline, resolution, and hardware
- Human fallback
- Pause, regenerate, edit manually, reject output, or switch to offline rendering.
- Failure boundary
- Hardware mismatch, quality collapse, temporal inconsistency, unsafe content, prompt drift, or missed interaction cadence.
- Evidence type
- experiment
- Interpretation
- Pipeline-level research demonstrates hardware-specific interactive diffusion generation.
- Caveat
- Demo throughput does not establish stable quality, safety, rights clearance, operational service, or scale.
Real-time object identification
- Boundary
- Classify or detect objects from benchmark-defined images; not open-world scene understanding.
- Scope
- ImageNet-2012 and OpenImages benchmark samples; not geographic or population inference
- Period
- MLPerf Inference v5.0 benchmark definitions, 2025
- Deadline
- Model- and scenario-specific latency constraint; ResNet50 server scenario specifies 15 ms
- Timing class
- operational-real-time
- Intended use
- Measure image classification or object detection throughput and latency while meeting stated accuracy targets
- Unit / denominator
- Benchmark queries on fixed validation datasets, quality targets, load scenarios, hardware, and software
- Human fallback
- Reject low-confidence output, stop downstream actuation, request another view, or send to human review.
- Failure boundary
- Open-set object, domain shift, poor lighting, occlusion, adversarial input, sensor fault, or deadline miss.
- Evidence type
- observation
- Interpretation
- A benchmark can demonstrate latency plus accuracy for a fixed task and dataset.
- Caveat
- Fixed-dataset accuracy does not prove field reliability under distribution shift, occlusion, novel classes, or sensor failure.
Warehouse sorting
- Boundary
- Pick and singulate packages from unstructured piles in a named robotic induction fleet; adjacent transport, packing, maintenance, and exceptions remain separate.
- Scope
- Amazon Robotics production facilities and handled packages; locations not enumerated in source
- Period
- Deployment evidence reported at IROS workshop in 2023
- Deadline
- Operational cycle must keep pace with induction flow; source reports daily fleet volume rather than per-item deadline
- Timing class
- operational-real-time
- Intended use
- Select feasible picks, manipulate packages, and route failed attempts within robotic induction
- Unit / denominator
- Vendor-authored extended abstract reports fleet handling up to 6 million packages per day and more than 2 billion cumulative manipulations
- Human fallback
- Operators handle damaged, unsafe, ungraspable, misrouted, or repeatedly failed packages and maintain equipment.
- Failure boundary
- Deformable or unknown package, clutter, bad grasp, damage risk, congestion, hardware fault, or downstream routing exception.
- Evidence type
- observation
- Interpretation
- Vendor research documents scaled production use of learned pick-success estimation in one bounded warehouse loop.
- Caveat
- Vendor-authored aggregate volume lacks independent reliability, exception-rate, economics, worker-outcome, and site-level denominators.
Independent gates: Demonstrated, operational, reliable, approved, and scaled answer different questions. Unknown is not failed. Approval may be N/A.
02 · What direction are we moving?
Case passports show
which gate evidence supports.
A green gate never spills into its neighbour. Authorization does not turn unknown field reliability into “reliable”; scale does not certify safety.
Authorized, bounded operation
Supported: demonstrated, operational, approved.
Unknown: field reliability and scale under comparable denominators.
Authorized diagnostic loop
Supported: demonstrated, operational, approved.
Unknown: long-run field reliability and broad scale.
Latency benchmark
Supported: demonstrated.
Unknown: operation, reliability, scale. N/A: approval.
Peer-reviewed demonstration
Supported: demonstrated.
Unknown: operation, reliability, scale. N/A: approval.
Fixed-dataset benchmark
Supported: demonstrated.
Unknown: operation, reliability, scale. N/A: approval.
Vendor-documented production loop
Supported: demonstrated, operational, scaled.
Unknown: independent reliability. N/A: approval.
03 · Who benefits or carries cost?
A task loop is not
a whole position.
Automatable loop
Only stages inside stated task boundary.
Possible position effects
Demand, workflow, authority, economics, and exception load mediate staffing.
Retained human work
Goals, context, consent, preparation, verification, care, maintenance, and accountability.
Exception handling
Out-of-domain cases, unsafe states, appeals, damaged inputs, and recovery.
Affected groups
Operators, patients, road users, workers, customers, bystanders, and data subjects.
Evidence basis: ILO–NASK task exposure index, 2025 ↗ · exposure is not observed job loss · Scope & caveats ↓
Position-removal threshold: full-loop coverage plus reliable operation, authority, integration, economics, and exception handling. Below that threshold, describe task change, exposure, or augmentation.
04 · Competing explanations
Model gains are
one system component.
- Constrained environment
- Hardware
- Sensors
- Human fallback
- Integration
- Regulation
- Model gains
Evidence basis: NIST AI RMF · MLPerf definitions · risk scope ↓ · benchmark caveat ↓
05 · What would change our conclusion?
Upgrade with field denominators.
Downgrade on hidden boundaries.
- Repeated field evidence reports deadline misses, failures, interventions, and recovery.
- Independent evaluation covers intended populations, domains, shift, and rare conditions.
- Authority, integration, economics, and exception handling hold across sustained volume.
- Downstream outcomes improve without shifting hidden cost or risk.
- Permit, benchmark, or demo is generalized beyond its operating domain.
- Latency improves while accuracy, safety, or outcomes worsen.
- Human intervention, exception work, or exposure denominators disappear.
- New conditions produce correlated failures, unsafe actions, or responsibility gaps.
06 · New-problem feedback map
Faster action creates
faster failure paths.
Late safe answer can become unsafe action.
Next input can leave intended-use distribution.
Observation loops can expand collection about people and places.
Sensors, models, interfaces, and actuators add attack surfaces.
Coupled systems can propagate one error before review.
Operator, deployer, vendor, and authority roles can diverge.
Reliance can weaken fallback capability.
Cheaper decisions can increase total actions and exception volume.
Evidence basis: NIST AI RMF 1.0 ↗ · risk nodes are editorial applications, not NIST maturity ratings · Scope & caveats ↓
07 · What might this make possible?
Possibilities stay
labelled and contestable.
Smaller local models may close more useful loops by reducing network latency and data transfer.
Exception-handling quality may become a stronger bottleneck than inference speed.
Jobs may reorganize around supervising several bounded loops rather than disappear whole.
Which actions should remain slow enough for contestability, consent, or human judgment?
Who bears proof and liability when a system acts before affected people can intervene?
Method & full source ledger
Scope travels
with every claim.
Sense → Interpret → Predict → Decide → Act → Observe must close before environment invalidates action.
Benchmark accuracy ≠ field reliability. Authorization ≠ scale. Scale ≠ safety.
Unknown is not failed. Approval may be N/A where no task-specific regime applies.
Monthly editorial replay. Curated JSON only; no visitor-time evidence calls.
Evidence vocabulary: observation · exposure estimate · association · experiment · forecast · scenario · editorial judgment · hypothesis · normative question
California DMV · Autonomous Vehicle Permit Holders
Scope: California permit categories and named operating constraints; reviewed 26 July 2026.
Caveat: permit status is not a safety rate, reliability verdict, or proof of unrestricted autonomy.
Open source ↗NHTSA · Standing General Order on Crash Reporting
Scope: reportable US ADS and Level 2 ADAS crashes; amended through 2025.
Caveat: possible duplicate reports and no common mileage, fleet, or ODD exposure denominator.
Open source ↗FDA · IDx-DR De Novo summary, DEN180001
Scope: indicated US adults with diabetes; retinal diagnostic software authorized 11 April 2018.
Caveat: authorization does not prove follow-up, outcomes, broad generalization, or scale.
Open source ↗npj Digital Medicine · IDx-DR pivotal trial
Scope: 900 enrolled people at 10 US primary-care sites; prospective diagnostic study.
Caveat: eligibility, enrichment, acquisition, and reference workflow limit generalization.
Open source ↗npj Digital Medicine · Johns Hopkins field deployment
Scope: more than 17,000 largely metropolitan adults with diabetes; 2019 versus 2021.
Caveat: retrospective association does not prove causality, completed referral, clinical outcomes, or broad scale.
Open source ↗MLCommons · MLPerf Inference v5.0 results
Scope: submitted systems under defined LLM workload, quality, TTFT, TPOT, and compliance rules.
Caveat: benchmark compliance does not establish factuality, field reliability, task completion, or service quality.
Open source ↗ICCV 2025 · StreamDiffusion
Scope: specified diffusion pipelines, inputs, resolutions, and GPU hardware.
Caveat: research throughput does not prove operational reliability, safe content, rights clearance, or scale.
Open source ↗MLCommons · MLPerf Inference definitions
Scope: fixed vision datasets, accuracy targets, scenarios, hardware, and software; v5.0 definitions.
Caveat: benchmark accuracy does not prove open-world reliability, sensor robustness, downstream decisions, or safety.
Open source ↗Amazon Science / IROS workshop · Robin pick planning
Scope: vendor-reported production fleet, up to 6 million packages per day and more than 2 billion cumulative manipulations.
Caveat: aggregate vendor claims lack independent site, reliability, intervention, economics, and worker-outcome denominators.
Open source ↗ILO–NASK · Refined global occupational exposure index
Scope: global employment mapped through tasks and harmonized microdata; 2025.
Caveat: modeled exposure is not adoption, full-loop automation, observed displacement, or forecast timing.
Open source ↗NIST · Artificial Intelligence Risk Management Framework 1.0
Scope: cross-sector lifecycle risk management, intended use, oversight, impacts, and emergent risk.
Caveat: voluntary guidance does not certify any case, maturity flag, reliability level, or legal compliance.
Open source ↗