Observation source
A retained camera observation.
See what supports each decision.
Aspinos brings observations, site instructions, agent work and human decisions into one record. Inspect the current demonstration, its sources and its limits.
A useful result explains what changed, why it matters and what the evidence still cannot settle.
Source identity travels with the work
The evidence keeps its original class.
Combining inputs preserves their limits. A simulated contribution cannot become live evidence through presentation.
Supporting input
An authored supporting observation.
Simulated
The simulated input remains part of the result. Combining it with recorded material cannot turn the record into recorded evidence.
Explanatory adapter of the repository’s evidence-floor rule. These selectors do not connect sources or change a real record.
Retained experiment · 10 August 2026
Read the retained observations.
This historical recorded-video experiment produced anonymous track summaries. Inspect their approximate source times and the limits of the original run.
Recorded source Historical experiment
Track 244
- Approx. video time
- 21.8 s
- Reported dwell
- 2.1 s
- Raw model score
- 0.43
Anonymous track numbers associate observations within this run. They do not identify a person.
Conditions and limits
This historical run used a wall-clock dwell implementation and approximate video timestamps. Its labels were provisional; the clip has editorial cuts and no person-free interval. The model score is uncalibrated.
These summaries establish neither detection accuracy nor false-alert performance, deployed capability or current processing speed.
Inspect the retained summaries ↗Engineering and compute
Compute follows the work.
Aspinos connects observations, site context, agent work, permissions and the resulting record. We evaluate existing computers and GPUs against the sources and responsibility a site needs.
6.4 fps on a recorded stream.
The Aspinos team reported 6.4 frames per second while processing a 640 × 468 recorded crowd clip at 12 fps over RTSP, using an Apple M2 CPU under heavy desktop load. No discrete GPU was used.
Processing lagged the source and the loader dropped frames. This was recorded footage played through the streaming pipeline, not a commissioned camera installation.
Size the workload with the site.
Camera count, resolution, scene detail and competing workloads affect the compute needed. Sustained processing, response delay, source recovery and power use need their own measurements under the proposed site conditions.
This historical run does not establish multi-camera capacity, detection accuracy, false-alert performance or a lower deployment cost.
Read the method and development boundary
The 10 August 2026 engineering report describes recorded-video streaming through an RTSP server, a person-detection and anonymous-tracking worker, and observation intake backed by PostgreSQL. It was an early experiment on a developer’s MacBook under concurrent desktop load, not a controlled deployment benchmark. The figure above is attributed to that historical report; this website does not present it as a newly reproduced run.
A separate synthetic clip read directly from a file tested a different path. It is not used here to answer how the streaming pipeline kept up with a camera-rate source.
Aspinos’s current contribution includes site context, workflow orchestration, authority controls and inspectable records. The prototype demonstrates those responsibilities with simulated work; outside-agent observation and proposal contracts remain partial and staged. Further model development and adaptation are planned around defined responsibilities and require their own provenance, rights and evaluation evidence.
What Aspinos builds. What the record proves.
Software that coordinates the work.
Aspinos’s current software joins source observations, site context, agent tasks and named-human authority. The shared demonstration retains ownership, evidence, decisions and unfinished work across its operating views.
Memory with a traceable source.
Questions can refer to available observations, times, locations, current instructions and reviewed site notes. A correction adds context to the record; it does not retrain a model or silently change a policy.
Models need their own proof.
Record-based answers and configured checks work in the demonstration. Further image-and-language reasoning, model adaptation and learned site patterns need approved data, evaluated versions and their own implementation evidence.
Inspect the demonstrated result.
Five retained application captures show held, approved, refused and unanswered decisions in separate simulated rehearsals. Their case, date and source class remain visible. Recorded camera experiments are separate; neither proves a customer deployment.
Open the captured product states ↗Preventing avoidable harm is our mission. We do not claim that every threat can be recognised or that a particular tragedy would certainly have been prevented.
Improve a version. Prove the difference.
This is the proposed evaluation path for a defined site responsibility. Data access, labels and future model adaptations require agreement before the work begins.
1. Use data we have permission to use.
Agree licensed or site-approved sources, permitted uses, privacy and retention. Keep each sample’s origin. A prospective introduction does not provide footage access or training rights.
2. Review what each example establishes.
People check the labels, location, time and outcome. Preserve disagreement and missing context. A supplied note, an observed fact and a reviewed result remain distinct.
3. Keep the test examples unseen.
Hold back ordinary activity, harmless lookalikes, genuine concerns and source gaps from tuning. Test continuity and conflicting inputs, not just selected clips where the answer is already known.
4. Measure under stated conditions.
Record misses, unnecessary alerts, operator effort and checked outcomes against an agreed reference. Report the sample, camera hours, hardware and source conditions. Measure processing and end-to-end response separately.
5. Review the version before use.
Compare the proposed change on the same held-out examples. Review its failures, data provenance and operating limits before approving a version. A recorded correction alone cannot promote new model weights.
6. Check the work with the team.
Follow source loss, task delivery, acknowledgement and human authority through the whole job. Test recovery and unanswered work. Continue, revise or stop based on the checked result.
The compute result above is a historical team measurement. New evaluation results should include their method, relevant comparison, limitations and actual operating conditions.
See how authority fits the modelResearch informs the questions.
The cited studies explain why context, community reports and coordinated response matter. They do not evaluate or endorse Aspinos.
Understand behaviour in context.
The U.S. Secret Service's 2019 study of targeted school violence stresses relevant behaviour, circumstances and information rather than a demographic profile. Our fictional student-origin scenario makes the visible concern specific; appearance or a backpack does not establish danger.
Protecting America's Schools · 2019 (PDF)People contribute essential information.
The Secret Service's 2021 analysis of averted school attack plots examines intervention and the importance of community members reporting concerning behaviour. Our design treats camera observations as one input to a case, alongside context and human judgement.
Averting Targeted School Violence · 2021 (PDF)Protect the freedom to gather.
The Christchurch Royal Commission describes lasting effects on Muslim communities, their wish to feel safe and welcome, and the limits of information available before the attack. That context informs a dignified fictional community-safety story, without claiming Aspinos could have prevented the historical event.
Christchurch Royal Commission · Executive summaryThe studies concern defined events and settings. Their findings should not be treated as universal incident rates or as proof that a camera can infer intentions or identify a fully concealed object.
Agree what useful means at your site.
Start with one responsibility and a proposed eight-week evaluation. The scope and start date depend on agreed source access, site participation and permitted actions.
Bring the conditions.
Share the site layout, camera or sensor constraints, relevant workflows and the problem worth solving. Identify the people affected and those responsible for decisions.
Define the work and its authority.
Choose what an agent may observe, prepare or notify, what needs approval and what remains unavailable. Include normal activity and uncertain cases in the evaluation.
Review evidence together.
Inspect the trace and measured outcomes, including failures. Decide what to change, what to evaluate next and whether the results justify a wider deployment.