Infer safety issues when no regex safety issues are found

[1jamesthompson1]

Safety Issues are sometimes listed explicitly but other times they are just implied throughout the text.

This means that I will need to read all of the analaysis and findings to find them all.

Important now that there will be exact and infered safety issues the output data will need to show this.

This inference is required to make sure that every report has atleast one safety issue because all acidents will have safety issues.

[1jamesthompson1]

There is a problem with the requirement of all accidents will have a safety issues.

As I find them I will compile a list of reports that don't have any safety issues

• 2017_007 (Not listed but LLM will infer safety issues)
• 2020_001 (Explicitly state that no new safety issues were identified) LLM will return something like this:

• | The pilot's potential incapacitation early in the take-off roll, which prevented any corrective action to align with the strip centreline, poses a significant safety issue by highlighting the need for mechanisms to detect and respond to pilot incapacitation.
• | The undercarriage's vulnerability to damage from striking uneven ground, leading to the dislodgement of the undercarriage assembly and subsequent uncontrollability of the aircraft, identifies a safety issue related to aircraft resilience and emergency response capabilities upon encountering rough terrain during take-off.

• 2013_203 (not listed but LLM infers safey issues)

[1jamesthompson1]

After this commit ed58cee

I ran it all the way through and found that the safety issues were infered at a rate 3 times higher.

This means that i need to control the range a bit better

[1jamesthompson1]

This commit 294ec924c38c5945135f5cb554d6c09e4befa8ce fixes the problem of having really high amount.

Now the number look something like this

        number

quality
exact 1.947368 inferred 2.791667 number quality
exact 0.00 1.0 0.25 1.0 0.50 2.0 0.75 2.5 1.00 4.0 inferred 0.00 2.0 0.25 2.0 0.50 2.0 0.75 3.0 1.00 8.0

When ran on a rnadom subset of 50 reports.

Now the problem is actually whether the safety issues are correct

Some good examples to look at are: 2011_102 these are the produced safety issues

• | A planned change to the train controller's workload was not assessed and managed in a way that recognized its higher-than-usual level of risk.
• | Train control's workplace culture at the time of the incident did not encourage train controllers to ask for help.
• | The train controller, who was performing a safety-critical function with a higher-than-usual level of risk, was able to become highly stressed and mentally fatigued during her shift without anyone being aware of, or anticipating, this.
• | Minimal training and education about detecting and managing stress and fatigue were provided to the train controller (and other train controllers) before the incident, even though she was performing a safety-critical function.
• | The train controller received insufficient support and supervision during her shift even though she was performing a safety-critical function with a higher-than-usual level of risk.
• | Train control protocols required train controllers to work their shifts without any certainty of reasonable breaks.
• | Poor planning and coordination of track infrastructure activities unnecessarily exacerbated the complexity of the train controller's work on the day of the incident. Train controllers are responsible for coordinating and managing high volumes of routine and unplanned track infrastructure activities. This increases the complexity of their role. In these situations, mistakes can occur.
• | A large number of rail movements and track activities are not electronically visible to train control at any given time.

2017_007 these are the produced safety issues

• |

  The train controller did not check the location of the coal train and mistakenly authorized the track engineer to on-track his alicart and travel towards it on the same section of track, contrary to standard protocol.

• | The train controller's workplace culture did not encourage train controllers to ask for help, contributing to an environment that increased the likelihood of errors.
• | The flight crew descended below the published minimum safe altitude for a segment of the arrival procedure due to inadequate situational awareness.
• | The flight crew elected to use an 'open descent' procedure rather than the available, fully automated 'managed descent' mode, requiring a higher level of human intervention to keep the airplane within permissible limits on the arrival route.
• | The operator's procedures did not encourage the appropriate use of the airplane's automated navigation systems, increasing operational risk by placing more reliance on human performance.
• | The air traffic controller observed the airplane's descent below the minimum safe altitude but did not follow the required procedures to alert the flight crew until the airplane had landed.

2020_001 Even though this report explicitly states that there are no new safety issues the model being coerced come up with these.

• | The pilot's potential incapacitation early in the take-off roll, which prevented the necessary corrective actions to align with the airstrip centerline.
• | The undercarriage assembly's vulnerability to dislodgement upon striking uneven ground, leading to loss of aircraft control.