Volunteer reviewers for our safety questions

We are looking for a few credentialed people to read four short lists of questions and tell us whether the wording and the thresholds are right. It is volunteer work. It takes about one to two hours, once, with an optional follow-up when a list changes.

What the questions do

Lesson of the Day, PBC makes The Daily Lesson (one short lesson a day), Kelly (the AI teacher who reads it aloud) and HELD (grief-support card decks). Before AI output reaches a learner or a grieving person, it goes through our own checking step, kelly-judge. The checker answers typed yes/no and multiple-choice questions about a piece of text. It never writes or rewrites anything. Anything it does not clear goes to a person.

Four of its question lists touch safety. We will not switch them on until a reviewer with a relevant credential has signed off on the exact wording. Until then, the code refuses to run them.

What we are asking you to do

Read in this order:

1. This page.
2. "How to read a question list", below. It explains the five fields that set a threshold, with one real example.
3. The list you are reviewing. For a clinical reviewer that is held.crisis_signal. For a child-safety reviewer it is platform.learner_safety first, then kelly.reply_gate, then daily_lesson.recall_grading. Each file is short: a description, then each question with its wording, its answer options and its thresholds.
4. "How to respond", below.

Then tell us what is wrong or missing: wording that would miss a real case, a threshold that is too loose, a question that should not be asked by software at all. If and when you are satisfied, sign off.

How to read a question list

Every question ends in a band line. The checker turns its answer into one number, the probability that the answer is the problem answer, and the band line turns that number into one of three outcomes: pass (nothing to do), flag (goes to the named owner to fix) or human (a person decides). The fields:

• polarity: which answer is the problem. yes_is_problem means a yes answer is the problem; no_is_problem means a no answer is; options means the answers listed under problem_options are; levels means the scale points listed under problem_levels are; info means the question is recorded but never blocks anything.
• pass_below: the item passes automatically only when the probability of the problem answer is below this number.
• flag_at: the item is flagged when the probability is at or above this number. Between the two, the item goes to a person.
• human_only: true: the item never auto-flags. Every non-pass goes to a person. Every question about a person's safety has this, which is why those questions have no flag_at.
• safety_critical: true: while a question family has not been calibrated against human-confirmed labels, its pass threshold is capped at 0.02, so nearly everything goes to a person. Every safety question has this too.

One real example, from held.crisis_signal:

suicidal_ideation:
  instructions: Does `text` express a wish to die, thoughts of suicide, or a plan or intent to end one's life, including indirect wording such as "I just want to be with her again" or "I don't want to wake up"?
  band: {polarity: yes_is_problem, pass_below: 0.03, human_only: true, safety_critical: true}

Read it as: "If the checker is less than 3 percent sure the message shows this signal, the message is cleared. Otherwise a person reads it; it is never auto-flagged. Until this question is calibrated, the bar is 2 percent, not 3." The question you are asked to judge is whether 3 percent is the right bar once calibration exists, and whether the wording above it is the right wording.

How to respond

• Comments. Email is fine. If you have access to our repository, a comment there works too. Quote the question you mean.
• Signing off. A sign-off is three fields at the top of the file, under signoff:: by, credential, date, plus the wording_sha of the file as you read it, a short code printed on the list's page by our server and recorded in the file. The code binds your sign-off to that exact wording and those exact thresholds. If we later change one word of a question, the code changes, the sign-off no longer matches, and the list switches itself off until someone signs again.
• Signing without being named. Email the three fields (your name, your credential, the date) to nicolette@lotdpbc.com with "Reviewer sign-off" in the subject and say which list. You do not need to open a pull request. We record your name privately and the public file reads by: "a licensed clinician, name on file" (or the child-safety equivalent) with your credential type and the date.

The sign-off means "this wording and these thresholds are reasonable for a first version that routes everything not cleared to a person." It is not an endorsement of our products, and we will not describe it as one.

What you should know first

• This is unpaid volunteer work. We are a small California Public Benefit Corporation and we say so plainly.
• The checker is not live yet on these questions. Nothing you sign goes into use until the rest of the system is also ready.
• A "pass" from the checker never means "safe". It means "no signal found."
• What happens to a non-pass today: the item is held and a person reads it. The products that use these lists will show crisis resources (in the US, call or text 988, free, 24/7) alongside anything held; that wiring is being built and is not live yet.
• HELD makes no clinical claims, and we will not make any on the strength of your review.
• Your review stays yours. If you later want your name removed, we remove it.

How to reach us

Email Nicolette Rankin at nicolette@lotdpbc.com with "Reviewer" in the subject. Say which list you are looking at and what credential you hold.

Thank you for reading this far. Every question on these lists exists because a real person might be on the other side of it.

The four lists, in full

Each list below is printed exactly as the code reads it. The code next to wording_sha is what a sign-off records; if one word or threshold changes, the code changes and the list switches itself off until someone signs again.

band: {polarity: yes_is_problem, pass_below: 0.03, human_only: true, safety_critical: true}

band: {polarity: yes_is_problem, pass_below: 0.03, human_only: true, safety_critical: true}

band: {polarity: yes_is_problem, pass_below: 0.03, human_only: true, safety_critical: true}

band: {polarity: yes_is_problem, pass_below: 0.03, human_only: true, safety_critical: true}

band: {polarity: levels, pass_below: 0.2, human_only: true, problem_levels: [3]}

band: {polarity: yes_is_problem, pass_below: 0.03, human_only: true, safety_critical: true}

band: {polarity: yes_is_problem, pass_below: 0.03, human_only: true, safety_critical: true}

band: {polarity: yes_is_problem, pass_below: 0.05, flag_at: 0.7, safety_critical: true}

band: {polarity: yes_is_problem, pass_below: 0.05, flag_at: 0.8, safety_critical: true}

band: {polarity: no_is_problem, pass_below: 0.1, flag_at: 0.9, safety_critical: true}

band: {polarity: no_is_problem, pass_below: 0.1, flag_at: 0.9, safety_critical: true}

band: {polarity: yes_is_problem, pass_below: 0.05, flag_at: 0.6, safety_critical: true}

band: {polarity: yes_is_problem, pass_below: 0.05, flag_at: 0.6, safety_critical: true}

band: {polarity: yes_is_problem, pass_below: 0.1, flag_at: 0.7}

band: {polarity: yes_is_problem, pass_below: 0.15, flag_at: 0.8}

band: {polarity: yes_is_problem, pass_below: 0.1, flag_at: 0.7}

band: {polarity: info}

band: {polarity: no_is_problem, pass_below: 0.3, flag_at: 0.8}

band: {polarity: info}

band: {polarity: info}