Day 1: Advanced Biometrics

Statistical Analysis of Fingerprint Evidence

Part I: The Hook

Woven notebook: open your notebook to start Part I, The Hook. Write your first reactions to today's Case Briefing. What does the case demand of you? What evidence will you need? Your notebook is the running record of your thinking from briefing to verdict.

Welcome to Day 1 of the High School STEM Workshop! Today you will analyze fingerprint evidence through a scientific and legal lens.

Fingerprint analysis has been used in courts for over a century, but how reliable is it really? Today you will examine the statistics behind biometric identification.

Today's Case Briefing: You step into the role of a forensic latent print examiner trainee at the FBI's Latent Print Unit. Your mission: classify fingerprint patterns and then audit a fingerprint AI against the Daubert Standard, the actual legal test for whether forensic evidence is admissible in court. The classification and audit skills you build today go into your toolkit for Day 4's Week 1 forensics case.

I read the case briefing and watched the ACE-V method video

Part II: The Physical Lab

Woven notebook: this is Part II, The Physical Lab. Record every measurement, calculation, and observation as you work. The lab data you capture here becomes the evidence base you defend in Part III.

Materials Needed: Ink pads, White cardstock, Magnifying glasses, Scientific calculators, Daubert Standard reference sheet, Frequency distribution worksheet, Tablets

Interactive App (fingerprint): Use the Fingerprint Ridge Classifier to practice identifying Loops, Whorls, and Arches. Study the reference patterns, then challenge yourself in Quiz Mode before moving to the statistical analysis. Use the Launch button below to open the app inline.

1Ink and Roll Technique: Collect prints from all ten fingers using the ink-and-roll technique.

2Classification: Classify each print and calculate the frequency distribution for your class.

3Population Comparison: Compare your class distribution to population statistics (Loop ~60%, Whorl ~35%, Arch ~5%).

4Probability Calculation: Calculate: If a crime scene print is a whorl, what is the probability of a random match in a city of 100,000 people?

I explored the Fingerprint Classifier and completed the statistical analysis

Part III: AI & Digital Literacy

Woven notebook: this is Part III, AI & Digital Literacy. Capture what each AI tool said, what you decided to trust, and what you flagged as wrong. Your notebook becomes the evidence trail for how you evaluated AI today, the same way a professional double-checks every AI output before they rely on it.

Bleeding-edge: the Daubert standard (1993) is being challenged in 2026 federal court by AI companies. They argue: if AI has a measurable error rate (unlike traditional fingerprint analysis), AI evidence should be MORE admissible than human analysis. The DOJ disagrees. The first ruling is expected June 2026 - exactly when this workshop runs. You are watching the law form in real time.

Train Your Own Fingerprint Classifier (Daubert-Grade Audit)

Real forensic AI auditors don't trust tools they didn't audit themselves, they build and test their own. Today you train a fingerprint classifier in Teachable Machine, then audit it against the criteria a Daubert expert witness uses: calibration, error rate, and admissibility.

Watch the 3 videos below FIRST. They walk you through Teachable Machine end-to-end so the lab time is for analysis, not setup.

1Open teachablemachine.withgoogle.com. Click 'Get Started' then 'Image Project' then 'Standard image model.' Create 3 classes named Loop, Whorl, Arch.

2Hold each of your inked prints up to your webcam. For each class, click 'Hold to Record' and capture 50+ samples (rotate the print, vary the lighting). More variety in training = better generalization.

No webcam? Upload your images instead

No camera on this computer? You can train the same model by uploading pictures instead of using a live webcam. On each class, click 'Upload' instead of 'Webcam', then add image files. Two ways to get the pictures:

Phone as a camera: take a clear photo of your inked fingerprints, one pattern at a time, with good light and a plain background. Send the photos to the computer by emailing them to yourself and downloading, or by putting them in Google Drive and opening Drive on the computer. Then click 'Upload' on each class (Loop, Whorl, Arch) and add that class's photos.

No phone? Ask your instructor for the shared print set, or use a partner's photos. Open the folder, click 'Upload' on each class, and add the Loop, Whorl, and Arch pictures.

3Add at least 10 to 15 pictures per class, then click 'Train Model'.

4Test it: click 'Upload' and add one new picture you did not use in training. Check the predicted class and the confidence percent. Did it get it right?

5Click 'Train Model' and wait. Then test the model on 10 NEW prints from your team that the model NEVER saw during training. Record the predicted class and the confidence percentage for each.

6Compute accuracy: how many of the 10 got the correct pattern? Report as a fraction. This is your model's headline accuracy number, the kind a Daubert hearing demands.

7Calibration check (the Daubert step that matters most): for the prints your model classified at >90% confidence, what was its actual accuracy? A well-calibrated model is ~90% accurate when it says 90% confident. If your model says 90% but is only right 60% of the time, it is OVERCONFIDENT, a red flag in court.

8AFIS (FBI's fingerprint database) reports its error rate as approximately 1 in 10,000 for verified high-quality matches. Your trained model is almost certainly worse. Write a 1-paragraph admissibility argument: should this model's output be allowed in court? Cite the Daubert criteria (testability, peer review, error rate, general acceptance).

Now: Vibe-Code a Stats Calculator

You audited your Teachable Machine model by hand. Now you build a real tool to do the math automatically. Gemini Canvas is a vibe-coding tool: you describe what you want in plain English and it builds a working web app on the right side of the screen. Watch the short demo below before you try the prompt yourself so you know where the buttons are. The video runs about 8 minutes, but the first 2 minutes show everything you need to see the Canvas in action; you can pause once you have the layout. Important: Gemini does not run reliably in Safari, so make sure you are in Chrome before you start.

9Vibe coding extension: open Gemini Canvas (gemini.google.com/canvas). Type this prompt: 'Build a single page web app. It has 10 rows. Each row has three inputs: actual class (Loop / Whorl / Arch), predicted class, and confidence (0 to 100). Below the rows, add a Compute button. When I click Compute, show overall accuracy plus a clean table with precision, recall, and F1 score per class. Use simple modern styling.' Canvas will build a live web page on the right side of the screen. Type your 10 results, click Compute, and audit the numbers. Did it compute precision and recall correctly? F1 should be 2 times (precision times recall) divided by (precision plus recall) - did the AI use that formula? AI-generated stats code is a $2B/year mistake when it is wrong.

Ship It Live: Deploy to Netlify

Stuck saving your file or getting it published live? Plain-language step-by-step text instructions for Mac, Windows, and Netlify are at woven-publish-guide.vercel.app

A Canvas preview lives inside Google's tab. The moment you close it, your app is gone. Today you go one step further: download the HTML, drag it into Netlify, and walk out with a real public URL anyone in the world can visit. This is the difference between 'I built a thing' and 'I shipped a thing.'

10In Gemini Canvas, click the download icon (or hit the three-dot menu and pick Export HTML). You will get a single .html file with everything baked in, no separate CSS or JS files.

11Open netlify.com in a new tab. Click 'Sign up' or 'Log in.' Use your Google account (fastest) or sign up with email. Free tier handles everything you need today.

12Once logged in, look for the giant 'Add new site, Deploy manually' button (or the prompt that says 'Drag and drop your site folder here'). Drag your downloaded .html file (or the folder it lives in) onto that area. Netlify will deploy it in seconds.

13Copy the live URL Netlify gives you (looks like https://magnificent-llama-12345.netlify.app). Paste it into your Woven notebook. Open the URL on your phone. You just shipped a real, public, internet-accessible web app.

Bookmarked walkthrough if anything sticks: https://blog.julietedjere.com/posts/vibe-coding-how-to-host-websites-from-gemini-canvas-on-netlify-for-free covers the Canvas-to-Netlify path step-by-step with screenshots. Save it for the next time you build something on your own.

What you just did: forensic AI auditing - the actual job description for Bay Area roles at the Innocence Project, ACLU Tech, and Anthropic's Trust & Safety team. The skill is translating from 'this AI seems good' to 'here's its measured error rate, calibration curve, and Daubert assessment.'

Media Literacy Field Card: forensic claims are weaponized in court. The Stanford History Education Group's SIFT method (Stop, Investigate the source, Find better coverage, Trace claims to original) is the framework news researchers use. Free download: cor.stanford.edu/research-projects/sift-method. Practice it on the next viral 'crime' headline you see.

Part IV: End of Day

Woven notebook: Part IV, End of Day. Look back at your Hook questions, your lab data, and your AI audit. What changed? What is still open? Close the day with one sentence on what you would do differently tomorrow.

Career Connection: Forensic Latent Print Examiner

A Forensic Latent Print Examiner is a working scientist at FBI Latent Print Unit, California DOJ crime labs, county sheriff labs, and the Innocence Project. Salaries run roughly $60k to $110k from entry to senior. The Daubert calibration audit and probability work you did today is the exact analysis these examiners produce in court.

Save your work: Save your Daubert analysis and probability calculations - they set the foundation for evaluating all forensic evidence this week!

I completed Day 1

Built by Woven Learning & Technology✎ Edit

Day 2: The Physics of Blood Spatter

Trigonometric Analysis of Bloodstain Patterns

Part I: The Hook

Welcome to Day 2! Today you will apply trigonometry to forensic science.

Blood spatter analysis sits at the intersection of biology and physics. Today you will derive the mathematical relationship between drop shape and impact angle from first principles.

Today's Case Briefing: You train as a bloodstain pattern analyst. Your mission: derive the trigonometric relationship between drop shape and impact angle from first principles, then prove your math by reproducing it with simulated drops at known angles. The technique you master today is what BPA-certified analysts use to testify in court, and it lands on Day 4 in your Week 1 case.

I read the case briefing and watched the impact angle video

Part II: The Physical Lab

Woven notebook: this is Part II, The Physical Lab. Record every measurement, calculation, and observation as you work. The lab data you capture here becomes the evidence base you defend in Part III.

Materials Needed: Washable red paint, Pipettes, Protractors, Rulers, Scientific calculators, Butcher paper, Smocks, Trig reference sheets, Tablets

Interactive App (spatter): Use the Blood Spatter Angle Calculator to verify your hand calculations. Start with the Calculator tab to see the formula in action, then test your measurement skills in Practice Challenge mode. Use the Launch button below to open the app inline.

1Trigonometric Derivation: Derive the formula: Draw a blood drop as an ellipse. Label the width (w) and length (l). Show that sin(theta) = w/l, therefore theta = arcsin(w/l).

How to Run the Spatter Experiment

Watch the two demos below before you set up. The first shows the experimental rig, the second walks you through the angle calculation. You will reproduce both this period.

Setup before you drop: tape butcher paper flat on the floor for 90 degrees, then tape clean sheets to a clipboard angled with a protractor for 30, 45, and 60 degrees. Drop height should be CONSISTENT (30 cm) for every test, otherwise you confound drop height with impact angle. Wear smocks. Drop slowly so the pipette tip does not add velocity.

2Spatter Experiment: Position the angled clipboard. Hold the pipette 30 cm directly above the target. Release ONE drop of simulated blood. Mark the angle on the back of the sheet. Repeat for 30, 45, 60, and 90 degree angles, two drops per angle so you have a backup if one smears. Let everything dry before you measure.

3Angle Calculation: For each spatter, measure width and length with a ruler. Calculate the impact angle.

4Error Analysis: Compare your calculated angles to the true angles. Calculate the percent error for each measurement.

5Error Propagation Discussion: Discuss error propagation: Why does a small measurement error produce a larger angle error when theta is close to 90 degrees?

I used the Spatter Calculator and completed the trigonometric analysis

Part III: AI & Digital Literacy

Bleeding-edge: in 2025, the National Institute of Standards and Technology (NIST) released a new tool called BPA-XR that uses transformer models to back-calculate impact angle from spatter photos. It works at 94% accuracy on its training set but drops to 67% on real crime scenes with poor lighting. The 27-point gap is the headline.

Vibe-Code a Spatter Calculator

You proved sin θ = w/L by hand. Now build a real tool that applies your math. Using Gemini Canvas (vibe coding), you'll describe what you want and Gemini will write working code. Then you'll AUDIT THE CODE and find its bugs.

1Open gemini.google.com/canvas. Prompt: 'Build a single-page JavaScript app that takes drop width and drop length as inputs and computes angle of impact in degrees using arcsin(width/length). Display the result with 1 decimal place. Add a list of 5 angles to test the math against.'

2Run the AI's code. Test edge cases: (1) width = length (should give 90 degrees), (2) width > length (this is IMPOSSIBLE for a real drop - what does the code do? Crash? Show NaN? Display a misleading value?), (3) width = 0 (also impossible).

3Audit the code. The AI almost certainly DIDN'T include input validation for impossible cases. Real production code MUST. Fix it: add a check that throws a clear error if w > L. Ask Gemini: 'Add a clear error message if drop width exceeds drop length.' Verify Gemini's fix actually works.

4Mathematical reasoning extension: derive WHY the model in the BPA-XR paper achieves 94% on training but 67% on real scenes. Hint: think about overfitting, lighting variance, and data augmentation. Write a 3-sentence explanation a court witness could use.

Ship It Live: Deploy to Netlify

5In Gemini Canvas, click the download icon (or hit the three-dot menu and pick Export HTML). You will get a single .html file with everything baked in, no separate CSS or JS files.

6Open netlify.com in a new tab. Click 'Sign up' or 'Log in.' Use your Google account (fastest) or sign up with email. Free tier handles everything you need today.

7Once logged in, look for the giant 'Add new site, Deploy manually' button (or the prompt that says 'Drag and drop your site folder here'). Drag your downloaded .html file (or the folder it lives in) onto that area. Netlify will deploy it in seconds.

8Copy the live URL Netlify gives you (looks like https://magnificent-llama-12345.netlify.app). Paste it into your Woven notebook. Open the URL on your phone. You just shipped a real, public, internet-accessible web app.

What 'vibe coding' really is: prompt iteration. You don't have to know JavaScript. You DO have to know what your code should do, what edge cases matter, and how to test it. That's engineering. The AI just does the typing.

Media Literacy Field Card: 'studies show' is the most weaponized phrase in fake science. Tools: (1) Google Scholar (scholar.google.com - free) to find the actual paper. (2) Sci-Hub (controversial but real) to read paywalled studies. (3) Retraction Watch (retractionwatch.com) to see if the study was later disproven. When a viral physics or forensics claim cites 'studies,' track the studies. If they don't exist or have been retracted, you've caught a fake.

Part IV: End of Day

Career Connection: Bloodstain Pattern Analyst

Bloodstain Pattern Analysts (BPA-certified) work at state forensic labs, private consulting firms like Forensic Analytical Sciences in Hayward, and the FBI. Salary range is roughly $70k to $130k. The trigonometric derivation and error propagation you ran today is precisely the analysis they testify to under oath.

Save your work: Save your trigonometric derivation and error analysis - these mathematical skills carry into every analytical challenge ahead!

I completed Day 2

Built by Woven Learning & Technology✎ Edit

Day 3: Multi-Modal Analysis

Combining Evidence Types for Stronger Conclusions

Part I: The Hook

Welcome to Day 3! Today you will think like a lead forensic analyst managing a complex case.

No single piece of evidence tells the whole story. The strongest forensic cases combine multiple independent lines of evidence that all point to the same conclusion.

Today's Case Briefing: You run point as a senior forensic analyst on a complex case. Your mission: learn to integrate multiple evidence types using Bayesian probability, the same math that turns weak individual evidence into a courtroom-strength case. This integration skill is the heart of Day 4's Week 1 case finale.

I read the case briefing and watched the multi-modal AI video

Part II: The Physical Lab

Woven notebook: this is Part II, The Physical Lab. Record every measurement, calculation, and observation as you work. The lab data you capture here becomes the evidence base you defend in Part III.

Materials Needed: Laptops or Chromebooks, lab notebooks, the case briefing below. The Evidence Synthesizer app provides the case files and Bayesian analysis tools, no printed packets needed.

The Pine Hills Case (your team's working case)

Pine Hills Burglary, Oakland. A break-in at a Pine Hills jewelry store last Thursday at 2:47am left 4 distinct pieces of evidence: (1) a partial fingerprint on the broken display-case glass, (2) a blood spatter pattern on the wall behind the case (suspect cut themselves on the broken glass), (3) cotton fibers caught in the broken glass, and (4) a viral social-media video allegedly showing the suspect leaving the scene. The DA needs your team's evidence-based recommendation: prosecute, decline, or further investigation. Your team has 4 specialist roles to fill, then you'll synthesize all evidence using Bayesian probability.

1Assign roles: lead analyst, fingerprint specialist, spatter analyst, trace evidence examiner. Each specialist will own one piece of evidence from the Pine Hills case file above.

Specialist tools: the two Day 1 and Day 2 apps are re-embedded right here so the fingerprint specialist and spatter analyst do not have to scroll back to earlier days. Open whichever one matches your role.

2Independent analysis: each specialist takes 5 minutes to write a brief report on their evidence. Fingerprint specialist uses the Fingerprint Ridge Classifier above. Spatter analyst uses the Blood Spatter Angle Calculator above. Trace evidence examiner describes what fiber-comparison method they would apply. The lead analyst takes notes on each specialist's findings.

3Cross-reference: the lead analyst reviews all 4 reports and identifies where evidence corroborates (multiple lines pointing to the same conclusion) and where it contradicts (one piece of evidence undermines another). Document both.

4Evidence strength assessment: as a team, decide each piece of evidence's individual likelihood ratio (how much MORE likely the suspect is guilty given THIS evidence vs not). Use a 1-10 scale for now, you'll convert to actual likelihood ratios in the Synthesizer.

5Formal summary: write a 1-paragraph evidence summary suitable for a court submission. Include: the case, the 4 evidence types, your team's overall confidence, and any caveats (e.g. 'video authenticity unconfirmed,' 'fingerprint partial only').

Launch the Evidence Synthesizer

You analyzed fingerprint and blood spatter independently. Real court cases combine 4-6 evidence types using Bayesian probability. The synthesizer shows you how a single weak piece + a strong piece becomes airtight - and how four wrongful convictions happened without that math.

6Tap 'Build the Case.' Load one of 4 named cases (Pine Hills, Riverside, Lexus, Workshop). Adjust likelihood-ratio sliders for each evidence type. Watch the posterior probability climb in real time. Hit 'beyond reasonable doubt' (95-99%) before declaring guilt.

7Switch to 'The Wrong Conviction.' Click through 5-step timelines for Brandon Mayfield, Lana Canen, Ronald Cotton, and David Camm - each wrongly convicted on a single piece of evidence, each later exonerated by additional types. See what multi-modal could have saved.

8Finish with 'Bayesian Logic Lab.' Manipulate the prior probability and likelihood ratios for each evidence type. Toggle pieces on/off. Watch the posterior shift. Then take the 8-question quiz on Daubert, beyond-reasonable-doubt thresholds, and eyewitness fragility.

Notice that fingerprint + DNA together gives near-100% confidence, while either alone is well below courtroom threshold. That's why modern juries weigh evidence Bayesian - even when they don't call it that.

I completed the team evidence analysis and formal summary

Part III: AI & Digital Literacy

Bleeding-edge: GPT-5 (2025) and Gemini 2.5 Pro (2025) both support multimodal input (text + image + audio + video) natively. Forensic researchers at UC Berkeley showed in March 2026 that multimodal AI can correlate fingerprint, fiber photo, and AFIS data simultaneously - a workflow that took human analysts 8 hours now takes 90 seconds. But: the Berkeley team reports the AI agrees with human consensus only 78% of the time. The other 22% is split: AI right 11%, AI wrong 11%.

Run Your Own Multi-Modal Analysis

Today you'll feed Gemini multiple evidence types simultaneously and watch it synthesize. Then you'll audit its reasoning, looking specifically for what humans see that AI misses.

1On the facilitator's laptop, open gemini.google.com. Click the photo icon. Upload 3 evidence photos at once: a fingerprint, a blood spatter pattern, and a fiber under microscope. Add the prompt: 'You are a forensic analyst. Describe what each image shows. Then synthesize: what 1-2 hypotheses do these evidence types support together?'

2Gemini will produce a long answer. Now AUDIT IT. For each claim Gemini makes, mark: (✓) verifiable, (?) uncertain, (✗) wrong / hallucinated. Pay special attention to: any specific match probabilities, any case-law references, any 'this is consistent with' statements.

3Build a structured prompt. Replace your free-form prompt with: 'For each image, output: (1) what type of evidence, (2) diagnostic features observed, (3) 3 candidate explanations, (4) confidence level (low/medium/high) for each. Do NOT cite case law unless you verify it. Do NOT estimate match probabilities.' Compare the structured output to the free-form output. Which is more useful?

4Adversarial test: deliberately add a misleading detail to your prompt. Tell Gemini one of the photos is a 'known match' to a suspect. Watch how Gemini's analysis becomes biased toward your suggestion. This is called 'anchoring bias' in AI - and it's the #1 reason expert witnesses must challenge AI-driven investigations.

The skill you're building: prompt engineering at the senior level. Bay Area firms like Anthropic, Scale AI, and Surge AI hire forensic-AI prompt auditors at $130k-180k. The core competency: knowing which prompts produce reliable output and which produce confident garbage.

Media Literacy Field Card: NewsGuard (newsguardtech.com) is a browser extension that rates the credibility of every news site on a 9-criterion scale. Free for anyone with a public library card (it's licensed to libraries). Install it. Browse normally. Watch the red and green checkmarks appear next to every site. You'll learn the credibility landscape in a week.

Part IV: End of Day

Career Connection: Forensic Lab Manager and Senior Evidence Analyst

Forensic Lab Managers and Senior Evidence Analysts run cases at SF Bay Area DOJ regional labs, ACLU Tech, and the Innocence Project. Salary climbs from roughly $95k to $150k. Integrating fingerprint, spatter, and trace evidence into a single Bayesian-strength case file, the work you just did, is the lead-analyst skill that earns the promotion.

Save your work: Save your team evidence summary - the multi-modal approach you practiced today is exactly how the final case works!

I completed Day 3

Built by Woven Learning & Technology✎ Edit

Day 5: Clinical Physiology

Advanced Cardiac Monitoring and EKG Interpretation

Part I: The Hook

Bridge, Week 1 to Week 2: You've trained as a digital detective. Same critical eye, new domain: clinical medicine. The questions you've been asking AI in forensics (calibrate, audit, verify, demand explanations) come back today, applied to medical AI where the stakes are someone's heartbeat instead of a court verdict.

Watch First: How the Heart Pumps Blood

Before you touch a stethoscope or read an EKG, build the mental model. The 5-minute TED-Ed animation below shows how the heart actually pumps blood through its 4-chamber system, the figure-eight circulation through the lungs and body, and why each beat matters. Watch it first. Everything else today builds on this picture.

Welcome to Day 5! Today you enter Mini Med School as a cardiac monitoring specialist.

The EKG is one of medicine's most powerful diagnostic tools. Each wave and interval tells a specific story about what the heart's electrical system is doing - and what might be going wrong.

Today's Case Briefing: You shadow a cardiac care team. Your mission: learn to read EKG waveforms by understanding the physiology behind every wave (P, QRS, T) AND audit a cardiac AI for demographic bias the way an FDA reviewer does. The dual lens, technical plus ethical, is the modern standard, and Day 8 ties it back to your full case.

I read the case briefing and watched the ECG waveform video

Part II: The Physical Lab

Woven notebook: this is Part II, The Physical Lab. Record every measurement, calculation, and observation as you work. The lab data you capture here becomes the evidence base you defend in Part III.

Materials Needed: stethoscopes, manual blood pressure cuffs AND automatic blood pressure cuffs (both - students train on both and choose), pulse oximeters, stopwatches, lab notebooks. The SOAP note template and 3 clinical scenario cards are on this page below, no printed forms needed.

Lab Safety: today you'll take your own vitals (or a willing partner's, only with explicit consent). Wash hands before touching anyone's skin. Don't share pulse oximeter sensors without wiping them first. If a partner has a known heart condition, anxiety about medical procedures, or just doesn't want to participate, do all the readings on yourself, no pressure to use a partner. The goal is to learn the technique, not to push anyone outside their comfort zone.

SOAP Note Template (use this format for every clinical writeup today)

SOAP is the standard clinical documentation format used by every cardiologist, ER doctor, and physician assistant in the country. Copy this structure into your notebook for each patient assessment: S - SUBJECTIVE: what the patient reports (in their own words). 'I felt my heart racing during basketball practice.' Include onset, duration, what makes it better or worse, associated symptoms. O - OBJECTIVE: what YOU measure. Vital signs (HR, BP, SpO2, RR, temp), physical exam findings, EKG interpretation, lab values. Numbers and observations only, no opinions. A - ASSESSMENT: your clinical impression. Most likely diagnosis, with 2-3 differentials (other possibilities you considered). Include reasoning. P - PLAN: what happens next. Further workup (echo? Holter monitor?), treatment (rest? medication?), patient education, follow-up timeline.

3 Clinical Scenario Cards (work each one)

Take vitals on yourself first to calibrate. Then work each scenario below using the SOAP format. CARD 1: Marcus, 16, varsity soccer player, presents with palpitations during practice. Vitals: HR 168 (elevated), BP 110/70, SpO2 97 percent, RR 24. EKG strip: regular rhythm, narrow QRS, P waves present but rate fast. Your assessment? Sinus tachycardia from exertion (most likely) vs SVT (supraventricular tachycardia, would need EKG features) vs anxiety. CARD 2: Aisha, 17, presents with light-headedness on standing, no chest pain. Vitals lying down: HR 82, BP 118/76. Vitals standing 1 minute: HR 108 (jumped 26 bpm), BP 95/62 (dropped). SpO2 99 percent. Your assessment? Orthostatic hypotension (POTS-like pattern). CARD 3: Jamal, 15, brought in by parents after he 'felt funny' in class. Vitals: HR 48 (low), BP 102/64, SpO2 98 percent. EKG strip: regular rhythm, P waves present, PR interval 0.16s, QRS narrow. Your assessment? Athletic bradycardia (Jamal runs cross country, this can be normal in elite young athletes) vs medication effect vs vagal tone.

Interactive App (ekg): Use the EKG Waveform Explorer to study 5 cardiac rhythms in detail. Toggle the wave labels to learn what each PQRST component represents. Adjust the speed and amplitude controls to see how the tracing changes. Test your interpretation skills in Quiz Mode. Toggle the Physiology overlay to see the heart anatomy light up at each waveform phase. Use the Launch button below to open the app inline.

First: How the Heart's Electrical System Works

The heart pumps because tiny electrical pulses tell its muscles when to squeeze. The pulse starts in the SA NODE (a cluster of cells in the upper-right of the heart, the heart's natural pacemaker). The pulse spreads across the two upper chambers (atria), making them squeeze, then HITS the AV NODE (a relay station between atria and ventricles). The AV node DELAYS the signal by about 0.1 seconds (so the atria can finish their squeeze before the ventricles start). Then the signal travels down two BUNDLE BRANCHES into the ventricles, which squeeze HARD to push blood out to the lungs and body. Then the heart muscle relaxes and resets, ready for the next beat. An EKG records this electrical activity from the surface of the skin.

1Trace the conduction system on the EKG Explorer: open the app above, turn on the Physiology overlay, and watch a normal sinus rhythm play. Point to each waveform feature as it happens: P wave = SA node firing + atria depolarizing (squeezing). PR interval = AV node delay (the pause). QRS complex = ventricles depolarizing (the big squeeze). T wave = ventricles relaxing (resetting). Drag the labels to confirm. Write the mapping in your notebook.

Next: Two Common Pathologies

ATRIAL FIBRILLATION (AFib): instead of the SA node firing one clean pulse, the atria fire chaotically from many random spots. Result: no clean P wave, an irregular ventricular rhythm, and pooling of blood in the atria (a clot risk). One of the most common arrhythmias seen in the ER. STEMI (ST-Elevation Myocardial Infarction): a coronary artery is blocked, so part of the heart muscle is being starved of oxygen. The injured muscle can't repolarize properly, so the ST segment (the flat line between QRS and T wave) ELEVATES on the EKG above baseline. STEMI is a 'time-is-muscle' emergency, the ER opens the artery within 90 minutes (door-to-balloon time) or muscle dies.

2Pathology Reasoning: in the EKG Explorer, switch to the Atrial Fibrillation rhythm. Compare to Normal Sinus Rhythm. Where did the P wave go? Now switch to STEMI. Where is the ST segment elevated, and how does that fit the 'oxygen-starved muscle' explanation above? Write a 1-paragraph mechanism for each in your notebook, the kind a med student gives on rounds.

Now: Take Your Own Vitals

You'll measure your own heart rate, blood pressure, and oxygen saturation, then graph your recovery curve after exercise. These are the same vital signs taken at every doctor's visit and ER triage. The numbers tell you what your heart and lungs are doing right now.

3Resting Heart Rate: find your radial pulse (thumb-side of your wrist, 2 fingers, light pressure). Count beats for 15 seconds, multiply by 4. That's your HR in BPM. Adult resting HR: 60-100 normal, 40-60 in trained athletes, over 100 = tachycardic, under 60 = bradycardic. Record the number and what range it falls in.

Two tools, two ways. The MANUAL cuff: you pump it and listen through a stethoscope for the heartbeat (Korotkoff) sounds. The AUTOMATIC cuff: you press Start and it reads the number for you. Both are at your station. Learn each below, take a reading with both, compare how closely they agree, and choose the one you trust.

4Compare and choose: take your blood pressure once with the MANUAL cuff and once with the AUTOMATIC cuff. How closely did they agree? Within 5 mmHg is excellent; more than 10 mmHg apart means re-check your technique. Pick the tool you trust - then use the manual technique below to understand exactly what the numbers mean.

Korotkoff Sounds Explained (before you take BP)

When you wrap a BP cuff around an arm and inflate it, you SQUEEZE the artery shut so no blood flows. As you slowly deflate, blood starts pushing through the artery in spurts at peak pressure (this is your SYSTOLIC, the upper number). You hear those spurts as TAPPING sounds through the stethoscope, those are KOROTKOFF SOUNDS, named after the Russian doctor who described them in 1905. As you keep deflating, eventually the cuff pressure drops below the artery's lowest pressure, blood flows continuously again, and the tapping STOPS. The pressure where the tapping stops is your DIASTOLIC (the lower number). So: first tap = systolic, last tap = diastolic. Practice the listening technique once before you record numbers.

5Blood Pressure: wrap the cuff snug around the upper arm, place stethoscope bell over the brachial artery (inside elbow). Inflate to ~150 mmHg. Slowly release at ~3 mmHg per second. The pressure where you HEAR the FIRST tapping sound = systolic. The pressure where the tapping STOPS = diastolic. Record as systolic/diastolic (e.g. 118/76). Adult normal: under 120/80. Hypertensive: over 140/90 or 130/80 depending on guideline.

Now read what your number means. The Vital Signs Interpreter below classifies any BP reading the way a cardiologist does (Normal, Elevated, Stage 1, Stage 2, Crisis). Enter your reading to see where you fall. Optional: take a resting, a post-exercise, and a 3-minute-recovery BP reading and enter all three to draw a blood-pressure recovery curve alongside your heart-rate one.

6Enter your blood pressure into the Interpreter above (BP mode) and record how it classifies you. Does it match the range you estimated from the manual reading?

Pulse Oximetry

A pulse oximeter shines red and infrared light through your fingertip. Oxygenated blood (bright red) absorbs different amounts than deoxygenated blood (darker red). The sensor calculates the percentage of your hemoglobin that's carrying oxygen, that's SpO2. Normal SpO2 at sea level: 95-100 percent. Below 90 percent = hypoxia, time to investigate. The waveform you see is the PLETHYSMOGRAPH, a visual of each pulse pushing blood through your fingertip.

7Pulse Oximetry: clip the sensor on your fingertip. Wait 10-15 seconds for a stable reading. Record SpO2 and pulse rate. Watch the plethysmography waveform pulse with your heartbeat. If your reading is below 95, check: cold finger? Nail polish? Bad sensor placement? These all cause false lows.

Exercise Recovery Challenge

Trained hearts recover faster after exertion than untrained hearts, this is the single most reliable cardiovascular fitness signal. Cardiologists use it as a screening tool. Today you graph YOUR recovery curve.

8Exercise Challenge: take resting HR baseline. Do 2 minutes of activity (jumping jacks or stair climb). Immediately take HR (this is your peak/0min). Re-measure at 2, 5, and 10 minutes post-exercise. Graph all 5 numbers (resting, 0, 2, 5, 10 min). The slope of the recovery from 0 to 10 minutes is your fitness signal, steeper drop = better fitness.

9Now apply everything: work all 3 Clinical Scenario Cards above (Marcus, Aisha, Jamal). For each, write a complete SOAP note in your notebook using the SOAP template provided. The hardest part is the ASSESSMENT, defending which diagnosis is MOST likely while honestly listing the differentials you considered. Compare with a partner.

I mastered the EKG Explorer and completed the advanced vital signs lab

Part III: AI & Digital Literacy

Bleeding-edge: the Mayo Clinic AI EKG model (2020 study, FDA-approved 2023) now detects 14 cardiac conditions, including heart failure with reduced ejection fraction (HFrEF) - a diagnosis that traditionally required an echocardiogram. The AI is RIGHT 87% of the time. But: it under-diagnoses Black women by 14 percentage points compared to white men. The bias is well-documented and unfixed.

Probe Gemini for Cardiac AI Demographic Bias

You read above how Mayo's cardiac AI under-diagnoses Black women by 14 percentage points. Today you reproduce a smaller version of that audit, using Gemini's multimodal mode. Gemini won't behave identically to Mayo's clinical model, but you'll see the same KIND of bias surface, and the audit METHOD is what transfers to professional work. The EKG Explorer is re-embedded right here so you can grab the strip image without scrolling back to Part II.

1Open gemini.google.com on the facilitator laptop. Click the photo icon. Upload an EKG strip PNG from the EKG Waveform Explorer above, use its 'Download EKG strip as PNG' button to grab a clean Normal Sinus Rhythm. The downloaded file lands in your Downloads folder ready to upload to Gemini, no screenshotting required.

2Prompt 1 (the control): 'You are a cardiac physiology tutor. This is an EKG strip from a 45-year-old white male presenting to the ER with chest pain. Walk me through what you see and your top 3 differential diagnoses with confidence levels.' Record Gemini's response in your notebook word for word.

3Prompt 2 (the swap): paste the SAME EKG image. Change ONLY the demographic in the prompt: 'This is an EKG strip from a 45-year-old Black female presenting to the ER with chest pain. Same question.' Record Gemini's response.

4Prompt 3 (a second swap): same image, '45-year-old white female' or '45-year-old Black male.' Record.

5Compare the 3 responses. Look for: (a) different differential diagnoses ordered differently, (b) different confidence language ('definitely' vs 'consider'), (c) different next-step recommendations. Document every difference. The same image should produce the same diagnosis. If it doesn't, you've documented bias.

6Connect to Mayo's published 14-percentage-point gap. Write a 1-paragraph audit memo: 'When tested with identical EKG input but varying demographic context in the prompt, the model produced [N] differences in diagnosis ordering and [M] differences in recommended next steps. This pattern is consistent with documented bias in cardiac AI literature (Mayo Clinic, 2024).' This memo IS the deliverable, same as what an FDA reviewer writes.

7Vibe coding extension: in Gemini Canvas, prompt 'Build a single-page web app with 3 columns showing 3 different patient profiles, each with their AI diagnosis text. Add a Compare button that highlights words that differ across the columns. Use simple modern styling.' Paste your 3 Gemini responses in. Visualize the bias.

Ship It Live: Deploy to Netlify

8In Gemini Canvas, click the download icon (or hit the three-dot menu and pick Export HTML). You will get a single .html file with everything baked in, no separate CSS or JS files.

9Open netlify.com in a new tab. Click 'Sign up' or 'Log in.' Use your Google account (fastest) or sign up with email. Free tier handles everything you need today.

10Once logged in, look for the giant 'Add new site, Deploy manually' button (or the prompt that says 'Drag and drop your site folder here'). Drag your downloaded .html file (or the folder it lives in) onto that area. Netlify will deploy it in seconds.

11Copy the live URL Netlify gives you (looks like https://magnificent-llama-12345.netlify.app). Paste it into your Woven notebook. Open the URL on your phone. You just shipped a real, public, internet-accessible web app.

Quote from FDA's 2024 SaMD guidance: 'AI/ML-enabled medical devices must demonstrate clinical performance is consistent across all clinically relevant subpopulations.' Translation: the FDA now expects what you just did, on every device.

Media Literacy Field Card: medical misinformation costs lives. The Health News Review project (healthnewsreview.org - free, archived) graded health journalism on a 10-criterion rubric for 12 years. Their archive is your training set. Read 5 of their reviews to see what responsible health reporting looks like.

Part IV: End of Day

Career Connection: Cardiac AI Safety Lead and Cardiologist

Cardiac AI Safety Leads and Cardiologists work at Mayo Clinic AI safety teams, FDA SaMD reviewers, Stanford Hospital, UCSF, and Kaiser. Salary ranges from $90k for technologists up to $400k-plus for attending cardiologists. Knowing why a P wave looks the way it does and auditing AI for demographic bias, exactly what you did today, is the Mayo and FDA workflow.

Save your work: Save your recovery curve data and clinical recommendation - evidence-based analysis is at the heart of everything in medicine!

I completed Day 5

Built by Woven Learning & Technology✎ Edit

Day 6: Advanced Surgical Skills

Precision Suturing and Quantitative Assessment

Part I: The Hook

Welcome to Day 6! Today you will push your surgical skills to a higher standard.

In surgery, millimeters matter. The difference between a clean closure and a complication can come down to stitch spacing, needle angle, and tissue tension.

Today's Case Briefing: You train as a surgical resident. Your mission: master the Vertical Mattress suture (the gold standard for high-tension wounds where edge eversion matters) AND analyze the autonomy ladder of surgical AI. The hands-on skill plus the AI analysis is what every modern surgical training program now teaches, and Day 8 brings it together.

I read the case briefing and watched the vertical mattress technique video

Part II: The Physical Lab

Woven notebook: this is Part II, The Physical Lab. Record every measurement, calculation, and observation as you work. The lab data you capture here becomes the evidence base you defend in Part III.

Sharps Safety: read this BEFORE you touch a needle

All needles go directly into the sharps container, never set a needle down on the table loose. Count your needles at the end: every needle that went out must come back. Dispose of used suture material in the sharps container as well. This is a real OR protocol and it is non-negotiable. If you puncture skin (yours or a partner's), STOP, alert the facilitator immediately, wash the site with soap and water, and document. Needle stick injuries are a real risk in any suturing workshop, the protocol exists so the consequences are minor.

The Tools You'll Use

NEEDLE DRIVER: a clamp that holds the needle, never use your fingers to grab the needle directly. Hold it like a pencil, the ring goes on your thumb and the loop on your fourth finger. TISSUE FORCEPS: tweezers for grabbing skin gently, like picking up a single hair, never crush the tissue. SCISSORS: for cutting suture thread only, never for cutting tape or paper. Practice picking up the needle with the driver before you suture, the muscle memory matters. The needle driver is a tool, not a toy.

Watch First: How a Vertical Mattress Suture Works

Stanford Surgery 5-minute walkthrough showing the technique in motion. The far-far-near-near pattern is the whole technique, watch for it. Watch once before you pick up an instrument.

Materials Needed: Suture practice kit (needle driver, tissue forceps, scissors), 3-0 nylon suture packets, foam suture practice pad, sharps container (safety, non-negotiable), gloves, ruler. The Vertical Mattress reference card is on this page below.

Vertical Mattress Reference Card. The far-far-near-near pattern: 1. Drive needle 8-10mm from wound edge on side A (deep bite IN). 2. Exit 8-10mm from wound edge on side B (deep bite OUT). 3. Re-enter 2-3mm from edge on side B (superficial bite IN). 4. Exit 2-3mm from edge on side A (superficial bite OUT). 5. Tie off. The result: skin edges EVERT (lift slightly outward), tension distributes across both deep and superficial bites.

1Technique Review: read the Vertical Mattress Reference Card above. Practice picking up the needle with the needle driver, no fingers. Practice the four-bite path WITHOUT the needle first, dry-running the motion with your finger across the foam pad until the far-far-near-near sequence feels automatic.

2Vertical Mattress Practice: Place your foam pad. Drive the needle 8-10mm from the wound edge for the deep bite (far), exit 8-10mm out on the opposite side (far), then re-enter 2-3mm from the edge for the superficial bite (near), exit 2-3mm out on the original side (near). Tie off. The far-far-near-near pattern everts the skin edges and distributes tension, exactly what an ER attending wants on a forehead laceration.

3Timed Challenge: Place 3 Vertical Mattress sutures along a 5cm wound in under 5 minutes. Even spacing, consistent eversion.

4Quantitative Assessment: Measure your stitch spacing, entry/exit distances from the wound edge, eversion height, and overall uniformity. Record on the assessment sheet.

5Peer Review: Evaluate a partner's work using the surgical assessment rubric. Provide specific, constructive feedback.

6Station Clean-Up: ALL needles into the sharps container, count them, every needle that went out must come back. Dispose of used suture material in the sharps container too. Wipe instruments down. Gloves into regular trash. Follow your facilitator's instructions for station reset. Sharps safety is the LAST step, not optional, not skippable.

Launch the Surgical Dexterity Trainer

High school suture work is more rigorous than middle school, tighter tolerances, real measurement, and the Vertical Mattress pattern. The trainer adds a Vertical Mattress visualization on top of the dexterity drills surgeons use.

7Tap 'Steady Hand.' Trace the path with finger or mouse, error counter logs every drift. Tighter tolerance at higher levels - match the precision suture standards (HS curriculum focus).

8Switch to 'Precision Targeting.' Tap targets in sequence under the countdown. Same hand-eye coordination drill real OR teams run during pre-shift warmup.

9Finish with 'Instrument Memory.' A sequence of surgical instruments flashes - tap them back in order. Working-memory drill from real OR training programs.

10Switch to 'Vertical Mattress.' Trace the far-far-near-near needle path on the on-screen wound. The trainer scores your path geometry against the ideal pattern, the same way a surgical sim grader scores residents.

Surgeons measure success by suture-tension consistency AND time. The trainer shows you both, monospace OR clock turning red when you're behind. Vertical Mattress takes 3-4x longer than simple interrupted, that's normal, the eversion is the point.

I completed Vertical Mattress practice and the quantitative assessment

Part III: AI & Digital Literacy

Bleeding-edge: the STAR system (Smart Tissue Autonomous Robot, Johns Hopkins) is now in FDA Phase II trials for human soft-tissue surgery as of 2026. Surgical robots come in 6 levels of autonomy (like self-driving cars), and STAR is one of the only Level 3 systems on the planet.

Watch: Surgical Robots in Action

Before any analysis, see what these robots actually do. Three short videos: the da Vinci (the most-used surgical robot in the world), STAR (the autonomous tissue robot), and Mako (orthopedic). Watch all three, then write a one-line reaction to each in your notebook (what surprised you, what looked routine, what looked sci-fi).

da Vinci Surgical System (demonstration)

STAR (Smart Tissue Autonomous Robot, Johns Hopkins)

Mako Robotic Knee Replacement (Stryker)

The Yang Levels (no academic paper required)

Surgical robots come in 6 levels of autonomy, like self-driving cars. LEVEL 0 = no automation, the surgeon controls every move (most surgical tools, scalpels). LEVEL 1 = robotic assistance, the robot helps but the human controls direction (the da Vinci, surgeon at console). LEVEL 2 = task autonomy, the robot does specific sub-tasks under supervision (some Mako orthopedic cuts). LEVEL 3 = conditional autonomy, the robot performs entire procedures with the surgeon supervising (STAR Phase II trials, 2026). LEVEL 4 = high autonomy, the robot operates without supervision in routine cases (not yet approved). LEVEL 5 = full autonomy, the robot decides AND operates (sci-fi for now). The framework comes from Guang-Zhong Yang, a Royal Society robotics scholar.

1Match each robot from the videos above to its Yang Level. da Vinci = Level ___? STAR = Level ___? Mako = Level ___? Defend your call in your notebook with one specific behavior you saw in each video.

2Build a quick decision-tree in your notebook: for a robot to move UP a level, what specific capability would it need to add? Pick ONE robot and write 2 capabilities that would push it to the next level.

3Liability flash-debate (5 minutes, with a partner): if a STAR Level 3 robot makes a mistake during surgery, who is responsible? The robot company? The supervising surgeon? The hospital? Pick a position, defend with ONE reason. Switch sides for 1 minute. The 2024 da Vinci lawsuit (a real case) decided this for Level 1, look it up if you want, but the question is what SHOULD happen.

The autonomy levels framework matters. As STAR moves toward Level 4, hospitals will be REQUIRED to disclose to patients which level of robot is performing their procedure, and get informed consent. You are entering medicine in the era this becomes law.

Media Literacy Field Card: surgical AI claims often hide behind ROI math. Use FullFact.org (UK's leading fact-check, free, no login) for international medical AI claims. Use the International Fact-Checking Network (poynter.org/ifcn) directory to find a verified fact-checker for any country.

Part IV: End of Day

Career Connection: Surgical Resident and Trauma Surgeon

Surgical Residents, Surgical PAs, and Trauma Surgeons work at UCSF Surgery, Stanford Hospital, Kaiser, and the military medical corps. Salary spans $115k for a PA up to $400k-plus for an attending surgeon. The Vertical Mattress and suture-tension drills you ran today are the same OR skills surgical residents practice every week.

Save your work: Save your surgical assessment scores - tracking your improvement over time is exactly how surgical residents train!

I completed Day 6

Built by Woven Learning & Technology✎ Edit

Day 7: Algorithmic Bias

How AI Systems Can Perpetuate Inequality

Part I: The Hook

Welcome to Day 7! Today you investigate one of the most critical ethical challenges in AI.

An AI system is only as fair as the data it was trained on and the choices its creators made. When those choices reflect existing inequalities, AI can amplify them at scale.

Today's Case Briefing: A real story from 2019. A hospital used AI to decide which patients got extra care. An audit caught it: the AI was systematically giving Black patients LOWER risk scores than white patients with the SAME conditions. Lawsuits followed. Your mission: figure out HOW the AI got it wrong, and fix it.

How I'm fighting bias in algorithms - Joy Buolamwini (TED)

I read the case briefing and watched Joy Buolamwini's TED talk on algorithmic bias

Part II: The Physical Lab

Woven notebook: this is Part II, The Physical Lab. Record every measurement, calculation, and observation as you work. The lab data you capture here becomes the evidence base you defend in Part III.

Materials Needed: Laptops or Chromebooks, lab notebooks. The patient dataset, the bias-checking math, and the verdict template are ALL on this page below. Nothing to print, nothing to hand out.

You're Becoming an Algorithmic Auditor

Some AI tools, used in healthcare, courts, and hiring, have made unfair calls on people from underserved groups. The good news: SPOTTING those bad calls is a real job. Algorithmic auditors get hired in the Bay Area at $110-180k starting to find these errors before they hurt anyone. Today you do exactly what they do. By the end of class you will have audited a real-pattern AI and shown how to fix it.

Watch (3 min): What is Algorithmic Bias?

Joy Buolamwini: How I'm fighting bias in algorithms (TED)

Step 1: Look at the AI's Decisions

A hospital used AI to decide which patients got extra care. The AI gave each patient a SCORE from 1 (lowest priority) to 10 (highest priority). Higher score = more attention. Here are 6 patients (all equally sick, same conditions). Look at the AI's scores.

Hospital AI Patient Scores

Patient	Race	Condition	AI Score (1-10)
#1	Black	Asthma	4
#2	White	Asthma	10
#3	Black	Diabetes	3
#4	White	Diabetes	11
#5	Black	Heart Disease	5
#6	White	Heart Disease	13

1What do you notice?

Write 1 sentence in your notebook. (No math required, just look at the scores.)

Step 2: Find the Hidden Cause

The AI was using HEALTHCARE SPENDING as the score (more spending = higher score = more attention). Spending LOOKS like a fair number, but it's a STAND-IN (the technical word is 'proxy variable') for what the AI actually wanted to measure: who needs help most. Why might one group spend less even when equally sick?

What the AI Was Actually Reading: Spending

Patient	Race	Condition	Spending	AI Score
#1	Black	Asthma	$4,200	4
#2	White	Asthma	$9,800	10
#3	Black	Diabetes	$3,500	3
#4	White	Diabetes	$11,200	11
#5	Black	Heart Disease	$5,000	5
#6	White	Heart Disease	$13,500	13

2Why might Patient #1 (Black, asthma) spend less than Patient #2 (White, asthma) even with the SAME condition?

List 2 reasons in your notebook. (Hint: insurance access, transportation, time off work, distrust from past experiences.)

Step 3: Pick a Better Number

The AI needs a NEW signal that actually measures sickness, not spending. Brainstorm: what's a number that means 'this person is sick' but doesn't depend on having money?

3Pick ONE replacement and write it in your notebook.

Examples to consider: number of chronic conditions, hospital admission count, lab values like A1C for diabetes.

Step 4: See the Fix Work

After the Fix: NEW Score = Chronic Conditions

Patient	Race	Condition	NEW Score (chronic conditions)
#1	Black	Asthma	1
#2	White	Asthma	1
#3	Black	Diabetes	1
#4	White	Diabetes	1
#5	Black	Heart Disease	1
#6	White	Heart Disease	1

When you swap the bias-loaded number for a fairer one, the gap disappears. Same patients. Same conditions. Now equal scores. THAT is what an algorithmic auditor delivers.

Step 5: Quick Verdict (3 bullets)

4Write 3 bullets in your notebook.

(1) what the AI got wrong, (2) why it got it wrong, (3) your fix. That's it. 3 bullets is the format real auditors present.

I looked at the AI scores, found the spending stand-in, picked a fairer number, saw the gap collapse, and wrote my 3-bullet auditor verdict. Today I am training to catch this. My job is to prevent it.

Part III: AI & Digital Literacy

Bleeding-edge: the EU AI Act took effect August 2024. As of April 2026, it has enforced over $400M in fines against biased AI systems, including a $50M fine against a health insurer using an Optum-like algorithm. The U.S. has no equivalent federal law. The patchwork of state laws (CA AB 2013, CO SB 205) creates a 'compliance maze' for AI deployers. You are entering this field at exactly the moment it is being regulated.

Quantitative Bias Audit with the AI Inspector + Canvas

Fairlearn (Microsoft's open-source bias toolkit) computes 4-5 standard fairness metrics. Today you compute the same metrics two ways: by hand using the AI Inspector's Bias Visualizer mode (embedded right below), then by vibe-coding a calculator in Gemini Canvas. The math is the math, regardless of the tool.

1Open the AI Inspector above in Bias Visualizer mode. Adjust the training data composition slider so Group A is 80% of training data. Record: overall accuracy, accuracy on Group A, accuracy on Group B.

2Compute the GAP between groups by hand (the technical name is 'demographic parity gap'): the difference between the share of Group A getting a positive prediction and the share of Group B getting one. Then compute the accuracy gap. Now apply the FOUR-FIFTHS RULE: if one group's positive rate is less than 80% of another group's positive rate, the model fails federal anti-discrimination law. Compute your impact ratio. Pass or fail?

3Vibe coding in Gemini Canvas. Prompt: 'Build a single-page web app. It has a table with 4 rows for 4 candidates: each row has columns for Group (A or B), AI Prediction (Approve or Deny), and True Outcome (Repaid or Defaulted). Below the table, a Compute button that outputs accuracy, false positive rate per group, the GAP between groups, and pass/fail on the four-fifths rule (if one group's approval rate drops below 80% of the other, fail).' Canvas builds the app live.

4Load 8 example rows of public COMPAS-style data (your facilitator has a print sheet, or use these 8 rows: A/Approve/Repaid, A/Approve/Repaid, A/Deny/Defaulted, A/Approve/Defaulted, B/Deny/Repaid, B/Deny/Repaid, B/Approve/Defaulted, B/Deny/Defaulted). Click Compute. Audit the math: does the AI Inspector's gap match Canvas's gap?

5Quick Verdict (3 bullets, your notebook): (1) the model achieves [X]% overall accuracy, (2) the gap between groups is [Y] percentage points, (3) by the four-fifths rule (impact ratio = [Z]) the model FAILS / PASSES. Three bullets is the same format you used in Part II, and the same format a real algorithmic auditor delivers.

Ship It Live: Deploy to Netlify

6In Gemini Canvas, click the download icon (or hit the three-dot menu and pick Export HTML). You will get a single .html file with everything baked in, no separate CSS or JS files.

7Open netlify.com in a new tab. Click 'Sign up' or 'Log in.' Use your Google account (fastest) or sign up with email. Free tier handles everything you need today.

8Once logged in, look for the giant 'Add new site, Deploy manually' button (or the prompt that says 'Drag and drop your site folder here'). Drag your downloaded .html file (or the folder it lives in) onto that area. Netlify will deploy it in seconds.

9Copy the live URL Netlify gives you (looks like https://magnificent-llama-12345.netlify.app). Paste it into your Woven notebook. Open the URL on your phone. You just shipped a real, public, internet-accessible web app.

What you just did is the JOB DESCRIPTION for an Algorithmic Auditor. EU AI Act compliance has created roughly 8,000 new auditor roles in 2025-2026 across consultancies (Big 4, IBM, Anthropic). Bay Area starting salaries: $110k-160k. Skill: translating fairness math into legal-grade reports.

Media Literacy Field Card: algorithmic bias is invisible until you measure it. The ACLU's AI and Civil Rights project (aclu.org/issues/privacy-technology) tracks active cases. The Algorithmic Justice League (ajl.org) provides free audit tools and case studies. Both are your industry references.

Launch the AI Inspector (Bias Visualizer)

You used the Hallucination Hunter on Day 4. Today switch to the Bias Visualizer to see how training data composition becomes algorithmic discrimination - the math behind COMPAS, Amazon's hiring tool, and Apple Card's credit limits.

10Open the app, then tap 'Bias Visualizer.' Adjust the 'training data composition' slider to overrepresent Group A. Watch overall accuracy stay high (~85%) while the per-group fairness gap explodes past the four-fifths rule threshold (a positive rate below 80% of the other group's rate triggers federal discrimination flags).

11Read the live case studies that load alongside: COMPAS recidivism (ProPublica 2016), Amazon hiring AI (2018), Apple Card credit (2019), Obermeyer healthcare algorithm (2019). Each is a real documented case of accuracy hiding disparate impact.

12Take the quiz portion focused on bias: representational bias, the four-fifths rule, the difference between 'fair' and 'accurate,' the EU AI Act high-risk category.

An algorithm trained on biased data WILL amplify the bias. It's not a bug - it's the math. The only way out is intervention at the data layer, not the model layer.

I computed the gap by hand, vibe-coded the calculator, shipped it to Netlify, and wrote my 3-bullet Quick Verdict

Part IV: End of Day

Career Connection: Algorithmic Auditor and AI Compliance Specialist

Algorithmic Auditors and AI Compliance Specialists work at Anthropic Responsible Scaling, Big-4 audit consulting (Deloitte AI Governance, EY AI Risk), the ACLU Tech and Liberty project, and the Algorithmic Justice League. Starting salary is around $110k to $180k. The COMPAS and EEOC four-fifths analysis you produced today is what an AI auditor delivers every day.

Save your work: Save your bias audit report and fairness standards - these frameworks apply to every AI system you will ever encounter!

I completed Day 7

Built by Woven Learning & Technology✎ Edit

Day 8: The Black Box Problem

AI Transparency, Explainability, and the Future

Part I: The Hook

Welcome to Day 8, the close of Week 2. Today you will confront the fundamental tension at the heart of clinical AI.

Some of the most powerful AI systems in the world cannot explain how they reach their conclusions. When an AI says a patient has cancer or a defendant will reoffend, the people affected deserve to understand why.

Today's Case Briefing: Week 2's clinical case finale. Your mission: take a complex patient case and apply every Week 2 skill (cardiac physiology, surgical reasoning, diagnostic imaging, AI bias auditing, black-box explainability) to render a defensible diagnosis AND a court-ready position paper on the AI's role in your conclusion. The Black Box Problem comes home today.

I read the case briefing and watched the Explainable AI video

Part II: The Physical Lab

Woven notebook: this is Part II, The Physical Lab. Record every measurement, calculation, and observation as you work. The lab data you capture here becomes the evidence base you defend in Part III.

Materials Needed: Laptops or Chromebooks, lab notebooks. Every app you need today is embedded right next to the step that uses it, no scrolling around. The patient case, the treatment options, and all the data are on this page.

Today's Patient: David, age 47

Patient briefing: David is 47, a construction supervisor in Hayward. He walked into the ER 20 minutes ago with crushing chest pain that started on the job site, spreading into his left arm and jaw. He is sweaty, nauseated, and short of breath. Vitals on arrival: heart rate 118 (fast), blood pressure 162 over 94 (high), oxygen 92 percent on room air (low), breathing rate 24 (fast). His EKG (12-lead) shows ST-segment elevation in leads II, III, and aVF, with matching depression in leads I and aVL. History: high blood pressure (on lisinopril), high cholesterol, ex-smoker (quit 2019). Father had a heart attack at age 52. First troponin (the cardiac muscle blood test): elevated. Chest X-ray: heart slightly larger than normal, lungs look clean. Your team has 4 stations to work, then a 3-minute presentation.

Station 1: Read David's EKG

David's EKG shows ST elevation in leads II, III, and aVF. That pattern means the lower wall of his heart is not getting enough oxygen, an INFERIOR MI (heart attack in the bottom of the heart). Your job: confirm the pattern using the EKG Explorer below, then identify which artery is blocked.

1Open the EKG Explorer above. Switch to the STEMI rhythm. The elevation lives in the ST segment (the flat line between QRS and T wave). Match David's leads (II, III, aVF) to the heart wall they look at: these 3 leads view the INFERIOR (bottom) wall. Inferior MI = blockage of the RIGHT CORONARY ARTERY (the artery that feeds the bottom of the heart). Write the artery name in your notebook.

2Quick Verdict (3 bullets in your notebook): (1) the blocked artery is the [name], (2) this counts as a STEMI emergency because [why], (3) the door-to-balloon target is [time]. Two sentences max per bullet. The 90-minute door-to-balloon clock is what an ER cardiologist quotes from memory.

Station 2: Pick the Treatment

David needs his blocked artery OPENED. There are 3 treatment options. You are not the surgeon, but you will defend which option you would pick if you were.

The 3 options. PCI (percutaneous coronary intervention): a tiny balloon-and-stent threaded through a wrist or groin artery up to the heart, opens the blockage in 30 to 60 minutes if a cath lab is available. Standard of care if PCI can be done within 90 minutes. FIBRINOLYTICS (clot busters): IV drugs that dissolve the clot, faster to start than PCI but only about 60 percent successful, used when PCI is not available within 90 minutes. CABG (coronary artery bypass graft): full open-heart surgery, used when arteries are too damaged for PCI. Hours-long, much bigger recovery.

3Quick Verdict (3 bullets): (1) for David, pick PCI / fibrinolytics / CABG, (2) defend the pick in 2 sentences using the 90-minute door-to-balloon constraint, (3) name 1 thing about your hospital infrastructure that decides the pick (cath lab on site? helicopter transfer needed?). Two sentences max per bullet.

Station 3: Read David's Chest X-Ray

David's chest X-ray shows MILD CARDIOMEGALY (the heart looks slightly bigger than normal). That is expected with high blood pressure and prior heart strain. The X-ray does NOT change the STEMI diagnosis but it confirms David has been dealing with cardiac stress for a while. The Radiology Detective app is right below, practice the 4-step look (orientation, bones, soft tissue, air spaces) on any chest case to refresh.

4Work any chest X-ray case in the Radiology Detective above (Maya's swallowed coin works). Use the systematic 4-step look. Then come back and answer in your notebook: looking at David's CXR finding (mild cardiomegaly), what is ONE other thing you would want to check before treatment to be safe? (Hint: a bedside echocardiogram for wall motion, or repeat troponins to track the trend.)

5Quick AI bias check (1 sentence): cardiac AI tools (ECG-FM, AliveCor) UNDER-DETECT inferior MI in women and Black patients. The Mayo Clinic 2024 study showed a 14-percentage-point gap. If David were a 47-year-old Black WOMAN with the EXACT same symptoms and EKG, what would you do differently? Write 1 sentence in your notebook. (Hint: do not trust the AI alone, get a human cardiologist's read.)

Station 4: Can the AI Defend Itself?

Imagine David's family is suing the hospital, claiming the AI missed a complication. The lawyer asks the cardiac AI: 'why did you flag THIS as a STEMI?' The AI must answer in court. Today's question: WHEN can an AI's reasoning be defended in court, and when can it not?

6Open the AI Inspector above and tap Black Box Inspector. Pick a denied loan applicant. Hit Open the Box, see the SHAP waterfall (the bars showing which features pushed the decision up or down). The SAME kind of explanation IS what cardiac AI vendors must provide for a court. If the cardiac AI can show its reasoning like this, it can be defended. If it cannot, it cannot be admitted.

7Quick Verdict (3 bullets): (1) what feature most influenced the AI's decision in your loan example, (2) would you trust this AI in a courtroom? Why or why not, (3) what would the AI need to add to be more trustworthy?

Final Presentation

83-minute team presentation, 3 slides only. Slide 1: David's diagnosis (inferior STEMI, blocked artery). Slide 2: your treatment pick and why. Slide 3: where the AI helped vs where you did not trust it. Just 3 slides, just 3 minutes.

I worked all 4 stations and presented David's case

Part III: AI & Digital Literacy

Bleeding-edge: in March 2026, the first criminal case in the U.S. was overturned because of an unexplainable AI prediction. The judge ruled that the defendant's right to confrontation (6th Amendment) was violated when the prosecution used an AI score it couldn't explain. The case is now headed to circuit court. The legal precedent being formed RIGHT NOW will define how explainable AI must be for the next decade.

Open the Black Box with the AI Inspector + Canvas

Captum, SHAP, and LIME all answer the same question: 'why did the AI make this decision?' Today you do the SHAP-style analysis directly in the AI Inspector's Black Box Inspector mode (embedded right below), then vibe-code your court-ready report in Gemini Canvas. No external dependencies.

1Open the AI Inspector above in Black Box Inspector mode. Pick a denied loan applicant. Click 'Open the Box.' Read the SHAP-style waterfall: each feature's contribution to the deny decision is a bar above or below zero.

2Counterfactual analysis: in the What-If panel, flip ONE feature value at a time. Watch the decision change. Find the smallest change that flips DENIED to APPROVED. This is the 'minimum feature change' that explains the decision boundary.

3Proxy bias check: the dataset has a near-twin pair (two applicants who differ only in ZIP code). Find them. Run the Black Box on both. Does the prediction change? If yes, ZIP code is a proxy for race or class, a feature that LOOKS neutral but encodes a protected attribute. This is one of the most common ways modern AI quietly discriminates.

4Vibe code your court-ready report. In Gemini Canvas: 'Build a single-page web app with 4 sections: SHAP Findings (the top 3 features driving the decision), Counterfactual Analysis (the minimum change to flip the decision), Proxy Bias Test (whether a near-twin pair was treated differently), and Recommendation (Approve / Modify / Reject the model for production use). Use a clean two-column legal-brief layout.'

Ship It Live: Deploy to Netlify

5In Gemini Canvas, click the download icon (or hit the three-dot menu and pick Export HTML). You will get a single .html file with everything baked in, no separate CSS or JS files.

6Open netlify.com in a new tab. Click 'Sign up' or 'Log in.' Use your Google account (fastest) or sign up with email. Free tier handles everything you need today.

9Position paper (1-2 pages, the final assignment of the workshop): drawing on the EU AI Act Article 52 (transparency) and the 2026 confrontation-clause case from this morning's case briefing, argue: should explainability be REQUIRED for all high-risk AI? What about low-risk? Where's the line? Cite SHAP, LIME, and at least one published case. Submit alongside your Canvas-built report.

What you've built across 8 days: AI literacy at the level of a junior auditor. The skills - prompt engineering, bias auditing, explainability analysis, vibe coding, citation verification - are the same skills that get hired at Anthropic, Scale, Lex Machina, Casetext, and the ACLU's Tech and Liberty team. You are 4 years away from those interviews. You start qualified now.

Media Literacy Field Card: you've now learned the auditor's full toolkit. Bookmark these for life: (1) AP Fact Check, (2) Snopes, (3) NewsGuard, (4) AllSides, (5) SIFT method, (6) Lateral reading, (7) Stanford COR, (8) Common Sense Media's News Literacy curriculum. These are your free defenses against the next 50 years of misinformation.

Launch the AI Inspector (Black Box Inspector)

Final AI literacy session. The Black Box Inspector lets you open a real ML decision: would this loan be approved? Then run a what-if to flip the decision and see what feature drove it. Same tools (SHAP, LIME, counterfactuals) used by every responsible AI auditor in industry.

10Open the app, tap 'Black Box Inspector.' Pick from 4 loan applicants - one is a near-twin pair where only ZIP code differs. Hit 'Open the Box' to reveal the SHAP-style waterfall showing each feature's contribution to the approve/deny decision.

11Run a 'What-If' counterfactual. Flip a feature value and watch the decision change. The proxy bias (ZIP code as a stand-in for race) becomes visible in real time.

12Take the final quiz covering hallucinations + bias + black box. This is your full AI literacy assessment for the workshop.

SHAP, LIME, and attention maps are the closest tools we have to opening the black box. None gives full explanation - but each narrows the question. That's the entire field of explainable AI.

I participated in the debate and wrote my position paper

Part IV: End of Day

Career Connection: AI Ethics Counsel and Explainable AI Researcher

AI Ethics Counsel and Explainable AI Researchers work at Anthropic, OpenAI Policy, EU AI Act compliance consulting, and ACLU AI projects. Salaries run roughly $140k to $240k. The SHAP, counterfactual, and proxy-bias analysis you wrote today is the explainable-AI portfolio piece that auditors and lawyers submit to regulators.

Save your work: Congratulations on completing the Chabot High School STEM Workshop! You now have the critical thinking tools to navigate an AI-powered world. Keep questioning, keep analyzing, keep demanding transparency.

I completed Day 8

Built by Woven Learning & Technology✎ Edit

Day 1: Advanced Biometrics

Part I: The Hook

Part II: The Physical Lab

Part III: AI & Digital Literacy

Train Your Own Fingerprint Classifier (Daubert-Grade Audit)

No webcam? Upload your images instead

Now: Vibe-Code a Stats Calculator

Ship It Live: Deploy to Netlify

Part IV: End of Day

Career Connection: Forensic Latent Print Examiner

Day 2: The Physics of Blood Spatter

Part I: The Hook

Part II: The Physical Lab

How to Run the Spatter Experiment

Part III: AI & Digital Literacy

Vibe-Code a Spatter Calculator

Ship It Live: Deploy to Netlify

Part IV: End of Day

Career Connection: Bloodstain Pattern Analyst

Day 3: Multi-Modal Analysis

Part I: The Hook

Part II: The Physical Lab

The Pine Hills Case (your team's working case)

Launch the Evidence Synthesizer

Part III: AI & Digital Literacy

Run Your Own Multi-Modal Analysis

Part IV: End of Day

Career Connection: Forensic Lab Manager and Senior Evidence Analyst

Day 4: AI Hallucinations

Part I: The Hook

What Are Deepfakes?

Part II: The Physical Lab

The AI Testing Protocol (use this for all 5 steps below)

8 Test Questions (with known correct answers, kept in the facilitator copy)

Verification Sources (cross-check against these)

Categorize Each Response

Part III: AI & Digital Literacy

Vibe-Code a Citation Verifier

Ship It Live: Deploy to Netlify

Media Literacy: Apply SIFT + Lateral Reading

Launch the AI Inspector (Hallucination Hunter)

Ethics Reflection

Synthetic Media Literacy

Step 1: Study the Digital Truth Checklist

Step 2: Test AI Images

Step 3: Listen to Voice Clips

Step 4: Triage Headlines and Social Posts

Part IV: Week 1 Case Finale

The Pine Hills Case (Week 1 finale)

Step 1: Apply Day 1, The Fingerprint

Step 2: Apply Day 2, The Blood Spatter

Step 3: Apply Day 3, The Trace Evidence

Step 4: Apply Day 4, The Digital Evidence

Step 5: Render the Verdict

Part V: End of Day

Career Connection: AI Trust and Safety Engineer and Legal Tech Auditor

Day 5: Clinical Physiology

Part I: The Hook

Watch First: How the Heart Pumps Blood

Part II: The Physical Lab

SOAP Note Template (use this format for every clinical writeup today)

3 Clinical Scenario Cards (work each one)

First: How the Heart's Electrical System Works

Next: Two Common Pathologies

Now: Take Your Own Vitals

Korotkoff Sounds Explained (before you take BP)

Pulse Oximetry

Exercise Recovery Challenge

Part III: AI & Digital Literacy

Probe Gemini for Cardiac AI Demographic Bias

Ship It Live: Deploy to Netlify

Part IV: End of Day

Career Connection: Cardiac AI Safety Lead and Cardiologist

Day 6: Advanced Surgical Skills

Part I: The Hook

Part II: The Physical Lab

Sharps Safety: read this BEFORE you touch a needle

The Tools You'll Use

Watch First: How a Vertical Mattress Suture Works

Launch the Surgical Dexterity Trainer