Computer Science Innovation and Entrepreneurship EN.660.410.01.FA26 -- Class Opportunity Bank

AI Trust and Verification Infrastructure combines 5 clusters, spanning 10 of 20 domains, extended in Week 5

Strong Signal

This groups three related clusters -- provenance/chain-of-custody, machine legibility, and confidence-gating -- because they're all versions of the same problem: proving something is real, unaltered, or safe before someone downstream acts on it. Chain-of-custody already exists as a formal discipline outside anything the class touched: pharma has it for anti-counterfeiting, diamonds and luxury goods have it for authenticity certification, courts have it for evidence, blood and organ transport have it because a broken chain is fatal, and food safety has it as cold-chain tracking. AI adds two problems those older disciplines never had to solve: content provenance (proving a video or image wasn't generated) and machine legibility (making something only a person could judge into something a system can query).

A team could take an existing, well-understood chain-of-custody discipline -- pharma, diamonds, courtroom evidence -- and rebuild it for the AI era, or take the newly urgent problem (media provenance, machine-queryable legibility) and build the standard nobody's built yet. Confidence-gating -- blocking an irreversible action until confidence clears a threshold -- is the enforcement layer this whole area eventually needs, whichever piece you pick.

Week 3 added a fourth cluster, "Gate the Act, Don't Just Log It," which is the confidence-gating piece applied to oversight. The Agency Under Surveillance team found that every existing control on plate-reader searches is an audit log, and a log only reports what already happened: one deputy searched one woman's plate 1,639 times before anyone noticed. The Cybersecurity team found that attackers now break out of a network in 29 minutes, faster than any after-the-fact review of an AI agent's actions. Both teams concluded that the check has to sit in front of the action. The other surveillance team's privacy-first plate readers rely on logging every search, so the class now holds both positions.

Week 5 added a fifth cluster, "Receipts for Clips and Agent Commitments," from three of the four Communication teams. The Feed Is the Newsroom team found that a political clip loses its source, date, and corrections as it moves through the feed, and picked a receipts layer that links any clip back to the original at the moment quoted. A provenance standard (C2PA) already exists for this, but the team reports that platforms strip its data on upload. The Machine-to-Machine team found the same gap for AI agents: open protocols such as Agent2Agent carry messages between agents but not proof of who sent an agent or on whose authority. The When No One Writes team named provenance and authorship as its first gap, since nothing shows how much of a workplace message a person actually stands behind. Two of the three teams used the word "receipt" independently.

The three original clusters -- Proof of Origin at Every Handoff, Making an Opaque System Remotely Legible, and Confidence-Gated Decision Layers -- surfaced independently across seven of the first eight domains the class explored: Smart Cities, AI Media, Social Media, Legal, Travel, Defense, and Accessibility. The Week 3 cluster added Surveillance and Cybersecurity, and the Week 5 cluster added Communication, for ten of the twenty domains explored so far. Sophia and Sylvia, on the Week 2 Social Media team that proposed verified-source labels, were on the Week 5 Machine-to-Machine team.

Domains: Smart Cities / Rapid Delivery AI-Generated Media and Entertainment Addictive and Opaque Social Media Algorithms Legal Travel Defense and National Security Accessibility and Assistive Technology Surveillance and Personal Data Cybersecurity in the Age of AI Agents Communication

AI Governance, Liability, and Insurance combines 6 clusters, spanning 9 of 20 domains, extended in Week 5

Strong Signal

This groups the class's "AI Outgrows the Institutions Built to Govern It," "Liability With No Holder," and "Verify-Before-Deploy" clusters, because all three are angles on the same open question: who's liable when an AI system is wrong, and how does anyone price that risk. Every industry that has adopted AI for a consequential decision is running into it. Healthcare has it with diagnostic algorithms. Auto insurers have it with autonomous vehicles. Hiring has it with resume-screening tools already facing discrimination lawsuits. Criminal justice has it with sentencing risk-assessment tools. Financial services has it with robo-advisors and automated underwriting. None of these have settled case law, and none have a mature insurance product.

A team could pick any one of those verticals -- healthcare AI liability, autonomous-vehicle insurance, hiring-algorithm audit compliance -- and build the accountability ledger or certification product for it specifically. Legal and Defense already proved two working versions of the underlying business model (verify the output, sell the verification to whoever underwrites the risk); the opportunity is applying that model somewhere nobody's tried yet.

Week 3 added two clusters. "Verify Before It Acts" joined the Who Watches the Watchers team, which wants independent accuracy and bias tests before a predictive policing model scores anyone (one model's predictions matched under 1% of crimes), with the Cybersecurity team, which found that nothing currently detects when an AI agent reaches its goal by a route nobody intended. "Risk No One Can Price Yet" came from the Supply Chain team's pick: no instrument lets importers, carriers, or marine insurers price or hedge the closure of a strait, even as Hormuz traffic fell from 100+ ships a day to about 5. That is the same pricing gap outside AI entirely, which suggests the underlying business (measure a new risk, sell the measurement to whoever underwrites it) is not limited to AI. Three of the four Supply Chain students were on the Week 2 teams behind "Liability With No Holder," so part of this convergence is the same people carrying an idea into a new domain.

Week 5 added "Who Answers for What an Agent Agreed To." The Machine-to-Machine team found that when one AI agent books or buys from another, the human click on "Buy" that used to prove consent is gone, and the Agent Payments Protocol (AP2) records the payment but not the terms agreed. Its well-defined problem is a plain-language receipt for every commitment an agent makes, with an alert when it goes past what the person approved. The When No One Writes team named the same gap from the workplace side, asking who owns what two agents agreed to when no person read the exchange. Both teams name firms that will carry the liability for agent errors, which is the pricing question this card started with.

The three original clusters surfaced across Public Approval, Social Media, AI Media, Legal, and Defense. The two Week 3 clusters added Surveillance, Cybersecurity, and Supply Chain, and the Week 5 cluster added Communication, for nine of the twenty domains explored so far.

Domains: Public Approval of AI Addictive and Opaque Social Media Algorithms AI-Generated Media and Entertainment Legal Defense and National Security Surveillance and Personal Data Cybersecurity in the Age of AI Agents Supply Chain and Logistics Communication

Personalization Built From the Edge Case Instead of the Average User combines 4 clusters, spanning 8 of 20 domains, extended in Weeks 5 and 6

Strong Signal

This groups "Declared Intent vs. Inferred Behavior" and "One-Size-Fits-All Personalization Breaks Down at the Edges." Most personalization gets built around an average user, then generalized outward -- and that ordering quietly breaks the product for anyone off the average. Healthcare treatment plans built around a clinical-trial-average patient underperform for anyone outside that population. Education software built for a typical learner fails students with different processing speeds. Financial products built around a standard credit history exclude the underbanked entirely. In every one of these fields, flipping the build order -- design for the hardest case first, generalize outward -- has repeatedly produced a better product for everyone.

A team could pick a personalization product built around the average case in a field that hasn't had this reckoning yet -- lending, education, healthcare, hiring -- and rebuild its interface starting from its hardest edge case. Or build the declared-preference layer as a standalone product: let people explicitly state a goal instead of a system inferring one from behavior it might be reading wrong.

Week 5 added "Choosing What the Feed Is For." The Many-to-One team found that feeds generally offer two settings, engagement-ranked or plain chronological, and that nothing lets a person choose what the feed optimizes for. That is the declared-preference layer this card describes, and it is close to Week 2's user-set feed goals; Jingchen was on both teams. The Many-to-One team also brought evidence that complicates the simple version: in the 2020 Meta experiments, the chronological feed showed more content from untrustworthy sources, so removing the ranking is not the same as letting a person set its goal. The EU Digital Services Act already requires the largest platforms to offer a feed not based on profiling, which gives a team a place to see what people do when they are offered a choice.

Week 6 added "Let the User Say What to Optimize." The Transport team found that routing apps rank by time while travelers want different weightings at night, with a child on board, or on a bike, and that an 8% faster route was found to carry 23% higher crash probability, a trade-off the traveler never sees. The Wearable team found that users can't tell what is normal for them from population ranges, and proposed personal baselines that flag a pattern only when it stays outside that person's own range for several nights. Both are versions of the declared-preference layer above: let people state what the system should optimize for them instead of applying one default to everyone. The Transport team also named the limit: a 2012 patent for routing around high-crime areas was criticized as digital redlining, so what a user is allowed to optimize for needs its own rules. The Cold Chain team found the same problem for products rather than people: one generic temperature threshold creates false alarms for hardy produce and misses risk for dairy or seafood, so the alert has to be judged against the specific product's range.

Two class teams reached this from opposite failure modes -- Social Media found inference actively wrong, Accessibility found the generic interface actively excluding -- and Week 5's Many-to-One team returned to it in Communication. Week 6's Transport, Wearable, and Cold Chain teams added three more, for eight of the twenty domains explored so far. Jingchen was on the Social Media, Many-to-One, and Transport teams, and Kendall and Sissi on both the Accessibility and Wearable teams.

Domains: Addictive and Opaque Social Media Algorithms Smart Cities / Rapid Delivery Travel Accessibility and Assistive Technology Communication Transportation and Mobility Wearable Health Tracking Food Systems and Agriculture Technology

Letting People See and Dispute the Scores Systems Keep About Them 1 cluster, spanning 2 of 20 domains, new in Week 3

Strong Signal

New in Week 3 as "Scored by a Profile You Can't See." Both surveillance teams, working separately, found people judged by scores they never see: government systems that classify and predict without saying what data shaped the decision, and private scores built from cars, faces, purchases, and chats that can end up with police. Week 2's Social Media team found the same thing in recommendation feeds, where the platform infers what you want and you can't see or correct the profile. Outside the class there is one mature precedent: credit reporting, where U.S. law has long given people the right to see their file and dispute errors. Most other scores that shape real decisions have nothing comparable, including retail face-recognition watchlists, workplace productivity scores, and the risk scores predictive-policing tools assign to people and places.

A team could pick one kind of score that already shapes decisions -- a retail watchlist, a driving-behavior insurance score, a workplace productivity metric -- and build the see-and-dispute layer for it. The open design question the Who Watches the Watchers team raised is who plays the role of the credit bureau when the scores are spread across thousands of companies.

Surfaced by both Week 3 surveillance teams and by Week 2's Social Media team, across two of the twenty domains explored so far. The two surveillance teams disagree on whether visibility is enough: Agency Under Surveillance argues that the audit logs already existed and did not stop the misuse.

Domains: Surveillance and Personal Data Addictive and Opaque Social Media Algorithms

Distributed Node Networks Replacing Centralized Infrastructure combines 3 clusters, spanning 4 of 20 domains, extended in Week 4

Strong Signal

This groups "The Neighborhood as a Distributed Network" and "Cheap Distributed Sensing Beats Expensive Centralized Systems." The underlying pattern -- many cheap, distributed, software-coordinated nodes beating one expensive centralized asset -- is one of the oldest architectures in infrastructure, and it keeps reappearing as sensors and connectivity get cheap enough for a new field to try it. Disaster response is moving from centralized command centers to self-organizing mesh communication after a hurricane or earthquake knocks out central infrastructure. Clean energy is moving from centralized power plants toward distributed microgrids. Agricultural monitoring for smallholder farmers is shifting from expensive centralized satellite contracts to cheap, distributed ground sensors.

A team could take the architecture directly into any of those fields, or build the general coordination software -- the fusion layer that makes a mesh of cheap, independent nodes act like one reliable system -- that any of them could plug into.

Week 4 added "The Grid as a Network of Cells." The Grid Modernization team, starting from Baltimore's utility, proposed turning each neighborhood into a cell that balances its own solar, batteries, and EVs, can island itself when the main grid fails, and shares power with neighboring cells over links that software controls. That is the clean-energy version this card already pointed to. The team also named the weakness that comes with it: more software on the grid means a larger attack surface, which is the Single Points of Control problem below, now inside the coordination layer.

Three class teams independently rebuilt pieces of this architecture -- a counter-drone sensor mesh, a peer-to-peer delivery network, and Week 1's Travel readability layer. Week 4's Grid team added a fourth, a network of neighborhood power cells, for four of the twenty domains explored so far.

Domains: Smart Cities / Rapid Delivery Travel Defense and National Security Grid Modernization

Single Points of Control That Become the Target combines 2 clusters, spanning 4 of 20 domains, new in Week 3

Strong Signal

Combines Week 3's "Single Points of Control Become Targets" with Week 1's "Redundant Multi-Vendor Sourcing Beats Single-Supplier Lock-In." The Supply Chain team found that most of the world's goods move through a few sea lanes and depend on one country for containers, port cranes, and new ships, so a single closure reprices everything downstream. The Cybersecurity team found that once AI can be shut down, the shutdown mechanism (the weights, the insiders, the key-holders) becomes the one thing an attacker needs to reach. Both describe a system that depends on one point of control, which then becomes leverage for whoever can reach it. The same problem is familiar outside the class: rare-earth processing concentrated in one country, advanced chips concentrated in a handful of fabs, and certificate authorities whose compromise breaks trust for everyone who relies on them.

A team could map the single points in one supply chain or control system before they fail, build the redundancy (a second source, a split key, a pre-arranged alternate route), or price the exposure for whoever carries it. Week 1's Travel team already built one version of the remedy for airline seats: capacity pre-arranged across several carriers so that one failure can't strand anyone.

Surfaced independently by Supply Chain and Cybersecurity in Week 3, extending Travel and Defense's Week 1 cluster, across four of the twenty domains explored so far.

Domains: Supply Chain and Logistics Cybersecurity in the Age of AI Agents Travel Defense and National Security

AI-Mediated Advocacy and Access for People Too Small or Underserved to Get Individual Attention combines 3 clusters, spanning 3 of 20 domains

Strong Signal

This groups "AI-Mediated Advocacy for Parties Too Small to Get Attention," "Bypassing the Gatekeeper That Traps the Powerless," and "High-Stakes Real-Time Communication for People Who Can't Wait." AI makes a specific kind of business newly viable: institutional-scale advocacy for someone whose case was never profitable enough for a human expert to take individually. That extends well past debt defense and small travel destinations. Immigration and asylum cases are overwhelmingly unrepresented because individual cases don't clear the bar for a lawyer's hourly rate. Medical billing disputes go unchallenged by patients who don't know they're allowed to fight a denied claim. Disability and veterans' benefits claims get denied on paperwork technicalities a specialist would have caught.

A team could pick any one of those -- immigration, medical billing, benefits claims, special education -- and build the AI-mediated advocacy layer for it specifically. The pattern repeats exactly: give one person the scale of institutional representation that used to require paying a specialist by the hour.

Legal and Travel proved two working versions of this business model in Week 1; Accessibility's real-time communication work solves the delivery half of the same problem, across three of the twenty domains explored so far.

Domains: Accessibility and Assistive Technology Legal Travel

Localized Infrastructure Costs of AI and Automation 1 cluster, spanning 2 of 20 domains

Strong Signal

Formerly "Convenience's Costs Land Locally." Any system with a remote benefit and a local, physical cost creates the same fight, and AI-datacenter buildout is only the newest version of it. Cryptocurrency mining already fought this battle over local grid strain and electricity rates. EV charging buildout is straining local distribution grids in specific neighborhoods faster than utilities can upgrade them. Semiconductor fabs consume enormous volumes of water, often in drought-prone regions, for a global supply chain that never touches the local community.

A team could pick a resource fight that's already happening in a specific place -- data-center water and power in a specific region, EV charging strain on a specific grid, fab siting in a drought-affected area -- and build the pricing, governance, or transparency layer for that fight specifically.

It surfaced independently in Public Approval's top-ranked pick and two separate Smart Cities opportunities, across two of the twenty domains explored so far.

Domains: Public Approval of AI Smart Cities / Rapid Delivery

Simulated Training and Practice Infrastructure for Disappearing Apprenticeships combines 3 clusters, spanning 6 of 20 domains, extended in Weeks 4 and 5

Strong Signal

Formerly "Rebuilding the Low-Stakes Practice Layer." Every field with an apprenticeship model is losing its low-stakes practice layer to automation at roughly the same rate, not just software engineering. Medical residents already train on cadavers and simulators because real patients can't absorb the risk of a first attempt -- that same logic is now overdue in fields that never had it. Pilots train on flight simulators; drone and autonomous-vehicle operators mostly don't have an equivalent yet. Skilled trades still rely on shadowing a senior tradesperson, a model that breaks as fewer seniors exist to shadow.

A team could pick one of these under-simulated fields and build its practice environment specifically, or build a general-purpose simulation platform -- the digital-twin approach Defense already validated -- that multiple fields could plug their own scenarios into.

Week 4 added a second cluster, "Where Experience Gets Made," from three teams working in different domains. The Junior Jobs team found that a junior engineer used to pay for their own training through routine work, and once AI took the routine work, firms stopped hiring at the level where experience is created: new-grad hiring at the largest tech companies is down 65% from 2019, while engineering hiring overall fell about 11%. The Construction team found the same gap in the trades, where most of the 349,000 workers needed in 2026 replace retirees and a new tradesperson takes years to reach full productivity. That is the skilled-trades case this card already predicted. The Higher Education team found it from the school's side: about 70% of employers say new graduates still need substantial training, and a degree teaches everything once, between 18 and 22. The open question all three left is who pays. The Junior Jobs team points to medical residency, which works partly because public funding covers much of it.

Week 5 added a third cluster, "Practice That Disappears When Communication Is Delegated." The Machine-to-Machine team found that AI agents now make the calls and small negotiations people used to learn from, and that nothing helps a person decide which of those conversations to keep doing themselves. The When No One Writes team found the same loss in workplace writing: writing is how people work out what they think, and an organization that hands it to AI may lose that judgment. This moves the card from work skills to everyday communication skills, where the low-stakes practice was never formally organized in the first place.

It showed up in the class's own work as both a talent-pipeline problem (Public Approval) and a systems-testing problem (Defense), independently, a week apart. Week 4 added Junior Jobs, Construction, and Higher Education, and Week 5 added Communication, for six of the twenty domains explored so far. Three of the four Week 2 Public Approval students carried the idea into Week 4 (Sissi and Evan on Construction, Kerry on Higher Education), so part of this convergence is the same people returning to an idea.

Domains: Public Approval of AI Defense and National Security Disappearing Junior-Level Software Jobs Construction Inefficiencies Higher Education and Curriculum Communication

Judgment Is the Skill AI Leaves Behind, and Grades No Longer Show It combines 2 clusters, spanning 3 of 20 domains, new in Week 4

Strong Signal

Combines Week 4's "Judgment Is the Skill AI Leaves Behind" and "The Signal Stopped Carrying Information." The Junior Jobs team found that as AI makes code cheap to generate, the ability to evaluate, debug, and decide about code matters more than producing it: 84% of developers use or plan to use AI, and 46% distrust its accuracy. The Higher Education team found that no course owns judgment or communication, where employers report 43-point gaps between how important these skills are and how proficient graduates are, and that assessment drifts toward whatever is easy to grade. The second cluster is about proof. Harvard's share of A grades rose from about 25% in 2005 to over 60%, and employers screening by GPA fell from 73% to 42%, with nothing replacing it. That is the same problem Week 2's AI Media team found for content: once a signal is cheap to produce, it stops telling anyone anything, and the value moves to whatever can certify the real thing.

Outside the class, fields that depend on judgment already assess it directly rather than through grades: medical schools use observed clinical exams, architecture requires supervised practice hours before licensing, and flight training grades decisions in a simulator. A team could build an assessment that grades how someone reached an answer with AI, not just the answer, or a portable record of demonstrated work that an employer can trust more than a GPA.

Surfaced by Junior Jobs and Higher Education in Week 4, with the signaling half extending Week 2's AI Media work, across three of the twenty domains explored so far.

Domains: Disappearing Junior-Level Software Jobs Higher Education and Curriculum AI-Generated Media and Entertainment

When the Constraint Is the Institution, Not the Technology combines 2 clusters, spanning 5 of 20 domains, extended in Week 6

Strong Signal

New in Week 4. This week's method asked each team to look above and below its focal point before deciding where to work, and three teams found the limit in an institution rather than a tool. The Construction team showed that machines, materials, and software improved sharply since 2000 while productivity grew about 1% a year, because every building is redesigned, re-permitted, and re-contracted from scratch. Katerra spent more than $2B on factories and an integrated supply chain and went bankrupt on state-by-state codes and contractors who would not change subcontractors. The Grid team listed institutional inertia as what is stuck at the distribution level, while renewables arrive faster than the grid adapts. The Higher Education team found that 39% of core skills change by 2030 while a curriculum change takes about two years to reach the catalog.

Outside the class, the fixes that worked for this kind of problem changed the institution's process rather than the technology. Some cities offer pre-approved standard plans for backyard cottages, so each one skips a fresh permit review, and utility regulators in several states have started paying utilities for performance rather than only for the capital they spend. A team could pick one approval or contracting step that every project repeats and build the pre-approved, reusable version of it.

Week 6 added "The Standard Exists, but Few Agencies Publish to It." The Transport team found that the data two of its opportunities need is already defined: the Curb Data Specification describes curb rules in machine-readable form, and the Work Zone Data Exchange has fed navigation providers since 2020. Only 40 work-zone feeds were active nationally as of June 2026, and cities such as Boston and Seattle are only now converting their curb rules. The permit behind a work zone exists weeks earlier with a city or state agency, outside the data chain entirely. As with Katerra, the technology is not what is missing; the constraint is publishing that is municipal and fragmented. The Cold Chain team reached the same conclusion in food logistics: real-time temperature sensing is mature, and what is missing is authority to act, agreed in advance in the contracts between shipper, carrier, and receiver.

Surfaced independently by Construction, Grid, and Higher Education in Week 4, and by Transport and Cold Chain in Week 6, across five of the twenty domains explored so far. The Construction team's Katerra case is the clearest evidence the class has for this pattern.

Domains: Construction Inefficiencies Grid Modernization Higher Education and Curriculum Transportation and Mobility Food Systems and Agriculture Technology

Restoring a Credible Signal That a Message Matters 1 cluster, spanning 3 of 20 domains, new in Week 5

Strong Signal

New in Week 5. Two Communication teams found the same problem from opposite ends of a message. The When No One Writes team used costly-signaling theory to explain it: a carefully written note used to prove that the sender cared, because writing it took effort, and once AI makes writing free, effort no longer separates an important message from a routine one. The team's notes call a replacement signal for "this matters" its most original gap, and its strongest data is on the receiving side: 40% of US desk workers received AI "workslop" in the past month, at about two hours each to deal with. The Many-to-One team found the sender's side of the same problem: schools, clinics, employers, and families compete with endless content and generally have no way to tell whether an important message landed, while knowledge workers face about 275 interruptions a day.

This is the same pattern as Week 4's "The Signal Stopped Carrying Information," where grades stopped telling employers who is ready, and Week 2's AI Media work, where synthetic content stopped proving that anything happened. Outside the class, some systems already keep a costly or scarce signal on purpose: certified mail and signature confirmation for legal notices, emergency alert channels that only authorized senders can use, and limits on how many high-priority flags a person can send. A team could pick one setting where important messages are generally lost among routine ones, such as a school's messages to families or a clinic's follow-up after a test result, and design a signal that stays credible because it is limited or costly to send.

Surfaced by When No One Writes and Many-to-One in Week 5, extending Week 4's signaling cluster from Higher Education and Week 2's AI Media work, across three of the twenty domains explored so far. Michael was on both the Week 4 Higher Education team and the Week 5 Many-to-One team.

Domains: Communication Higher Education and Curriculum AI-Generated Media and Entertainment

The Gap Sits at the Handoff Nobody Owns 1 cluster, spanning 5 of 20 domains, new in Week 6

Strong Signal

New in Week 6. This week's method asked each team to map a job step by step and rate where it is underserved, and all four teams found the worst step at a boundary. The Transport team found that waiting, transferring, and parking all happen between modes, and every current product (a map app, a transit agency, a scooter company) owns only one mode, so "nobody owns the handoffs." The Wearable team's largest gap was briefing the doctor: the user has weeks of their own data and a 15-minute visit with a clinician who rarely has tools to use it. The Education Technology team found the job of assigning, grading, and answering questions split across Canvas, Gradescope, and course forums. The Cold Chain team named the pattern most directly, "an alert without an owner": sensors detect a temperature problem in a refrigerated shipment in real time, but the alert does not reliably reach someone with the authority to re-route, re-ice, or reject the load, because that authority sits across three companies. Week 4's Construction team found the same pattern in project delivery, where design, permitting, and building are contracted separately and no single party is accountable for the finished outcome.

Outside the class, some fields have assigned an owner to the handoff. Hospitals use structured handoff protocols at shift change because errors cluster there. Airline interline agreements let one ticket cover several carriers, so the airline that sold the ticket is responsible when a connection is missed. Freight forwarders exist because no single carrier owns a shipment from end to end. A team could pick one handoff in a job it has mapped and design who owns it: a protocol, a shared record that crosses the boundary, or a party paid when the handoff succeeds. The Week 6 teams already split on the approach: Education Technology puts the whole job in one place, Wearable builds a better crossing between parties that stay separate, and Cold Chain agrees in advance who may act, so the handoff has an owner before anything goes wrong.

Surfaced independently by all four Week 6 teams, extending Week 4 Construction's fragmented workflows, across five of the twenty domains explored so far. Week 2 Smart Cities' chain-of-custody work (see AI Trust and Verification Infrastructure above) is the verification half of the same problem.

Domains: Transportation and Mobility Wearable Health Tracking Education Technology Food Systems and Agriculture Technology Construction Inefficiencies

Provenance and Authenticity Verification Across Delivered Packages, AI Video, and Sponsored Content Labels combines 3 connections, extended in Week 5

Combines "One Chain-of-Custody Standard for Parcels and Pixels" and "Motive Labels That Can Be Audited." Proving an instantly delivered package hasn't been tampered with, proving a video wasn't generated, and proving a "paid political ad" label is honest are the same underlying claim: something is what it says it is, made when and by whom it says. All three currently fail the same way -- the claim is self-asserted, and nothing separates an honest one from a cheap one. The fix is identical everywhere: a signed record written at the moment of creation or handoff, checked at each point the object or content changes hands. This kind of infrastructure already exists outside anything the class touched -- pharmaceutical anti-counterfeiting, diamond certification, courtroom chain-of-custody -- proof the pattern works, just not yet applied to delivery, media, or advertising.

A team could build a single audited-claims layer that certifies more than one type of claim at once ("this is human-made," "this hasn't been tampered with," "this is who paid for it"), since the underlying verification -- signed records, checked at handoff -- is identical regardless of which claim it's attached to. Delivery fleets, which already log every transfer and already face regulatory pressure to prove custody, are a more plausible first customer than media companies, which have less operational infrastructure to build on.

Week 5's Feed Is the Newsroom team found where the media version breaks in practice. A political clip loses its source, date, and corrections as it spreads, and although a provenance standard (C2PA) exists and TikTok joined it in 2024, the team reports that platforms strip its data on upload. Its pick, a receipts layer that opens the original video or document at the exact moment quoted, starts as a creator tool so that it needs no platform permission. That raises the question this card already pointed to: a self-applied label is only as honest as whoever applies it, and the creators least likely to add a receipt are the ones whose clips most need one.

Built from: Smart Cities' trust layer (Yash Chadda, Sichen Deng, Kendall Frederick, Jonathan Locala); AI Media's provenance gap and authenticity premium (Daniel Sung, Flora Huang, Steven Li, Owen Bartlett); Social Media's motive labels (Jingchen Jiang, Sahana Raja, Sophia Mai, Sylvia Yam); Communication's receipts layer (Feed Is the Newsroom, Team 2, names not on deck)

Using an AI Accountability Ledger to Underwrite AI Liability Insurance 1 connection

Formerly "The Tracker Feeds the Underwriter." Insurance can't price a risk it has no loss data for, and loss data for AI-caused harm barely exists because nobody consistently records who built, deployed, and approved the system that caused it. Public Approval's AI tracker is exactly that attribution record. Paired with AI Media's diagnosis -- that machine-scale publishing needs liability pricing before anyone absorbs its tail risk -- the tracker stops being just an accountability tool and becomes the data asset an insurance product gets built on.

The same pairing -- an attribution ledger plus an underwriter who needs it -- generalizes to any AI-driven decision with no insurance product yet: automated hiring calls, AI-assisted medical diagnoses, algorithmic loan denials. A team could build the ledger first, in whichever vertical has the most acute liability gap right now, and sell access to insurers directly rather than waiting for one to build its own.

Built from: Sissi Lai, Kerry Wang, Penn Riney, Evan Batten (Public Approval of AI); Daniel Sung, Flora Huang, Steven Li, Owen Bartlett (AI Media)

Applying Defense's Simulated-Practice Method to Fields Beyond Engineering combines 3 connections, extended in Weeks 4 and 5

Formerly "Simulated Apprenticeship." Nobody becomes senior without supervised, low-stakes practice, and AI is removing exactly that work in software engineering first. Defense's Week 1 digital twin -- replaying real system data so tools can be tested without touching a real ship -- is a ready-made answer to a problem Public Approval only diagnosed. Give junior engineers a realistic, disposable replica of production where they can break things, get reviewed, and build judgment, recreating the on-ramp automation closed.

The same fix applies anywhere else the low-stakes practice layer is thinning or never existed: medical residency already does this with cadavers and simulators; skilled trades never built an equivalent for apprentices who used to shadow a senior tradesperson. A team could take Defense's digital-twin method into one of those fields directly instead of staying in engineering, where the idea is more obvious but also more contested.

Week 4 brought both of the fields this card named. The Junior Jobs team found that new-grad hiring fell 65% because AI took the routine work juniors learned from, and proposed new pathways for building experience outside the traditional internship-to-junior pipeline. The Construction team found tradespeople retiring faster than their knowledge can be transferred. A disposable replica of production, or of a jobsite, is one concrete form those pathways could take.

Week 5's Machine-to-Machine team found a field that never had a formal practice layer: everyday conversation. As AI agents take over calls, bookings, and small negotiations, students and early-career workers lose the low-stakes conversations they used to learn from, and 72% of US teens have already used AI companions. A practice environment for conversation, where a person rehearses a negotiation or a difficult call against a realistic counterpart before doing it for real, would be the same method applied to a softer skill. Jonathan, Sophia, and Penn were each on a team behind one of the earlier connections on this card.

Built from: Sissi Lai, Kerry Wang, Penn Riney, Evan Batten (Public Approval of AI); Group 3 -- Sahana, Kerry, Michael, Penn, Evan (Defense and National Security, Week 1); Junior Jobs' new pathways for building experience (Owen Bartlett, Jingchen Jiang, Jonathan Locala, Sophia Mai, Daniel Sung); Construction's experience leaving before it transfers (Sissi Lai, Yash Chadda, Flora Huang, Steven Li, Evan Batten); Communication's who still practices talking (Jonathan Locala, Sophia Mai, Penn Riney, Sylvia Yam)

Applying Defense's Distributed-Sensor Architecture and Travel's Multi-Vendor Sourcing Model to Each Other combines 2 connections

Combines "Distributed Ground-Truth Mesh for Conflict/Disaster-Zone Travel" and "Redundant Multi-Vendor Defense Supply Mesh." Travel and Defense independently rebuilt the same architecture twice, in two directions. Defense's counter-drone mesh replaces one interceptor battery with many cheap radar and interceptor nodes; the same distributed-sensing logic, applied to safety data instead of radar, would give Travel a mesh of low-cost, locally placed reporting nodes that keeps trip-planners current in conflict or disaster zones no single "readable" feed could track. Running the other direction, Travel's cross-carrier rebooking pre-arranges swappable capacity across multiple vendors so no single failure can strand a traveler -- and Defense is exposed to the identical single-supplier risk on hardware without having built the fix yet.

A team could pick either direction -- port Defense's sensing architecture into unstable-region travel data, or port Travel's multi-vendor sourcing model into defense parts procurement -- and the same underlying redundancy pattern would carry either one.

Built from: Sylvia, Steven, Flora, Yash (Travel); Group 3 -- Sahana, Kerry, Michael, Penn, Evan (Defense and National Security)

Wiring Accessibility's Real-Time Communication Tools Into Legal Triage and Courtroom Hearings combines 2 connections

Combines "Accessible Crisis & Court Navigator" and "Live-Captioned Debt & Eviction Hearings." Legal's pretrial triage service reminds people of court dates and routes them to what they need instead of a bail bond -- but a deaf, hard-of-hearing, or communication-disabled defendant is disproportionately likely to miss exactly this kind of time-critical, high-stakes interaction in the first place. And Legal's mass AI defense service represents defendants at a scale no attorney could match one-on-one, but the hearing itself is still spoken, real-time, and unforgiving of a missed word. Wiring Accessibility's confidence-gated, real-time-speed communication and low-latency captioning directly into both -- the court-facing triage and the live hearing feed -- turns the accessibility layer from a generic add-on into the thing that makes the legal service actually reach the population most likely to fall through the gap it was built to close.

The same pairing generalizes to any legal, medical, or government process that's real-time, high-stakes, and currently assumes everyone can follow spoken language at full speed -- immigration hearings, medical consent conversations, DMV or benefits-office interactions.

Built from: Jingchen Jiang, Kendall Frederick, Sissi Lai, Sophia Mai (Accessibility and Assistive Technology); Owen Bartlett (Legal)

Applying Legal's Verified-Filing Model to Travel Compensation Claims and Overlooked-Party Representation combines 2 connections

Combines "Travel-Disruption Rights Autofiler" and "The Overlooked-Party Legibility Engine." Two separate connections between Legal and Travel point at the same underlying opportunity: giving a party too small or too rushed for individual attention the scale of representation an expert would normally provide. The moment Travel's rebooking layer detects a disruption -- a cancellation, a denied boarding, a carrier-caused delay -- that event is very often also a compensation-eligible legal claim almost no traveler actually files, because doing so requires legal knowledge they don't have and time they don't want to spend. Attach a verified-filing service, built the same way as Legal's malpractice-insurer certification, so the rebooking product doesn't just get someone rebooked -- it auto-files what they're owed.

The second connection is more structural: strip away the domain, and Legal's debt/eviction defense and Travel's small-destination readability layer are the same engine -- take a party too small, poor, or obscure for individual institutional attention, and give it a standardized way to become legible and actionable to an AI system at scale. That's a reusable pattern with an open third instantiation: a gig worker invisible to a benefits system, a rural clinic invisible to a hospital network -- built twice this class already, generic enough to build a third time.

Built from: Owen Bartlett (Legal); Sylvia, Steven, Flora, Yash (Travel)

Putting an Authorization Gate in Front of Police Database Searches and AI Agent Actions 1 connection, new in Week 3

Agency Under Surveillance found that every existing control on plate-reader searches is an audit log, and logs only report what already happened: one deputy searched one woman's plate 1,639 times before anyone noticed. The Cybersecurity team found the same failure at machine speed. Attackers now break out of a network in 29 minutes, far faster than any human can review an AI agent's actions after the fact. Both teams need the same component: an authorization layer that sits in front of the action and scales friction to risk. Routine requests pass, and sensitive ones need a stated case number, a second approver, or a warrant before they run. The Cybersecurity team also sketched this for agents as an intent-verification gateway, and the same design could sit in front of a police database.

The same gate is reusable anywhere a sensitive lookup or action is logged but not controlled: clinicians opening patient records they have no treatment reason to see, employees querying customer financial data, or AI agents calling payment and deletion tools. It is a close relative of the confidence gate proposed above for machine-scale AI publishing. A team could build the gate once and sell it to whichever of these buyers is under the most pressure to prove that misuse can't happen, not just that it was recorded.

Built from: Agency Under Surveillance's make-the-search-costly (Sophia Mai, Sissi Lai, Yash Chadda, Flora Huang, Daniel Sung); Cybersecurity's oversight at machine speed (Sahana Raja, Evan Batten, Kendall Frederick, Michael Squeri)

Giving People a Credit-Report-Style Right to See and Correct the Profiles Systems Infer About Them 1 connection, new in Week 3

Who Watches the Watchers wants people to be able to see, dispute, and delete the private risk scores built from their car, face, purchases, and chats, the way credit reports already allow. Week 2's Social Media team found that recommendation platforms infer what people want from their behavior and often get it wrong, with no way for the user to see or correct the profile. Both describe the same object: a profile built about a person, used to make decisions about them, that the person can't see. A single disclosure layer could do both jobs: show people the profile a system holds, let them dispute it, and let them declare what it should say instead.

Credit reporting is the one place this already works, and it took a federal statute to get there. A team could start where disclosure pressure is highest -- insurers that price from driving data, retailers running face-recognition watchlists, employers scoring productivity -- and build the see-and-dispute layer there first. The hard design question is the one the Watchers team raised: who plays the role of the credit bureau when the scores are spread across thousands of companies.

Built from: Who Watches the Watchers' right to your score (Owen Bartlett, Jonathan Locala, Jingchen Jiang, Sylvia Yam); Social Media's user-set feed goals (Jingchen Jiang, Sahana Raja, Sophia Mai, Sylvia Yam)

Funding Experience as Shared Infrastructure, the Way Medicine Funds Residency combines 2 connections, new in Week 4

Three Week 4 teams found that the way beginners become experienced is breaking, and each left the same question open: who pays for it. The Junior Jobs team explained why employers won't: training is an industry-level externality, since a firm that develops a junior may watch a competitor hire them away. The Construction team found the same in the trades, where a new tradesperson takes years to reach full productivity and supervision thins while they get there. The Higher Education team approached it from the school's side, proposing a degree that keeps skills current across a career instead of ending at 22. Medical residency, which the Junior Jobs team names as its model, addresses exactly this problem: training happens inside real work, under supervision, and much of its cost is carried by public funding rather than by the hospital that might lose the doctor afterward.

Other fields have built their own versions. Germany's dual apprenticeship system splits training between employers and vocational schools, and the UK funds apprenticeships through a levy that every large employer pays into. A team could design the funding and placement layer for one field (software, or a specific trade) where the externality is sharpest, and test whether employers, schools, or a pooled fund would carry the cost.

Built from: Junior Jobs' industry talent investment (Owen Bartlett, Jingchen Jiang, Jonathan Locala, Sophia Mai, Daniel Sung); Construction's experience leaving before it transfers (Sissi Lai, Yash Chadda, Flora Huang, Steven Li, Evan Batten); Higher Education's degree that stops at 22 (Michael Squeri, Sahana Raja, Kerry Wang)

Grading How the Answer Was Reached, for Students and for AI Agents 1 connection, new in Week 4

The Higher Education team found that grading capacity sets what gets assigned, and that schools responded to AI by returning to handwritten exams chosen because they are easy to grade, not because they measure what matters. Week 3's Cybersecurity team found a similar gap in AI agents: nothing detects when an agent reaches its goal by a route nobody intended. Both are asking for an evaluation that checks the process, not only the result. For a student working with AI, that means grading how they directed, checked, and corrected the tool. For an agent, it means checking whether it got there the intended way.

The Junior Jobs team's second opportunity points at the same skill from the employer's side: engineering judgment, the ability to evaluate and direct AI output, is becoming the differentiator. A team could build a process-level assessment that records how a person or an agent worked through a task with AI and grades the decisions along the way, starting in one large course where grading capacity is the binding constraint. Two Higher Education students (Sahana and Michael) were on the Week 3 Cybersecurity team, so this connection is already within one team's reach.

Built from: Higher Education's what's easy to grade gets taught (Michael Squeri, Sahana Raja, Kerry Wang); Cybersecurity's intent gap (Sahana Raja, Evan Batten, Kendall Frederick, Michael Squeri)

Letting People Set What Their Feed Is For, Tested Where the Law Already Requires a Choice 1 connection, new in Week 5

Week 2's Social Media team proposed letting people tell their recommendation algorithm what to optimize for instead of inferring everything from engagement. Week 5's Many-to-One team reached the same opportunity from gatekeeping theory: the editor who used to decide what reached people has been replaced by a feed ranked for engagement, and the only alternative most platforms offer is a plain chronological feed. The Week 5 team added two things the Week 2 version did not have. Firstly, it focused on teens and young adults, whose views are still forming and who were not studied in the 2020 Meta experiments. The other addition is a complication: those experiments found that the chronological feed showed more content from untrustworthy sources, so a choice of goal has to be more than a choice to turn ranking off.

The EU Digital Services Act already requires the largest platforms to offer at least one feed option not based on profiling, and some newer platforms let users pick among custom feeds. A team could start by looking at what people do when those choices are offered, and then design the goal-setting layer for one group, such as first-time voters. Jingchen was on both teams.

Built from: Communication's the feed decides for you (Flora Huang, Michael Squeri, Jingchen Jiang, Yash Chadda); Social Media's user-set feed goals (Jingchen Jiang, Sahana Raja, Sophia Mai, Sylvia Yam)

A Signed Record of What an AI Agent Agreed To, Kept Like the Accountability Ledger 1 connection, new in Week 5

The Machine-to-Machine team found that when a person's AI agent books or buys from another company's agent, no person reads the exchange, and the payment protocol records the payment but not the terms. Its well-defined problem is a plain-language receipt for every agent commitment, built from a signed log of the agent-to-agent messages, with an alert when a commitment goes past what the person approved. Week 2's Public Approval team proposed a tracker for who is accountable when an AI system causes harm, and Week 3's Cybersecurity team found that nothing detects when an agent reaches its goal by a route nobody intended. The Machine-to-Machine receipt is a specific case of both: an attribution record kept at the moment of the decision, and a check on whether the agent did what the person meant.

Consumer protection already has versions of this for human transactions, such as required disclosures before a purchase and a written record of the terms. A team could narrow the receipt to one kind of commitment where the terms matter more than the price, such as cancellation rules on travel bookings, and test whether the person, the merchant, or an insurer would pay for the record. The When No One Writes team named the same gap as its third opportunity, accountability for agent actions.

Built from: Communication's what did my agent agree to (Jonathan Locala, Sophia Mai, Penn Riney, Sylvia Yam); Public Approval's AI tracker (Sissi Lai, Kerry Wang, Penn Riney, Evan Batten); Cybersecurity's intent gap (Sahana Raja, Evan Batten, Kendall Frederick, Michael Squeri)

Routing Drivers to a Priced, Legal Curb Instead of an Address Pin 1 connection, new in Week 6

The Transport team found that navigation ends at the address pin, not at a legal place to stop, so drivers learn the curb rules on arrival, when they have the fewest options. In San Francisco's SFpark evaluation, demand-based curb pricing cut average parking search time by 43%. Week 2's Smart Cities team proposed the other half: an exchange for reserving, pricing, and governing curb space, rooftops, and drop points as instant delivery turns them into scarce urban assets. Combined, the curb is priced and reserved on one side and routed to on the other, so a driver or a delivery van arrives at a space it already knows is open and legal.

The Curb Data Specification, which Boston and Seattle are beginning to use, is the shared format both halves would need. A team could start in one city that already publishes curb rules in that format and build the routing-to-curb layer first, since it produces the demand data a pricing exchange would later rely on. Jonathan and Sichen were on both teams.

Built from: Transport's know the curb (Michael Squeri, Jingchen Jiang, Jonathan Locala, Sichen Deng); Smart Cities' logistics real-estate exchange (Yash Chadda, Sichen Deng, Kendall Frederick, Jonathan Locala)

Letting Travelers Declare What a Route Should Optimize, the Way Social Media's Team Proposed for Feeds combines 2 connections, new in Week 6

The Transport team found that routing apps rank by time and expose little else, while an 8% faster route was found to carry 23% higher crash probability, a trade-off the traveler never sees. Week 2's Social Media team found the same design in feeds, where the algorithm infers what you want from engagement, and proposed letting people state the goal instead, and Week 5's Many-to-One team returned to it (see the card above on letting people set what their feed is for). Week 3's Cybersecurity team described the failure all three are reacting to: a system that does what it was told to optimize rather than what was meant. The Week 5 team's complication carries over as well: removing the ranking is not the same as letting a person set its goal, and a route sorted only by distance is not automatically a safer one.

A team could build the declared-preference layer for routing: let a traveler set what a trip should weigh (risk, lighting, a child on board, a bike) and show the trade-off being made. The Transport team's own caution applies: define safety from road-level evidence such as crash history, lighting, and sidewalks, not from neighborhood crime data. Jingchen was on the Social Media, Many-to-One, and Transport teams, and Michael on the Cybersecurity, Many-to-One, and Transport teams.

Built from: Transport's routes that weigh more than time (Michael Squeri, Jingchen Jiang, Jonathan Locala, Sichen Deng); Social Media's user-set feed goals (Jingchen Jiang, Sahana Raja, Sophia Mai, Sylvia Yam); Communication's the feed decides for you (Flora Huang, Michael Squeri, Jingchen Jiang, Yash Chadda)

Agreeing Who Can Act Before the Alert Fires, for Food Shipments and for AI Agents combines 2 connections, new in Week 6

The Cold Chain team found that when a refrigerated load warms at 2 a.m., the driver has no authority to decide, the shipper's quality team is offline, and the dispatcher doesn't know the receiver's rejection rules, so the window to save the load closes before anyone with authority answers. Its third opportunity is to agree before the truck leaves who may re-route, re-ice, or reject. Week 5's Machine-to-Machine team proposed the same structure for AI agents: record what the person approved in advance, and alert when an agent goes past it. Week 3's Cybersecurity team found why it matters: AI security decisions happen faster than a person can read them, so authority has to be set before the event rather than requested during it.

Outside the class, pre-agreed authority already exists in a few places: standing orders let a nurse act on a defined set of conditions without calling the physician, and a trading desk's limits let a trader act up to a set size without approval. A team could write the standing order for one kind of shipment (a produce load with a defined temperature band, for example), test which decisions shippers, carriers, and receivers will agree to delegate, and see whether the same template fits an agent acting for a person.

Built from: Cold Chain's cross-boundary authority (Sophia Mai, Sahana Raja, Daniel Sung, Kerry Wang); Communication's what did my agent agree to (Jonathan Locala, Sophia Mai, Penn Riney, Sylvia Yam); Cybersecurity's oversight at machine speed (Sahana Raja, Evan Batten, Kendall Frederick, Michael Squeri)