Ideas - verification and trust#
Contents
Game 5 and the compressed version: generation is free; verification is scarce. These companies sell the scarce side.
1. Provenance at the sensor#
Why now. Deepfakes and synthetic media made post-hoc "is this real?" inference a losing arms race. Attestation has to happen at capture, not in moderation queues after the flood. → C8, media
Build this. Hardware + firmware + cloud that cryptographically signs photo/video/audio at the device, with a verifier API banks, newsrooms, courts, and insurers can call. Start in one high-stakes vertical (insurance claims, remote notarization, clinical imaging chain of custody).
Own this. Device trust roots, verifier default placement, and the audit trail data - not a "detector model."
Not this. Another AI deepfake classifier trained on last month's generators.
Dead if. Platforms ship free, universal capture attestation that commodity hardware adopts - you become a checkbox, not a company.
2. Proving a human is present#
Why now. Voice and video clones turn "I saw them / heard them" into a non-signal. Fraud, hiring, dating, capital markets, and customer support all need a live human-presence layer that is not a CAPTCHA. YC's Fall 2026 RFS names the same problem; the corpus frames it as costly signaling collapse (Game 5).
Build this. A privacy-preserving presence protocol: liveness + optional institutional identity, usable inside Zoom/phone/bank flows, with liability terms for false accepts.
Own this. The protocol default, the certified device or biometric path, and the indemnity relationship with relying parties.
Not this. A chatbot that "guesses" bot probability from text.
Dead if. One OS vendor or payments network makes presence a free platform primitive and locks the API.
3. Review as a product (software)#
Why now. Generation throughput made senior review the schedule bottleneck (software). Teams that buy more coding agents without buying review capacity just lengthen the PR queue.
Build this. Tools and services that shrink time-to-trusted-merge: behavioral eval harnesses on the customer's suite, risk-ranked review, incident-linked learning, optional insured review for high-severity paths. Sell to orgs drowning in AI PRs and to open-source maintainers paying the bogus-report tax.
Own this. Customer eval corpora, production outcome labels, and (if you underwrite) the risk book - not "LGTM from a model."
Not this. Autocomplete that files more PRs.
Dead if. End-to-end agents merge to prod with measured incident rates at human or better without external review scaffolding - then review was a transition rent (U5).
4. Eval infrastructure with skin in the game#
Why now. Benchmarks are marketing; buyers need continuous, task-specific, adversarial evals tied to contracts (B5 outcome pricing).
Build this. An eval cloud where enterprises define workflows, run scheduled red-teams, and gate model/version promotion - with optional warranty products that pay when evals pass but production fails.
Own this. Proprietary task suites, longitudinal failure data, and warranty capital if you go there.
Not this. A leaderboard website.
Dead if. Model vendors give away credible, independent, per-customer eval that buyers trust as much as third parties - rare, but watch enterprise contracts.
5. Accountability routing for agents#
Why now. Agent chains blur causation across user, deployer, model vendor, and tool provider (insurance). Claims and governance need a legible "who signed."
Build this. Runtime that forces designated human or corporate principals at high-severity actions, logs decision provenance for insurers and courts, and exports the artifact underwriters already ask for.
Own this. The enterprise default for agent governance, the log format regulators recognize, and partnerships with carriers.
Not this. A policy PDF generator.
Dead if. One hyperscaler's agent platform becomes the de facto auditable runtime and refuses interop.
6. Institutional vouching markets#
Why now. Asynchronous credentials (degrees, follower counts, polished applications) are cheap to fake. High-stakes matching needs reputation that costs something to stake (Game 5, education).
Build this. Markets or networks where professionals and firms stake bond-like reputation on claims (this candidate can do X; this vendor hit SLA Y), with slashing or insurance when wrong. Start in hiring, vendor procurement, or expert networks.
Own this. The stake ledger, dispute resolution, and the density of stakers in a vertical.
Not this. Another endorsement button with no downside.
Dead if. Synchronous evaluation (live work trials) fully replaces reputation intermediaries in your vertical.
7. Self-maintaining integration tissue#
Why now. APIs churn; agentic coding can open PRs; the missing layer is provider-to-customer change application (YC RFS: self-maintaining APIs). This is verification of compatibility, not of truth.
Build this. Agents with scoped repo access that watch vendor changelogs, propose and test upgrades, and only merge when customer eval suites pass - Dependabot for semantic API and agent-tool breakages.
Own this. Cross-customer break graphs and vendor-specific repair playbooks.
Not this. A prettier changelog email.
Dead if. Every major API ships formal contracts + automated customer migration as a free vendor feature.
8. Multiplayer agent sessions#
Why now. Long-running agents are team work stuck in single-player chat (YC RFS: multiplayer AI). Shared state, handoff, and audit beat a thousand private threads.
Build this. Real-time shared agent workspaces for eng, support, sales, or legal - presence, permissions, replay, and compliance export.
Own this. Team graph + workflow data inside the session substrate; become the default place agent work happens.
Not this. Transcript sharing links.
Dead if. The dominant productivity suite ships good enough multiplayer agents as a free tier.
Related: Compressed · Game 5 · Software · Insurance · C8