# Thesis stress test

Date: 2026-07-20

## Verdict

Aeonora has enough evidence for a technical proof of concept. It does not yet have evidence for broad market demand.

The repository proves that a Markdown agent brain can be encrypted, restored exactly, migrated across named runtime profiles without changing agent identity, and protected by rotating runtime keys. It does not prove that teams experience this problem often enough to pay, that on-chain anchoring is required for the first buyer, or that the current storage topology is permanent.

## Strongest version of the thesis

Long-lived agents should not be owned by the runtime that happens to execute them. Their portable identity, owner authority, encrypted state, and lineage should survive model, harness, cloud, and device changes.

## Falsifiable assumptions

| Assumption | Evidence now | Next test | Failure signal |
|---|---|---|---|
| Operators lose useful agent state when changing runtimes | Anecdotal and inferred | Interview 10 operators with agents older than 30 days | Fewer than 3 report a painful migration or lock-in event |
| Markdown captures enough portable state for a useful wedge | Local multi-file roundtrip passes | Move 3 real agents between two unrelated harnesses | Important behavior depends on runtime-private state that cannot be reconstructed |
| Neutral identity matters independently of execution | Identity survives deterministic test migrations | Ask design partners to recover and verify the same agent on a new device | Users only want backup/export, not durable identity |
| On-chain proof improves the product | Contract and optional Base Sepolia flow work | Compare migration demo with and without public proof | Buyers cannot name a counterparty who needs independent verification |
| Managed continuity is monetizable | No revenue evidence | Ask for 3 paid design-partner commitments | Users only accept free self-hosting or existing storage tools |

## Adversarial alternatives

1. **Git is enough.** For technical users, a private repository plus secrets management may solve most portability needs. Aeonora wins only if recovery, encryption, cross-runtime normalization, policy, and public lineage reduce meaningful work.
2. **Harnesses standardize first.** If major runtimes adopt a common complete export format, Aeonora becomes a thin compatibility layer. The protocol should embrace that format rather than compete with it.
3. **On-chain identity is premature.** The first useful product may be encrypted continuity with optional proof. Chain anchoring should remain optional until a real verifier needs it.
4. **State is not identity.** Copying files may reproduce context without reproducing tool permissions, credentials, scheduled work, embeddings, or behavior. Aeonora must define explicit adapter boundaries and never imply perfect behavioral equivalence.
5. **Recovery creates custody risk.** A seed phrase is portable but unforgiving. Passkeys, multiple devices, revocation, and documented recovery drills are required before sensitive use.

## Go/no-go gates

Continue building the hosted product only after:

- three external agents migrate between unrelated harnesses;
- exact file restoration and stable identity pass every time;
- at least three operators ask to keep using it;
- at least one operator pays for managed continuity;
- one buyer explains why public lineage is valuable without prompting.

Do not expand into a marketplace, token, generalized runtime, or decentralized compute network before those gates.

## Automated evidence

Run `npm run eval`. `evals/thesis-eval.mjs` tests multi-file capture, encrypted-at-rest plaintext absence, exact nested restoration, three sequential runtime migrations, identity continuity, old-key rejection, signed lineage, symlink rejection, and the no-Markdown boundary. Its successful verdict is deliberately `TECHNICAL_POC_SUPPORTED_MARKET_UNVALIDATED`.