Course Drift Oracle

Course Drift Oracle

A signed drift report for Microsoft's ai-agents-for-beginners course. It scans the course for two things that go wrong quietly:

• model pins that name a deployment which is deprecated or scheduled to retire
• package version floors whose open >= requirement now resolves to a newer major version than the course was written against

Both are invisible in a diff. Nothing in a course repo tells you a pinned model is about to stop serving, or that pip install today pulls a package two majors ahead of what the lessons assume. Answering that question takes someone actually checking each pin against the platform's own retirement calendar and the package's real release history. This repository does that checking, continuously, and sells the result as a signed report.

Why trust it

Every report is signed, not just asserted:

• Ed25519 signatures over RFC 8785 canonical JSON. The exact bytes that were signed can be reconstructed and re-hashed by anyone, in any language. A Python-signed receipt verifies against the TypeScript verifier with no shared trust between them - see oracle/receipt.py (signs) and worker/src/logic.ts (verifies).
• Append-only publication chain. Every report ever published is a link in oracle/chain.jsonl, each entry pointing at the hash of the one before it. Drop an entry, edit one, or reorder two and every later link stops matching. oracle/verify_chain.py checks this independently of the code that writes it.
• `verify_report` is free. Call it before paying to confirm the free summary's claims are backed by a real signature, and after paying to confirm the paid report matches the hash the free summary already committed to. A seller who serves fewer findings than promised gets caught by arithmetic, not by trust.
• This repository is the source. Nothing here is a description of the product; it is the product's actual scanner, signing code, and server, in the same form that runs the live service.

Use it in five minutes

The service is an MCP server over streamable HTTP:

https://signetworks.atelieri.workers.dev/mcp

Copy-paste quickstart, plain curl, no client library (every command here was run against production before being written down):

# 1. Initialize and capture the session id from the response headers.

SESSION_ID=$(curl -si -X POST https://signetworks.atelieri.workers.dev/mcp \ -H 'Content-Type: application/json' \ -H 'Accept: application/json, text/event-stream' \ -d '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2024-11-05","capabilities":{},"clientInfo":{"name":"quickstart","version":"1.0"}}}' \ | tr -d '\r' | awk 'tolower($1)=="mcp-session-id:"{print $2}')

2. Complete the handshake.

curl -s -X POST https://signetworks.atelieri.workers.dev/mcp \ -H 'Content-Type: application/json' \ -H 'Accept: application/json, text/event-stream' \ -H "Mcp-Session-Id: $SESSION_ID" \ -d '{"jsonrpc":"2.0","method":"notifications/initialized"}'

3. Call a free tool. Responses are SSE frames; strip the data: prefix.

curl -s -X POST https://signetworks.atelieri.workers.dev/mcp \ -H 'Content-Type: application/json' \ -H 'Accept: application/json, text/event-stream' \ -H "Mcp-Session-Id: $SESSION_ID" \ -d '{"jsonrpc":"2.0","id":2,"method":"tools/call","params":{"name":"drift_summary","arguments":{}}}' \ | grep '^data:' | sed 's/^data: //'

The same pattern calls every tool in the table above; only the name and arguments change.

Ten tools:

The twelve audit checks, so you know what a score means before paying: robots-exists, robots-ai-directives, robots-not-blanket, llmstxt-exists, llmstxt-shape, llmstxt-links, mcp-advert, structured-data, title-meta, charset, content-type-sanity, homepage-response. Ten are scoreable; homepage-response and robots-not-blanket are informational readings by construction and never count for or against a site, which is why a flawless site scores 10 of 10, not 12 of 12.

A typical run:

1. Call drift_summary. It costs nothing and returns a signed receipt describing how many findings exist, their severity, and which lessons are affected. Decide from that whether the full report is worth buying.
2. Call verify_report with the receipt from step 1. Confirm the signature is valid before you pay anything.
3. Pay $0.25 and call drift_report (or drift_report_x402 if you are paying with USDC). You get every finding with a file, a line number, what is wrong, and the fix.
4. Call verify_report again, this time passing the findings array you just received alongside the same receipt. Confirm the hash matches - this is what proves you got everything you paid for, not a trimmed version of it.

Use it as a GitHub Action

Add the scanner to a course repo's CI in one line:

- uses: 710git/course-drift-oracle@v1

That is the whole add - every input below has a default.

Inputs

Outputs

Why offline is the default

PyPI lookups are a live-network dependency, and a one-line CI add should not force one on a stranger by default. With offline: true (the default), the Action only checks model pins against the catalog bundled in this repo - no outbound calls beyond the checkout itself. Set offline: false to also catch package version drift against the live PyPI JSON API.

What it never touches

The Action never touches the Ed25519 signing key or the publication chain (oracle/chain.jsonl). It only reads the checkout it is given and the catalog bundled with the Action, and writes nothing back to the repo it runs in.

Verifying a receipt yourself

Every response carries a receipt object: a signed statement of what was found, when, and by which key, chained to every report published before it. verify_report does the check for you over MCP, but nothing about it is a black box - the two files that do the actual work are short and readable:

• oracle/receipt.py builds and signs a receipt: canonicalize the payload (RFC 8785), hash it (SHA-256), sign the hash (Ed25519).
• worker/src/logic.ts does the same steps in TypeScript to check a signature. Read both side by side and there is nothing left to take on faith.

oracle/verify_chain.py walks the full publication history in oracle/chain.jsonl and confirms every signature is valid and every entry links correctly to the one before it - the check a buyer or an outside auditor would run against the whole record, not just one receipt.

A free, signed skill

skills/course-drift-check/SKILL.md is the correct procedure for using this oracle, written for agents: audit the seller before paying, pay on either rail, verify the goods after. It is free, and it is signed - the receipt.json beside it commits to the file's exact SHA-256 under the same Ed25519 key that signs every report, so an agent loading the skill can prove it is running the procedure the operators published, unaltered (oracle/sign_skill.py is the 100-line signer; verification needs nothing from us). Signature proves provenance and integrity, not correctness - the same boundary every receipt here draws.

Payments

Two rails carry identical goods:

• MPP (Machine Payments Protocol) is designed to settle by card or stablecoin; this deployment currently charges on the Tempo testnet.
• x402 settles in USDC and exists mainly for discovery: agents that search for services through the x402 Bazaar only find ones that settle through an x402 facilitator.

Both rails currently run on testnets - MPP against the Tempo testnet, x402 against Base Sepolia. The x402 rail's payment mechanics are proven end to end: on 2026-08-28 a real testnet settlement cleared against the live worker (the buyer client in clients/ signed a $0.25 USDC payment, the facilitator settled it on Base Sepolia, and the client independently verified the findings hash it paid for - 7 passed, 0 failed, 0 skipped). No mainnet money has moved, and MPP settlement has not yet been exercised; both stated plainly rather than glossed over. You can reproduce the proof yourself: fund a throwaway wallet with faucet USDC (no ETH needed) and run the buyer per "Wiring a real x402 payer" in clients/buyer/README.md.

What this repository contains

oracle/     the scanner and the signing code

(to run it locally: clone the course repo it scans, then python3 oracle/drift_scan.py --repo-root /path/to/ai-agents-for-beginners with no course checkout, --offline runs the model-pin checks only) worker/ the MCP server: four free tools, six paid tools, HTTP 402 web/ the storefront pages served alongside the endpoint clients/ an independent buyer client - run it yourself before you pay

Project-Office, the private repository this mirror is drawn from, holds the day-to-day working process, deploy automation, and credentials. None of that is needed to use or verify this service, so none of it is here.

Honest limits

• Payments settle on testnets, not mainnet, as of this writing.
• The model and package catalog this scanner checks against is maintained by hand and covers a limited set of models today. Coverage expands every cycle; it is not exhaustive yet.
• Signature verification proves a report was produced by this service and has not been altered. It does not prove the report is correct - that claim rests on the sources cited inside each finding, which you are free to check yourself.

License

No license file is included. Absent one, all rights are reserved by default; this is a decision for the repository owner to make explicitly, not an oversight.

Contact

pm.agent.svc@gmail.com