Trigger checks on learner PRs and submit without merging

.github/workflows/grade.yml

@@ -1,12 +1,13 @@
 name: Check intent submission
 on:
-  push:
+  pull_request:
+    types: [opened, synchronize, reopened]
     branches: [main]
 permissions:
   contents: read
   pull-requests: read
 jobs:
   grade:
-    uses: TensorOpsAI/tensorlearn-intent-grader/.github/workflows/grade.yml@9e40d82165c88f9028c6b119f5833dac5d48da9b
+    uses: TensorOpsAI/tensorlearn-intent-grader/.github/workflows/grade.yml@46de78b3462c3d3a18c2f55e81d645a816815302
     with:
-      grader-sha: 9e40d82165c88f9028c6b119f5833dac5d48da9b
+      grader-sha: 46de78b3462c3d3a18c2f55e81d645a816815302

CLAUDE.md

@@ -6,4 +6,7 @@ Read @intent.md for the student's proposed implementation instructions.
 This round is about writing a useful intent document, not implementing the feature.
 Change only intent.md. Help the student reason about scope, constraints and testable
 acceptance criteria. Do not silently commit, push, open or merge a PR. The student
-must explicitly choose to merge their solution PR into their own fork's main branch.
+opens a PR inside their own fork, targeting that fork's main branch. Opening or
+updating the PR starts advisory checks. Only the student's explicit /submit request
+submits the selected pushed revision to TensorLearn. Keep the PR open; no merge is
+required. Never present an Actions check as a saved official score.

README.md

@@ -8,16 +8,30 @@ Read [the product brief and published rubric](docs/product-brief.md).
 
 1. Fork this repository to your personal GitHub account. Use fictional information
    only, keep the fork public, and keep `main` as its default branch.
-2. Clone your fork, open it in Claude Code, and enable GitHub Actions in the fork.
-3. Create a branch such as `intent/first-draft`. Read the brief and ask Claude to
+2. Clone your fork and enable GitHub Actions in it. Your facilitator enrolls you in
+   the lab, approves your exact fork, and confirms the GitHub App connection.
+3. Download the compatible TensorLearn client (v0.5.1 or later) from the lab page
+   and extract it outside your fork. Follow the page's `--code` connection command
+   from your fork folder. Paste your learner token only into the hidden terminal
+   prompt. After the facilitator launches the round, use `/load` in Claude Code.
+4. Create a branch such as `intent/first-draft`. Read the brief and ask Claude to
    help you write `intent.md`. Change only that file. Review the document yourself.
-4. Start a fresh Claude session. Ask it to explain the scope and acceptance criteria
-   it loaded before doing any implementation. Refine your document as needed.
-5. Commit and push your branch. Open a PR **inside your own fork**, targeting its
+5. Start a fresh Claude session if needed and ask it to explain the loaded scope
+   and acceptance criteria before doing any implementation. Refine your document.
+6. Commit and push your branch. Open a PR **inside your own fork**, targeting its
    `main` branch. GitHub may suggest the upstream repository: select your fork instead.
-6. Review and merge your PR. **The merge is your submission for this exercise.**
-7. Open the Actions run for the merged commit. Download `intent-score` and inspect
-   `score.json`. Discuss the feedback with your facilitator and repeat with a new PR.
+   Opening or updating this PR starts the document checks automatically.
+7. Keep the PR open. When ready, run **`/submit`** in the connected Claude Code session
+   to submit that exact pushed revision for the current round. The client can also
+   create the same-repository PR if it does not exist yet. **No merge is required.**
+8. Use `/check` or refresh the lab page to read the saved result. The Actions run
+   and `intent-score/score.json` are useful feedback, but an Actions check alone
+   does not submit your work or prove that the server saved a score.
+
+A change to your answer requires a new commit, push, and explicit `/submit`.
+Keep the PR open until your facilitator finishes reviewing the exercise. Your
+facilitator controls the official attempt allowance; advisory CI runs do not count
+as official submissions or cap GitHub compute.
 
 ## What is verified
 
@@ -32,17 +46,16 @@ Claude imports are described in [the official memory documentation](https://code
 
 The separate [document checker](https://github.com/TensorOpsAI/tensorlearn-intent-grader)
 is pinned to an immutable commit in `.github/workflows/grade.yml`. No server secrets
-belong in this starter or student forks. The workflow emits the platform's score.json
-format; GitHub App delivery and admission of the submission must also be configured
-on the platform before a score can appear there.
-
-**Rehearsal blocker:** the currently observed platform requires a private repository
-and explicit submission of an open PR. It does not currently accept this exercise's
-public-fork merge as an official submission. The exercise can be prepared and its
-checker tested, but the website score step remains unverified and blocked by that
-mismatch. Do not treat an Actions score as proof that the server saved it.
-
-The exercise should remain unlaunched until the facilitator has resolved that setup
-and checked one actual learner result end to end. See the checker repository's
-rehearsal checklist. No customer identity, credentials, or participant data is stored
-in these public repositories.
+belong in this starter or student forks. The workflow grades the PR head commit,
+not GitHub's synthetic merge preview. Both PR branches must belong to the learner's
+approved repository; do not open a PR against the upstream starter.
+
+Before launch, verify that the deployed platform accepts approved public repositories
+and pull-request check results, the matching client is downloadable, and the learner
+repository's GitHub App installation is approved. Configure the grading allowance.
+Observe an actual saved learner score before claiming end-to-end readiness. A client
+or server that only supports private repositories is incompatible with this public
+fork exercise. See the checker repository's rehearsal checklist.
+
+No customer identity, credentials, or participant data is stored in these public
+repositories. All product examples are fictional.