TPI-006 eval-first monkey model provider metadata pass

notes/pr5_update_summary.md

@@ -0,0 +1,27 @@
+# PR #5 Update Summary
+
+## What this PR adds
+
+- a provider-metadata / endpoint-discovery plan for the selected Runpod path
+- a concrete endpoint contract template
+- a provider checklist and metadata decision surface
+- a sharper blocked-state record showing which local discovery sources were exhausted
+
+## What still remains
+
+- actual pod identifier or display name from the provider
+- actual host / endpoint from the provider
+- actual attach or SSH command from the provider
+- explicit username / port if required
+- runtime confirmation that the attach route lands in `/workspace/parameter-golf`
+
+## First required provider handoff
+
+The next turn should begin from the exact provider-supplied attach command or SSH tuple that identifies the selected pod explicitly.
+
+## Evidence added in this turn
+
+- PR #5 has no review comments or embedded provider metadata
+- no Runpod endpoint or attach command was found in shell history or Windows PowerShell history
+- no Runpod URL history was found in Chrome, Edge, Brave, or Firefox
+- local SSH material exists, but provider-specific endpoint data does not

notes/tpi_006_endpoint_contract.md

@@ -0,0 +1,82 @@
+# TPI-006 Endpoint Contract
+
+## Objective
+
+Record the concrete provider-supplied endpoint data needed to attach to the selected Runpod pod.
+
+## Required fields
+
+- pod identifier or pod display name: not yet obtained
+- host or endpoint: not yet obtained
+- exact attach command or SSH command: not yet obtained
+- username: not yet obtained
+- port (if required): not yet obtained
+- expected landing path: `/workspace/parameter-golf`
+- first verification commands after attach: fixed below
+
+## Concrete facts available now
+
+- local SSH client exists at `/usr/bin/ssh`
+- reusable SSH keys exist at:
+  - `/mnt/c/Users/eb245/.ssh/id_ed25519`
+  - `/mnt/c/Users/eb245/.ssh/id_rsa`
+- no Windows-side SSH config file was present
+- Windows-side `known_hosts` contained GitHub hosts only
+- no Runpod-specific endpoint, hostname, or attach command was found in:
+  - PR #5 body or reviews
+  - shell history
+  - Windows PowerShell history
+  - browser history for Chrome, Edge, Brave, or Firefox
+
+## Provider-adjacent public reference
+
+The public repo still points to the official launch template:
+
+```text
+https://console.runpod.io/deploy?template=y5cejece4j&ref=nl2r56th
+```
+
+This is not sufficient to resume TPI-004, because it does not identify the concrete pod, endpoint, username, or port for the already selected instance.
+
+## Target landing path
+
+Preferred:
+
+```bash
+/workspace/parameter-golf
+```
+
+If the attach route lands elsewhere, record the exact correction steps needed to reach the repo.
+
+## Exact attach command
+
+Still unavailable from provider metadata in this workspace.
+
+Current reusable template once provider metadata is handed off:
+
+```bash
+ssh -i /mnt/c/Users/eb245/.ssh/id_ed25519 <user>@<host> -p <port>
+```
+
+## First verification commands
+
+```bash
+pwd
+ls /workspace
+cd /workspace/parameter-golf
+git rev-parse --abbrev-ref HEAD
+python3 -c "import torch, datasets, sentencepiece; print('deps-ok')"
+nvidia-smi
+```
+
+## Resume condition
+
+Once all required fields are filled with real values, the branch is ready to resume the unchanged TPI-004 evidence pass.
+
+## Attach failure fallback
+
+- require one concrete provider handoff item set:
+  - exact SSH command, or
+  - exact host + username + port tuple, or
+  - provider console attach route tied to the selected pod id
+- until then, endpoint discovery remains blocked rather than partial

notes/tpi_006_metadata_decision.md

@@ -0,0 +1,89 @@
+# TPI-006 Metadata Decision
+
+## Status
+
+blocked
+
+## Objective
+
+Record whether concrete provider-supplied endpoint metadata has been obtained for the selected Runpod path.
+
+## Required decision fields
+
+- pod identifier known or not: not known
+- host / endpoint known or not: not known
+- exact attach / SSH command known or not: not known
+- username known or not: not known
+- port known or not: not known
+- landing path confirmed or not: expected path known, but not runtime-confirmed
+
+## Classification
+
+- `confirmed`
+- `partial`
+- `blocked`
+
+## Concrete metadata obtained
+
+- no pod-specific provider metadata was obtained in this turn
+- expected landing path remains `/workspace/parameter-golf`
+- local auth material remains reusable:
+  - `/mnt/c/Users/eb245/.ssh/id_ed25519`
+  - `/mnt/c/Users/eb245/.ssh/id_rsa`
+- public launch template reference remains available in the repo README
+
+## Discovery sources checked
+
+- PR #5 body and review surface
+- shell history
+- Windows PowerShell history
+- Windows-side SSH directory, config, and known hosts
+- Chrome history (`Default`, `Profile 2`)
+- Edge history (`Default`)
+- Brave history (`Default`, `Profile 1`)
+- Firefox history (`default-release`)
+
+## Still missing
+
+- pod identifier or display name
+- host / endpoint
+- exact attach command or SSH command
+- username
+- port if non-default
+
+## Classification result
+
+- `blocked`
+
+## Interpretation
+
+- The blocker has narrowed to provider-supplied pod metadata that is absent from the current workspace.
+- Further local discovery would likely repeat the same negative checks rather than create a runnable attach route.
+- TPI-004 still cannot resume from this workspace alone.
+
+## Can `/workspace/parameter-golf` be reached now?
+
+- No
+
+## Can TPI-004 resume now?
+
+- No
+
+## First command once provider metadata is supplied
+
+```bash
+ssh -i /mnt/c/Users/eb245/.ssh/id_ed25519 <user>@<host> -p <port>
+```
+
+Then:
+
+```bash
+pwd
+ls /workspace
+cd /workspace/parameter-golf
+git rev-parse --abbrev-ref HEAD
+```
+
+## Resume condition
+
+TPI-004 can resume unchanged only once the classification is at least `confirmed` for the concrete attach route.

notes/tpi_006_provider_checklist.md

@@ -0,0 +1,33 @@
+# TPI-006 Provider Checklist
+
+## Goal
+
+Obtain the minimum provider-supplied metadata needed to attach to the selected Runpod pod and resume TPI-004 unchanged.
+
+## Required items
+
+- [ ] pod identifier or display name
+- [ ] host / endpoint
+- [ ] exact attach command or SSH command
+- [ ] username
+- [ ] port (if needed)
+- [x] expected landing path
+- [x] first verification commands confirmed
+
+## Unfilled items and why
+
+- pod identifier or display name: not present in PR #5, local history, or browser history
+- host / endpoint: not present in PR #5, local history, SSH state, or browser history
+- exact attach command or SSH command: no saved provider command was found
+- username: depends on the missing provider command or endpoint tuple
+- port: depends on the missing provider command or endpoint tuple
+
+## Acceptance rule
+
+TPI-006 is successful only if a future turn can begin from concrete provider metadata rather than guessing endpoint details.
+
+## Resume target
+
+After endpoint discovery is concrete enough, the next turn should resume TPI-004 with:
+- baseline `EVAL_STRIDE=1024`
+- candidate `EVAL_STRIDE=128`

notes/tpi_006_provider_metadata_plan.md

@@ -0,0 +1,46 @@
+# TPI-006 Provider Metadata Plan
+
+## Objective
+
+Obtain the concrete provider-supplied metadata needed to attach to the selected Runpod pod and land in `/workspace/parameter-golf`.
+
+## Public-facing name
+
+`MonkeyModel_EvalFirst_EndpointDiscovery`
+
+## Required provider metadata
+
+- pod identifier or display name
+- host / endpoint
+- exact provider-supplied attach command or SSH command
+- username
+- port (if needed)
+- expected landing path
+
+## Why this loop exists
+
+TPI-005 established that local auth material exists, but pod-specific metadata is still missing. This loop narrows the blocker from generic attachability to concrete provider data.
+
+## Success condition
+
+The provider metadata is concrete enough that a future turn can begin by executing the exact attach route instead of inferring it.
+
+## Discovery sources for this turn
+
+- PR #5 body and review surface
+- local shell history
+- Windows PowerShell history
+- Windows-side SSH directory and known-hosts state
+- browser history for Chrome, Edge, Brave, and Firefox
+- public Parameter Golf README guidance for the Runpod template
+
+## Stop rule
+
+If pod-specific values are still absent after checking the sources above, classify the branch as blocked on external provider handoff rather than spinning on more local discovery.
+
+## Non-goals
+
+- no model changes
+- no tokenizer changes
+- no environment reselection
+- no score claims

runs/TPI-004/run_notes.md

@@ -92,3 +92,21 @@ torchrun --standalone --nproc_per_node=1 train_gpt.py
 ## Status
 
 blocked before pod attachment
+
+## TPI-006 provider metadata discovery update
+
+- branch checked for discovery handoff: `exp/eval-first-006`
+- current local commit during discovery: `0b981990cc6d2d21e2e49e8bb71ed1a70691342f`
+- provider-specific metadata checked in:
+  - PR #5 body and review surface
+  - shell history
+  - Windows PowerShell history
+  - Windows-side SSH directory and known-hosts state
+  - browser history for Chrome, Edge, Brave, and Firefox
+- result:
+  - no pod identifier found
+  - no Runpod host or endpoint found
+  - no exact attach command found
+  - no username or port found
+- implication:
+  - TPI-004 remains blocked on external provider handoff, not on the eval-first monkey-model branch itself