Research LLM I/O capture for future distillation #12

Open

lost-rob0t wants to merge 2 commits from research/llm-io-corpus-distillation into main

lost-rob0t commented

2026-08-18 02:29:45 +00:00

(Migrated from github.com)

What changed

adds research/llm-io-corpus-for-distillation.org
indexes the new Org-roam research node
grounds the design in primary/OpenRouter/OpenTelemetry/OpenAI sources and distillation papers

Research conclusions

treat complete model boundary traffic as a corpus, not merely token telemetry
keep an encrypted append-only raw ledger separate from normalized analytics and compiled training datasets
preserve exact text/structured messages as canonical data; token IDs are optional derived/cache data
record tool calls/results and user/verifier outcomes as first-class causal events
distinguish black-box sequence-level distillation from logit-level KD, which requires distributions/providers to expose more than ordinary API I/O
record per-generation distillation eligibility/provenance; OpenRouter exposes is_trainable_text and provider.enforce_distillable_text
align normalized telemetry with OpenTelemetry GenAI semantic conventions where practical
use a Prolog policy layer to compile only provenance-complete, secret-clean, policy-eligible examples into SFT/preference/tool/eval datasets

Validation intent

The new node has a stable UUID-backed Org-roam ID, a title, source grounding, and an index.org link. No existing node IDs are changed.

Merge only when all required CI checks are green.

## What changed - adds `research/llm-io-corpus-for-distillation.org` - indexes the new Org-roam research node - grounds the design in primary/OpenRouter/OpenTelemetry/OpenAI sources and distillation papers ## Research conclusions - treat complete model boundary traffic as a corpus, not merely token telemetry - keep an encrypted append-only raw ledger separate from normalized analytics and compiled training datasets - preserve exact text/structured messages as canonical data; token IDs are optional derived/cache data - record tool calls/results and user/verifier outcomes as first-class causal events - distinguish black-box sequence-level distillation from logit-level KD, which requires distributions/providers to expose more than ordinary API I/O - record per-generation distillation eligibility/provenance; OpenRouter exposes `is_trainable_text` and `provider.enforce_distillable_text` - align normalized telemetry with OpenTelemetry GenAI semantic conventions where practical - use a Prolog policy layer to compile only provenance-complete, secret-clean, policy-eligible examples into SFT/preference/tool/eval datasets ## Validation intent The new node has a stable UUID-backed Org-roam ID, a title, source grounding, and an `index.org` link. No existing node IDs are changed. Merge only when all required CI checks are green.

This pull request can be merged automatically.

This branch is out-of-date with the base branch

You are not authorized to merge this pull request.

View command line instructions

Checkout

From your project repository, check out a new branch and test the changes.

git fetch -u origin research/llm-io-corpus-distillation:research/llm-io-corpus-distillation

git switch research/llm-io-corpus-distillation

Merge

Merge the changes and update on Forgejo.

Warning: The "Autodetect manual merge" setting is not enabled for this repository, you will have to mark this pull request as manually merged afterwards.

git switch main

git merge --no-ff research/llm-io-corpus-distillation

git switch research/llm-io-corpus-distillation

git rebase main

git switch main

git merge --ff-only research/llm-io-corpus-distillation

git switch research/llm-io-corpus-distillation

git rebase main

git switch main

git merge --no-ff research/llm-io-corpus-distillation

git switch main

git merge --squash research/llm-io-corpus-distillation