Integrate #7: the local model evidence, and where it came from - #21
Merged
Merged
Conversation
…model evidence The README and docs/local-llms.md said one contributor had run a local model and no transcript was in the repository. Two are now, under docs/assets/transcripts/: seed 6, where the previous run's episode sits in the system prompt under "What you remember" and the model never calls `remember`, and seed 5 after `remember` became strategy step 5, where it calls it once with a fact from the verb results. Both successes, every turn a native tool call, no text fallback, 14 s and 11 s of wall clock. Only the contributor's home directory was removed from `duck_path`. docs/local-llms.md reads them in a table and says what they cannot show; docs/assets/README.md records how they were made; the README's two sentences point at the folder; CHANGELOG under Unreleased. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The local-llms table called 14 s and 11 s "wall clock"; that is the simulator clock the budget counts on sim2d. The transcript timestamps give 33 s and 29 s, three to nine seconds per LLM call, and the table now says both. The seed 5 row read as if `remember` came alone and the bundle followed; the model returned all three tools in one response and the loop kept the first. The CHANGELOG entry no longer sits between the Unreleased intro and the audit it introduces. assets/README says why `memory.path` is relative, and memory.md points its "read but never wrote" claim at the two files that show it. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GrWzSyBBfB33MFGYR8qUf9
…the transcripts change main rewrote the flock bullet and the contributors row while this branch was open; both are taken verbatim. The two lines that said no transcript is in the repository are the ones this branch exists to change, so they are updated in the shorter voice main now uses.
…ame from Bayway's PR #7, merged whole. His commits land as written; the corrections are in the commits after this one. The README conflict was resolved to MAIN's side, not to his, which is the opposite of how #13 was handled and deliberate. His branch is from ea271b0 and the conflicting hunks are nine table rows and four Limitations bullets he never edited: taking his side would have put back "seven adapters" as eight (the CLI lists seven), four providers as eleven cloud vendors, eight robot_* tools as nine, and deleted main's registry, pilot-flock and coordinator-flock rows along with the model-catalogue bullet. Resolving to his side is right when his side is his work. Here it was a stale copy of somebody else's, so main's text stands and his two line changes were applied onto it by hand, unaltered. Those two lines are still exactly his words, including "one of them reading the previous run's memory". That sentence is wrong, and the next commit says why and fixes it there rather than editing it here. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
…sample Four sentences said seed 6 read what seed 5 wrote. It did not. The two transcripts ran against different memory directories: seed 5's run_start says `memory-qwen3` with `episodes: 0` and a prompt reading "(nothing yet, this is the first run on this robot)", seed 6's says `memory-qwen2` with `episodes: 1`. The episode seed 6 remembers reads "Ball moved 0.38 m after kick, meeting the success criteria." and the published seed 5 declares "the ball was kicked successfully, moving 0.38 m to the right with the right leg". Different runs. So seed 5 wrote a fact nothing here reads, seed 6 read a fact nothing here wrote, and the pair are the two ends of a chain whose middle is not in this repository. "the previous run's episode" is now "an earlier run's episode" in local-llms.md, assets/README.md and the README, and local-llms.md says outright that the two are not a chain, so nobody tries to verify one against the other and concludes the docs are lying. Also: "There are exactly two data points" was a selection described as a sample. docs/design/memory.md records the same model reading the memory block and never writing to it across four runs before `remember` moved into the strategy, so there were more than two runs and two were kept. Said so, and added the missing line to "what they cannot show": neither transcript shows a note surviving from the run that wrote it into the run that reads it, which is the one thing the pair looks like it demonstrates. One more, caught by the repo rather than by reading: the sentence added to the README carried a semicolon, and test_readme_punctuation_style forbids semicolons and dashes in README prose. Now a full stop. The test was already there and it works. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The new entry auto-merged with no conflict into line 1538, which is inside `## [0.7.0] - 2026-09-07`, four lines above `## [0.6.0]`. Nothing would have flagged it. Merging it there would have claimed 0.7.0 shipped two transcripts it did not ship, in the one file where a retroactive edit does the most damage. Moved to Unreleased, folded into the `### Added` section that was already there rather than carrying its own heading. While it was being moved it also gained the credit in the house form, and the caveat the entry needs to be true: the two runs are not a chain. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
`grep -rn "no transcript" --include=*.md` on main returned five live copies of the same sentence. The PR rewrote three of them (README twice, local-llms.md) and correctly left two alone, because CHANGELOG.md and docs/design/memory.md are history and history is not edited to match the present. PLAN.md was the fifth, and it is neither. Its own header says "What is still open. Everything that has shipped is in CHANGELOG.md, the ADRs and the git history", so the carve-out that protects the other two does not reach it. On merge both halves of its sentence became false at once: the transcript is in the repository now, and the README it cited as saying otherwise no longer says it. Rewritten rather than deleted, because the item is not finished. "None on the dev machine" is still true, LM Studio is now covered and the two other servers are not, and the pair does not demonstrate a note surviving from one run into the next. That last part is the honest remainder and it is what the item now asks for. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
10 tasks
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Integration branch for #7 (@Bayway), handled per CONTRIBUTING: merged with
git merge, his commits verbatim, corrections on top in named commits.What checks out
Every number in the new
local-llms.mdtable matchesrun_end: seed 6 is 8 steps / 9 calls / 29,403 + 244, seed 5 is 4 steps / 6 calls / 17,939 + 265. All 15 LLM rows carrystop_reason: tool_calls, so "no text fallback" is true. "Two kicks fall short before the third connects" is true: 0.15, 0.18, 0.44 m.quackd tracerenders both files. Privacy scan is clean, no home directories, no/Users/, no IPs, hostnames, email or machine name. Hisduck_pathscrub held.His docs are in one place more careful than his own PR description, which said "14 s and 11 s of wall clock" where
local-llms.mdcorrectly separates sim clock from wall clock.The merge resolved README to main's side, not his
The opposite of how #13 was handled, and deliberate. His branch is from
ea271b0and the conflicting hunks are nine table rows and four Limitations bullets he never edited. Taking his side would have put back "seven adapters" as eight (quackd list-adapterslists seven), eleven cloud vendors as four, ninerobot_*tools as eight, and deleted main's registry, pilot-flock and coordinator-flock rows along with the model-catalogue bullet. Resolving to the contributor's side is right when his side is his work. Here it was a stale copy of somebody else's, so main's text stands and his two line changes were hand-applied onto it, unaltered.Corrections on top
6d0acb8the two runs are not a chain. Four sentences said seed 6 read what seed 5 wrote. It did not. Seed 5'srun_startsaysmemory-qwen3withepisodes: 0, seed 6's saysmemory-qwen2withepisodes: 1. The episode seed 6 remembers reads "Ball moved 0.38 m after kick, meeting the success criteria." and the published seed 5 declares "the ball was kicked successfully, moving 0.38 m to the right with the right leg". Different runs. Seed 5 shows the write, seed 6 shows the read, neither shows the other's half.Also "There are exactly two data points" was a selection described as a sample:
docs/design/memory.mdrecords four runs beforeremembermoved into the strategy.af272cbthe CHANGELOG entry was landing inside 0.7.0. It auto-merged with no conflict at line 1538, four lines above## [0.6.0]. Nothing would have flagged it. Moved to Unreleased, credit added in the house form.8fa6ee3PLAN.md was the fifth place that said no transcript. He rewrote three and correctly left CHANGELOG anddocs/design/memory.mdalone as history. PLAN.md is neither: its header says "What is still open". Rewritten rather than deleted, because the item is not finished.Caught by the repo, not by reading
The sentence added to the README carried a semicolon, and
test_readme_punctuation_styleforbids semicolons and dashes in README prose. Fourteen review agents read that line and none flagged it. The test failed in half a second.Gate, on the merged result
ruff check,ruff format --checkagainstgit archivebytes,mypyon 3.11 and 3.12, and the full suite. All green locally. This PR exists to get macOS, Linux and the physics job too.🤖 Generated with Claude Code