Skip to content

Fold peterpme/lev (educational Kev pointer head) - #76

Merged
24601 merged 2 commits into
mainfrom
cursor/fold-peterpme-lev-4eec
Sep 21, 2026
Merged

24601 merged 2 commits into
mainfrom
cursor/fold-peterpme-lev-4eec

Conversation

@24601

@24601 24601 commented Sep 21, 2026 •

Copy link
Copy Markdown
Owner

Summary

Catalogs peterpme/lev as a novel HIGH peer on the open Kev line (notes.md §149). Educational Qwen2.5-0.5B + LoRA + pointer head. Caller-supplied options become softmax probabilities. FastAPI POST /v1/systemone is choice only; noul and score are later. Not production. License file absent at HEAD 9b379262e45f4bc2ebecccc2c7927800c61bdc99.

GitHub release Lev 1.0.0 was published 2026-09-21T17:30:54Z. Tag 1.0.0 points at 49d23ff319dacc266e44b55ed1f6c74fb18c34b9. HEAD is one commit ahead of that tag. Fingerprint release_tag is 1.0.0. HEAD, README SHA, and baseline SHA stay on 9b379262.

Author banking77-v1 baseline is accuracy 0.88 on 150 records, with NLL, Brier, ECE, and latency from baselines/lev-banking77-1500.json (theirs). GOAL.md's 88.67% on 150 held-out rows is a separate author sentence and is not averaged with 0.88.

Soft probabilities are for rank or threshold. They are not a safety gate. Lev is not hosted TypeSafe Jev, not a replacement for jaredpalmer/kev (§45), not trained Laya, not an OpenJev scalar head, and not yoheinakajima/glance (vision logit harness, §147). No Augustus runtime call site.

Uniqueness lock lives in research/notes.md only. The human README still ends at ## License. Lev stays §149. This PR does not absorb hourly 1256.

Left as a draft for ADV. Do not merge.

Files

  • research/notes.md §149 dated HIGH card and uniqueness lock
  • research/sources.json primary repo
  • research/revisit_fingerprints.json last-look fingerprints
  • research/refresh-log.md short pointer
  • docs/ecosystem.md peer index line
  • .agents/skills/augustus/references/judgment-class.md class placement (not the lock)

Test

python3 .agents/skills/augustus/scripts/uniqueness_gate.py passed (uniqueness-gate ok).
python3 research/revisit_fingerprints.py --self-test passed.

Open in Web Open in Cursor 

Catalog the choice-only systemone surface, Banking77 author metrics, and the split from Kev, trained Laya, OpenJev heads, and glance. Soft scores stay off hard gates. The uniqueness lock stays in research notes.

Co-authored-by: Basit Mustafa <24601@users.noreply.github.com>
@cursor

cursor Bot commented Sep 21, 2026

Copy link
Copy Markdown

ADV #76 lev (draft, not merged)

Reviewed cursor/fold-peterpme-lev-4eec @ 192c6902ecaf1bc4bfb67fc05dcbb6b3c5b42790 against base main @ c9b744ef (#74). Live recapture of peterpme/lev at HEAD 9b379262e45f4bc2ebecccc2c7927800c61bdc99 (pushed 2026-09-21T18:13:30Z). README blob 5c2e44b1c62faac4902c517f893fbcfa4006c353. GOAL blob dc5fd89ac4ceac5d5e4158c7f95bc5448eeb8dd9. Baseline blob 6b90b5f51e7fd4264402b1290a2751b4e9b5871f (benchmarks/banking77-v1/baselines/lev-banking77-1500.json). Description hash 77faa8872a8d. GitHub license null. LICENSE HTTP 404. Repo id 1379945478, size 157, created 2026-09-21T14:45:37Z, 1 star, 0 forks.

ADV_FAIL

Honesty gates that hold

  1. Live README, GOAL, and the checked-in baseline match the card. Kev-line Qwen/Qwen2.5-0.5B, LoRA rank 16, pointer head, Banking77 (legacy-datasets/banking77). POST /v1/systemone is choice only. README: noul and score are future steps. GOAL step 15 leaves them later. README: learning project and working prototype, not a production decision system. License file absent. Do not invent an SPDX.
  2. Baseline metrics on 150 records / 77 options are theirs: accuracy 0.88, NLL 0.517905, Brier 0.187623, ECE 0.040830, mean confidence 0.881221, coverage at 0.9 0.686667, accuracy at 0.9 0.980583, median latency 117.241 ms. Those digits are the JSON rounded. GOAL step 12 is a separate sentence: 88.67% on 150 held-out rows. The card keeps both and does not average them.
  3. Soft scores stay soft. Pointer softmax is for rank or threshold. The card says that is not a safety gate, and softmax over options is not a calibrated Noul. No Augustus runtime call site. lev is not imported into the skill.
  4. Human README still ends at ## License. This PR does not edit README.md. The lev lock lives in research/notes.md and the fingerprint row. python3 .agents/skills/augustus/scripts/uniqueness_gate.py exits 0 (uniqueness-gate ok). Inserted prose adds no em dashes.
  5. Class framing holds. notes.md §149 names the decision-model class and TypeSafe Jev (Choice, Score, Noul) as the dominant exemplar. Lev sits beside jaredpalmer/kev (§45). It is not hosted TypeSafe, not trained Laya, not hf:thaitea/laya-vision, not Zefan-Cai/Open-Jev, not IamBusy/OpenJev, and not yoheinakajima/glance (§147). Namesakes heapwolf/lev, YasserAsmi/lev, connectFree/lev, and rgrinberg/lev are other projects.
  6. Occupancy on this base is the next free id. Glance is §147. Hourly 1203 is §148. This fold is §149 for lev alone. It does not absorb hourly 1256.

Spot-recapture failure (fix)

Release tag 1.0.0 exists. The card says it does not. GitHub release Lev 1.0.0 was published 2026-09-21T17:30:54Z (created 2026-09-21T17:14:42Z, not a draft, not a prerelease). Lightweight tag 1.0.0 points at 49d23ff319dacc266e44b55ed1f6c74fb18c34b9 ("Rewrite README for public users"). HEAD 9b379262 is one commit ahead of that tag ("Document benchmark iteration workflow", 2026-09-21T18:13:28Z). The release predates this fold's last look (~19:31Z).

The card prose says "No release tag." research/revisit_fingerprints.json stores "release_tag": null. Revisit stores the latest release tag. Record 1.0.0 at 49d23ff, and say HEAD is the later commit. Do not retarget the card's HEAD, README SHA, or baseline SHA onto the tag.

Concurrent #75 (merge-order)

Draft PR #75 (cursor/hourly-1256-fold-981a @ b8f48cfa793c7995864ad65d5a07492abbb6c9dc) is open, draft, and dirty against this same main @ c9b744ef. Its published head still claims notes.md §148 for hourly 1256. The heading is ## 148. Hourly 1256 HIGH. The lock still says notes.md §148 and "do not push onto open #74". That branch does not contain §149. The rebase onto post-#74 that would retarget 1256 to §149 has not landed.

§149 on this base is lev's. Do not absorb the 1256 batch into #76. Merge #76 first once it reaches ADV_PASS, so the single peer keeps §149. Then rebase #75 and move hourly 1256 to §150, including its uniqueness lock and fingerprints. This review is ADV_FAIL, so #76 is not ready to merge. Do not merge #76. Do not merge #75.

Draft left open.

ADV_FAIL

Concrete fix, in the card and in the fingerprint: replace "No release tag" / release_tag: null with GitHub release 1.0.0 at 49d23ff319dacc266e44b55ed1f6c74fb18c34b9, one commit behind HEAD 9b379262e45f4bc2ebecccc2c7927800c61bdc99. Leave the draft open.

Sources:

cursor Bot pushed a commit that referenced this pull request Sep 21, 2026
Card the Boise 1256 decision-model class signals, densify jevframe (§48) and roster-shapes (§134), and lock composition items 777-792 plus findings batch #130. Rebased onto c9b744e (merged #74 hourly 1203 §148). Leave §149 for peterpme/lev draft #76. Third-party numbers stay theirs. Does not bump 0.5.0 or 0.5.1.

Co-authored-by: Basit Mustafa <24601@users.noreply.github.com>
HEAD 9b379262 stays one commit ahead of the tag. Fingerprint release_tag is 1.0.0. Lev remains notes.md §149.

Co-authored-by: Basit Mustafa <24601@users.noreply.github.com>
@cursor

cursor Bot commented Sep 21, 2026

Copy link
Copy Markdown

Release fingerprint corrected. GitHub release Lev 1.0.0 is tag 1.0.0 at 49d23ff319dacc266e44b55ed1f6c74fb18c34b9 (published 2026-09-21T17:30:54Z). HEAD 9b379262e45f4bc2ebecccc2c7927800c61bdc99 is one commit ahead. release_tag is 1.0.0. Lev stays notes.md §149. Draft left open. Do not merge.

@24601

24601 commented Sep 21, 2026

Copy link
Copy Markdown
Owner Author

ADV #76 lev re-check (draft, not merged)

Re-reviewed cursor/fold-peterpme-lev-4eec @ 1e9cec7d537248cc3091b8fcfcd58f4a09568945 (release-tag commit on top of 192c690) against base main @ c9b744ef. Live recapture of peterpme/lev still at HEAD 9b379262e45f4bc2ebecccc2c7927800c61bdc99 (pushed 2026-09-21T18:13:30Z). README blob 5c2e44b1c62faac4902c517f893fbcfa4006c353. GOAL blob dc5fd89ac4ceac5d5e4158c7f95bc5448eeb8dd9. Baseline blob 6b90b5f51e7fd4264402b1290a2751b4e9b5871f. Description hash 77faa8872a8d. GitHub license null. LICENSE HTTP 404. Repo id 1379945478, size 157, created 2026-09-21T14:45:37Z, 1 star, 0 forks.

The prior ADV_FAIL was the missing release tag. That fix holds.

Release tag

GitHub release Lev 1.0.0 is the only release. Tag name 1.0.0, published 2026-09-21T17:30:54Z, created 2026-09-21T17:14:42Z, not a draft, not a prerelease. The git ref is a lightweight tag whose object is commit 49d23ff319dacc266e44b55ed1f6c74fb18c34b9 ("Rewrite README for public users"). Compare of that commit to HEAD is ahead by 1, behind by 0: 9b379262 ("Document benchmark iteration workflow", 2026-09-21T18:13:28Z).

Card, uniqueness lock, research/sources.json, and research/revisit_fingerprints.json agree: release_tag 1.0.0, tag commit 49d23ff319dacc266e44b55ed1f6c74fb18c34b9, HEAD 9b379262e45f4bc2ebecccc2c7927800c61bdc99 one commit ahead. README SHA and baseline SHA stay on the HEAD blobs. The phrase "No release tag" is gone from the lev card. The baseline file is absent on the tag tree and present at HEAD, so keeping its SHA on 9b379262 is the right pin.

Reconfirm

  1. Live README, GOAL, and the checked-in baseline still match the card. Kev-line Qwen/Qwen2.5-0.5B, LoRA rank 16, pointer head, Banking77 (legacy-datasets/banking77). POST /v1/systemone is choice only. README: noul and score are future steps. GOAL step 15 leaves them later. README: learning project and working prototype, not a production decision system. License file absent.
  2. Baseline JSON on 150 records / 77 options is theirs, and the card's digits are that JSON rounded: accuracy 0.88, NLL 0.517905, Brier 0.187623, ECE 0.040830, mean confidence 0.881221, coverage at 0.9 0.686667, accuracy at 0.9 0.980583, median latency 117.241 ms. GOAL step 12 is a separate sentence: 88.67% on 150 held-out rows. The card keeps both and does not average them.
  3. Soft scores stay soft. Pointer softmax is for rank or threshold. The card says that is not a safety gate, and softmax over options is not a calibrated Noul. The 0.9 slice is a rank or threshold fact. No Augustus runtime call site. This diff does not import lev. README.md is untouched and still ends at ## License.
  4. python3 .agents/skills/augustus/scripts/uniqueness_gate.py exits 0 (uniqueness-gate ok). The lev lock is one substring in research/notes.md §149. Inserted prose adds no em dashes.
  5. Class framing holds. §149 names the decision-model class and TypeSafe Jev (Choice, Score, Noul) as the dominant exemplar. Lev sits on the open Kev line beside jaredpalmer/kev (§45) and does not replace that card. It is not hosted TypeSafe, not trained Laya, not an OpenJev scalar head, and not yoheinakajima/glance (§147).
  6. §149 on this branch is lev alone. The section has no hourly 1256 text. Draft PR #75 @ a48ff112 is ## 150. Hourly 1256 HIGH, and its lock says notes.md §150. It records that draft Fold peterpme/lev (educational Kev pointer head) #76 owns §149 and does not reclaim lev. Merge Fold peterpme/lev (educational Kev pointer head) #76 first, then Fold hourly 1256 HIGH into notes §150 #75.

Draft left open. Do not merge.

ADV_PASS

Sources:

@24601
24601 marked this pull request as ready for review September 21, 2026 19:59
@24601
24601 merged commit 6ef49bc into main Sep 21, 2026
1 check passed
cursor Bot pushed a commit that referenced this pull request Sep 21, 2026
Card the Boise 1256 decision-model class signals, densify jevframe (§48) and roster-shapes (§134), and lock composition items 777-792 plus findings batch #130. Rebased onto c9b744e (merged #74 hourly 1203 §148). Leave §149 for peterpme/lev draft #76. Third-party numbers stay theirs. Does not bump 0.5.0 or 0.5.1.

Co-authored-by: Basit Mustafa <24601@users.noreply.github.com>
24601 added a commit that referenced this pull request Sep 21, 2026
Card the Boise 1256 decision-model class signals, densify jevframe (§48) and roster-shapes (§134), and lock composition items 777-792 plus findings batch #130. Rebased onto c9b744e (merged #74 hourly 1203 §148). Leave §149 for peterpme/lev draft #76. Third-party numbers stay theirs. Does not bump 0.5.0 or 0.5.1.

Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: Basit Mustafa <24601@users.noreply.github.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants