Skip to content

Densify SemIf rename + MLX + accuracy ladder #37

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
24601 merged 2 commits into from
Sep 20, 2026
Merged

Densify SemIf rename + MLX + accuracy ladder #37

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
47a870a
Densify SemIf rename + MLX + ladder onto main (2275★, §117 / 330–336).
cursoragent Sep 20, 2026
fbe0ba5
Relock SemIf live REST 2282★ on §117 densify.
cursoragent Sep 20, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
5 changes: 3 additions & 2 deletions .agents/skills/augustus/SKILL.md
View file
Open in desktop

Large diffs are not rendered by default.

3 changes: 3 additions & 0 deletions .agents/skills/augustus/references/agent-self-assessment.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -936,3 +936,6 @@ User-provided 0915 uniqueness lock: TianyuCodings/NanoJev unified-games-v1 densi
**0920 jcr:** pre-action lookup is not a done-check and not a permission gate. Ambiguity / no-match / depth-limit are explicit abstention paths. NiazMorshed2007/jcr. `notes.md` §116.

User-provided 0920 jcr uniqueness lock: NiazMorshed2007/jcr MIT; site https://jcr.niazmorshed.dev; topics ai-agents,jev,mcp; **4★**; HEAD `138b3832`; README SHA `2a49dbc1`; LICENSE SHA `46231303`; size **14850**; Jev Capability Resolver; one tool to find documented deterministic commands in a nested capability tree; returns context; **does not execute**; skills = workflow+judgment; capabilities = individual operations; format independent of Jev; proposed open standard exploration; classify (Jev) → optional OpenAI decompose compound → beam search geometric mean of routing probs; keep up to 3 paths ≥60% of best (JCR_BAND_RATIO 0.6); ambiguity / no-match / depth-limit explicit; soft scores ≠ hard gates; 0.6 band is application policy; routing ≠ permission; docs ≠ authority to run; sol-vs-opus5-20 *theirs*: 20 scenarios × 4 variants = 80 runs; lookup+explain only, no execution; Claude Opus 5: agent input 108,585→15,819 (−85%), cost $0.3700→$0.1222 (−67%), wall 105.5s→77.7s; Codex GPT-5.6-Sol: 61,952→47,669 (−23%), $0.1377→$0.1151 (−16%), wall 25.3s→62.4s (Sol slower with JCR in 19/20); One Sol outlier 372.6s / 193 Jev calls; n=1 per cell; Not Harbor task-execution; Claude/Codex harnesses; compare mode; 50 scenarios bundled; 11 groups, 960 nodes, 11,360 items; 16 routing rounds per step; NiazMorshed2007/jcr ≠ skill-broker ≠ skillranker ≠ jev-sift ≠ jev-lens ≠ jevusher ≠ jev_select_capability; do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33/#34; notes.md §116

User-provided 0922 (`notes.md` §117 / items 330–336 / batch #100). SemIf was formerly OpenJev; MLX backend; 5.21× systems≠semantic; Softmax over options ≠ calibrated Noul; live REST 2282★. Do not reopen #23–#36/#38; do not push onto open #39/#40.
User-provided 0922 uniqueness lock: SemIf was formerly OpenJev; independent; not affiliated with Jev or TypeSafe; homepage openjev.com; default master; MIT; HEAD ca3ba65f1429; Tolerate float roundoff in MLX evidence verification; pushed 2026-09-19; live REST 2282★ / 140 forks; size 9177; README SHA 74ab7f7f; LICENSE SHA ca562883; interface pattern reproduction with open models; does not reproduce Jev undisclosed model/training; Direct option logits; 0 output tokens; shared-state parallel; MLX backend for Apple Silicon (`--backend mlx`); Qwen3.5-4B 3090 direct 1.023s vs AR JSON 5.332s (**5.21×**); argmax agree 18/21; systems comparison ≠ semantic equivalence; Parallel suffixes 20.03 dec/s on 777 decisions; Browser ladder Qwen3.5-4B authored BA 0.813, pert 0.766, TypeSafe subset agreement 0.845 vs Published Jev 0.883 (102 across 20 cases); Softmax over options ≠ calibrated Noul; typed output does not guarantee semantic correctness; wire/agreement ≠ replica of TypeSafe; SemIf ≠ kw2828/OpenJev playground ≠ zhihz/openjev ≠ apiplant/semif-rs port ≠ dddanielliu/semif-serve; rename is densify not a second census; JevBench 74.6 is §78 not this ladder; do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33/#34/#35/#36/#38; do not push onto open #39/#40; notes.md §117
3 changes: 3 additions & 0 deletions .agents/skills/augustus/references/applied-mappings.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2505,3 +2505,6 @@ User-provided 0915 uniqueness lock: TianyuCodings/NanoJev unified-games-v1 densi


User-provided 0920 jcr uniqueness lock: NiazMorshed2007/jcr MIT; site https://jcr.niazmorshed.dev; topics ai-agents,jev,mcp; **4★**; HEAD `138b3832`; README SHA `2a49dbc1`; LICENSE SHA `46231303`; size **14850**; Jev Capability Resolver; one tool to find documented deterministic commands in a nested capability tree; returns context; **does not execute**; skills = workflow+judgment; capabilities = individual operations; format independent of Jev; proposed open standard exploration; classify (Jev) → optional OpenAI decompose compound → beam search geometric mean of routing probs; keep up to 3 paths ≥60% of best (JCR_BAND_RATIO 0.6); ambiguity / no-match / depth-limit explicit; soft scores ≠ hard gates; 0.6 band is application policy; routing ≠ permission; docs ≠ authority to run; sol-vs-opus5-20 *theirs*: 20 scenarios × 4 variants = 80 runs; lookup+explain only, no execution; Claude Opus 5: agent input 108,585→15,819 (−85%), cost $0.3700→$0.1222 (−67%), wall 105.5s→77.7s; Codex GPT-5.6-Sol: 61,952→47,669 (−23%), $0.1377→$0.1151 (−16%), wall 25.3s→62.4s (Sol slower with JCR in 19/20); One Sol outlier 372.6s / 193 Jev calls; n=1 per cell; Not Harbor task-execution; Claude/Codex harnesses; compare mode; 50 scenarios bundled; 11 groups, 960 nodes, 11,360 items; 16 routing rounds per step; NiazMorshed2007/jcr ≠ skill-broker ≠ skillranker ≠ jev-sift ≠ jev-lens ≠ jevusher ≠ jev_select_capability; do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33/#34; notes.md §116

User-provided 0922 (`notes.md` §117 / items 330–336 / batch #100). SemIf was formerly OpenJev; MLX backend; 5.21× systems≠semantic; Softmax over options ≠ calibrated Noul; live REST 2282★. Do not reopen #23–#36/#38; do not push onto open #39/#40.
User-provided 0922 uniqueness lock: SemIf was formerly OpenJev; independent; not affiliated with Jev or TypeSafe; homepage openjev.com; default master; MIT; HEAD ca3ba65f1429; Tolerate float roundoff in MLX evidence verification; pushed 2026-09-19; live REST 2282★ / 140 forks; size 9177; README SHA 74ab7f7f; LICENSE SHA ca562883; interface pattern reproduction with open models; does not reproduce Jev undisclosed model/training; Direct option logits; 0 output tokens; shared-state parallel; MLX backend for Apple Silicon (`--backend mlx`); Qwen3.5-4B 3090 direct 1.023s vs AR JSON 5.332s (**5.21×**); argmax agree 18/21; systems comparison ≠ semantic equivalence; Parallel suffixes 20.03 dec/s on 777 decisions; Browser ladder Qwen3.5-4B authored BA 0.813, pert 0.766, TypeSafe subset agreement 0.845 vs Published Jev 0.883 (102 across 20 cases); Softmax over options ≠ calibrated Noul; typed output does not guarantee semantic correctness; wire/agreement ≠ replica of TypeSafe; SemIf ≠ kw2828/OpenJev playground ≠ zhihz/openjev ≠ apiplant/semif-rs port ≠ dddanielliu/semif-serve; rename is densify not a second census; JevBench 74.6 is §78 not this ladder; do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33/#34/#35/#36/#38; do not push onto open #39/#40; notes.md §117
43 changes: 43 additions & 0 deletions .agents/skills/augustus/references/composition-algebra.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2088,6 +2088,48 @@ User-provided 0915 uniqueness lock: TianyuCodings/NanoJev unified-games-v1 densi
positions 8 (Metric) × 11 (Bounds). Claude/Codex harnesses. compare mode. 50 scenarios bundled.
Full cards: `validation.md`.


330. **SemIf rename densify** (TheoLeeCJ/SemIf):
position 1 (Operand). SemIf was formerly OpenJev. rename is densify not a second census.
independent; not affiliated with Jev or TypeSafe. homepage openjev.com. default master. MIT.
live REST 2282★ / 140 forks. HEAD ca3ba65f1429.
Full cards: `judgment-class.md`, `faq.md`, `mental-models.md`.
331. **Interface pattern ≠ replica** (TheoLeeCJ/SemIf):
position 1 (Operand). interface pattern reproduction with open models;
does not reproduce Jev undisclosed model/training.
wire/agreement ≠ replica of TypeSafe.
Full cards: `judgment-class.md`, `faq.md`.
332. **Direct logits / 0 tokens / shared-state** (TheoLeeCJ/SemIf):
positions 1 (Operand) × 8 (Metric). Direct option logits; 0 output tokens; shared-state parallel.
Full cards: `mixed-architecture.md`, `judgment-class.md`.
333. **MLX Apple Silicon densify** (TheoLeeCJ/SemIf):
position 1 (Operand). MLX backend for Apple Silicon (`--backend mlx`).
Tolerate float roundoff in MLX evidence verification.
typed output does not guarantee semantic correctness.
Full cards: `judgment-class.md`, `faq.md`.
334. **Speed 5.21× systems comparison** (TheoLeeCJ/SemIf):
position 8 (Metric). Qwen3.5-4B 3090 direct 1.023s vs AR JSON 5.332s (**5.21×**);
argmax agree 18/21; systems comparison ≠ semantic equivalence.
Full cards: `validation.md`, `faq.md`.
335. **Parallel suffixes 20.03 densify** (TheoLeeCJ/SemIf):
position 8 (Metric). Parallel suffixes 20.03 dec/s on 777 decisions.
Full cards: `validation.md`.
336. **Browser ladder + cousins** (TheoLeeCJ/SemIf):
positions 8 (Metric) × 11 (Bounds). authored BA 0.813, pert 0.766,
TypeSafe subset agreement 0.845 vs Published Jev 0.883 (102 across 20 cases).
Softmax over options ≠ calibrated Noul. JevBench 74.6 is §78 not this ladder.
SemIf ≠ kw2828/OpenJev playground ≠ zhihz/openjev ≠ apiplant/semif-rs port ≠ dddanielliu/semif-serve.
Full cards: `validation.md`, `faq.md`, `mental-models.md`.

User-provided 0922 items 330–336 (`notes.md` §117). Do **not**
re-fold merged #36 §115 items 303–308 / §114 items 289–302 / §113 items 273–288
/ merged #38 §116 items 309–316. Skip open #40 items 322–329.
Soft Noul ≠ hard safety. Ranking ≠ calibration theater.
SemIf 2282★ (+41 vs §113 2241; +45 vs §114 2237);
do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33/#34/#35/#36/#38.
do not push onto open #39/#40.


User-provided 0920 jcr items 309–316 (`notes.md` §116). Do **not**
re-fold 0743 items 273–288 / merged #30 items 268–272 / 0646 items 248–267.
Merged #35 owns §114 / items 289–302 / batch #97. Merged #36 owns §115 / 303–308 / #98. Open #37 owns §117 / 315–321 / #100 (item overlap 315–316 is #37's remap).
Expand Down Expand Up @@ -2192,3 +2234,4 @@ Soft Noul ≠ hard safety.
Hourly 0843 uniqueness lock: A hunch is a probability with a policy attached; { enter: 0.8, exit: 0.6 } is hysteresis; replay a policy change without inference; Decision models are providers, not the product; huncho ≠ Kungie/gut ≠ carldaws/hunch ≠ tpellet/hunch; pretrained Qwen2.5 base ECE 0.030 (0.5B) / 0.040 (7B); instruct 0.302 / 0.269; 70.9% → 70.0% mean conf 74.1% → 96.7%; temperature scaling still matches it in-distribution; No Jev API was called; Qwen2.5 ≠ Archer; Qwen/Qwen3.8-27B ≠ Archer; 学習済みモデル v0.1 は準備中です; bool AUROC 0.523; 先頭だと0件、末尾だと250件; 温度を渡さない場合、確率は較正されていません; このリポジトリには Jev を呼ぶコードが存在しません; g0runmezadam/what-is-jev IS tunahansahin897/what-is-jev (same GitHub id 1378007307); 947 repos scored; A 273 · B 302 · C 372; LLM rubric ≠ benches; Data as of 2026-09-20; HEAD 895b9498; README SHA 3ae98c56; 13 focused checks and one mutually exclusive outcome; Probabilities are advisory, not calibrated guarantees; omni-/ask-jev ≠ pedroknigge/mcp_jev; pd.cut bins by equal width while jeval bins by quantile; ECE 0.113 and ECE 0.076; jeval drift is not implemented yet; rlaope/jeval ≠ dayhaysoos/jevals; calibration does not compose; ECE has exactly zero statistical power to detect the failure mode that kills trajectories; 25–60× headline withdrawn; P(all-correct): 0.0071 vs 0.0001; TCE / AMS; Qwen 3.8 sparring ≠ Archer; Deferred Crispification; light_cutoff_applied_to_combination 0; recorded run, kinematic animation; BANKING77 Accuracy BERT-Base 93.02 Jev 79.90; Analyse jev calibration (NLL, ECE) backlog; BERT figures are published supervised references, not zero-shot; 档位措辞效应 分数极差中位 0.50、最大 1.32; 修好后对照组是 0.01; 不是 benchmark; 概率没做 calibration; ~1,430 API calls, about $0.15; xiaohuaxi/jev-study ≠ wjdjdakf17/jev-study ≠ baekenough/jev-study; AND: product (independence assumed and recorded in the trace); chat model's stated confidence is not calibrated; circuit-vl-4b ≠ Archer; Bring your own API key; vamsikrishna2421/jev-usecases ≠ whyashthakker/awesome-jev-use-cases; catalog ≠ endorsement; SemIf 2237★ (+30 vs §111 2207); jevlike 1049★ (+6 vs 1043); TypeLLM/TypeLLM 16★; AnotiaWang 98★; yibie/awesome-jev 520★ (+14 vs 506); Laya likes 861 (was 822); tracker likes 66 (+2 vs 64) lastModified UNCHANGED; Blackwood likes 2 gated manual; Archer still promised_not_landed; Hub archerhume/4rcherhume HTTP 401; do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33; notes.md §114

User-provided 0920 jcr uniqueness lock: NiazMorshed2007/jcr MIT; site https://jcr.niazmorshed.dev; topics ai-agents,jev,mcp; **4★**; HEAD `138b3832`; README SHA `2a49dbc1`; LICENSE SHA `46231303`; size **14850**; Jev Capability Resolver; one tool to find documented deterministic commands in a nested capability tree; returns context; **does not execute**; skills = workflow+judgment; capabilities = individual operations; format independent of Jev; proposed open standard exploration; classify (Jev) → optional OpenAI decompose compound → beam search geometric mean of routing probs; keep up to 3 paths ≥60% of best (JCR_BAND_RATIO 0.6); ambiguity / no-match / depth-limit explicit; soft scores ≠ hard gates; 0.6 band is application policy; routing ≠ permission; docs ≠ authority to run; sol-vs-opus5-20 *theirs*: 20 scenarios × 4 variants = 80 runs; lookup+explain only, no execution; Claude Opus 5: agent input 108,585→15,819 (−85%), cost $0.3700→$0.1222 (−67%), wall 105.5s→77.7s; Codex GPT-5.6-Sol: 61,952→47,669 (−23%), $0.1377→$0.1151 (−16%), wall 25.3s→62.4s (Sol slower with JCR in 19/20); One Sol outlier 372.6s / 193 Jev calls; n=1 per cell; Not Harbor task-execution; Claude/Codex harnesses; compare mode; 50 scenarios bundled; 11 groups, 960 nodes, 11,360 items; 16 routing rounds per step; NiazMorshed2007/jcr ≠ skill-broker ≠ skillranker ≠ jev-sift ≠ jev-lens ≠ jevusher ≠ jev_select_capability; do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33/#34; notes.md §116
User-provided 0922 uniqueness lock: SemIf was formerly OpenJev; independent; not affiliated with Jev or TypeSafe; homepage openjev.com; default master; MIT; HEAD ca3ba65f1429; Tolerate float roundoff in MLX evidence verification; pushed 2026-09-19; live REST 2282★ / 140 forks; size 9177; README SHA 74ab7f7f; LICENSE SHA ca562883; interface pattern reproduction with open models; does not reproduce Jev undisclosed model/training; Direct option logits; 0 output tokens; shared-state parallel; MLX backend for Apple Silicon (`--backend mlx`); Qwen3.5-4B 3090 direct 1.023s vs AR JSON 5.332s (**5.21×**); argmax agree 18/21; systems comparison ≠ semantic equivalence; Parallel suffixes 20.03 dec/s on 777 decisions; Browser ladder Qwen3.5-4B authored BA 0.813, pert 0.766, TypeSafe subset agreement 0.845 vs Published Jev 0.883 (102 across 20 cases); Softmax over options ≠ calibrated Noul; typed output does not guarantee semantic correctness; wire/agreement ≠ replica of TypeSafe; SemIf ≠ kw2828/OpenJev playground ≠ zhihz/openjev ≠ apiplant/semif-rs port ≠ dddanielliu/semif-serve; rename is densify not a second census; JevBench 74.6 is §78 not this ladder; do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33/#34/#35/#36/#38; do not push onto open #39/#40; notes.md §117
Loading
Loading