Fold Abide as productized soft-rule lint; kev Hub/NOTA delta

.agents/skills/augustus/references/composition-algebra.md

@@ -16,8 +16,8 @@ mappings.md conventions.
 |---|---|---|---|---|---|
 | 1 | **Operand** | F(Jev(...)) — judgment's number feeds the function | Nouls as CatBoost features; Score as PUCT leaf value | It's a calibrated belief in your rubric's units, not a natural quantity; version feature/question defs with the consumer | **Empirical recipe** |
 | 2 | **Post-judge** | F(x) → Jev judges the result | Output judge (leaks_secret, failure_class); citation check on generated text | Only the post-judge sees what the call printed; the pre-gate cannot | **Empirical recipe** |
-| 3 | **Gate** | if Jev(x): apply F — Jev decides *whether* F runs, or whether F's result is admitted | Pre-action gates (destructive .90/exfil .70); winnow context sieve | A gate is a filter, not authorization — validate operation+target in code. Error paths fail **per action**, not always open: an advisory guard fails open *because* a hard interlock or sandbox sits underneath; a gate that selects or authorizes a side effect fails closed (`mixed-architecture.md` prefilter table; `mappings.md` §18) | **Empirical recipe** |
-| 4 | **Selector (of F or its parameters)** | Jev picks which F runs: Choice over functions/models/effort levels | jev-router (cheapest model), jev-codex-router (model+effort), DiffJury review_depth | Dispatch stays in code; per-option consequences are your cost model; confidence-gate the selection | **Empirical recipe** |
+| 3 | **Gate** | if Jev(x): apply F — Jev decides *whether* F runs, or whether F's result is admitted | Pre-action gates (destructive .90/exfil .70); winnow context sieve; pi-heed side-effect check | A gate is a filter, not authorization — validate operation+target in code. Error paths fail **per action**, not always open: an advisory guard fails open *because* a hard interlock or sandbox sits underneath; a gate that selects or authorizes a side effect fails closed (`mixed-architecture.md` prefilter table; `mappings.md` §18) | **Empirical recipe** |
+| 4 | **Selector (of F or its parameters)** | Jev picks which F runs: Choice over functions/models/effort levels | jev-router (cheapest model), jev-codex-router (model+effort), DiffJury review_depth, jeffrey next-tool | Dispatch stays in code; per-option consequences are your cost model; confidence-gate the selection. **Pick ≠ fill:** the LLM may write args; Jev does not | **Empirical recipe** |
 | 5 | **Comparator** | Replace a semantic comparator inside sort/rank: "more relevant / more severe" as a key | Rerank; skillranker; order statistics over semantic keys | Comparability needs a shared rubric; measure recall separately from rerank quality | **Empirical recipe** |
 | 6 | **Prior / initializer** | Jev distribution seeds a deterministic method that refines it | MCTS PUCT priors; beam-search branch priority | It's a heuristic prior, not a posterior; refine with real observations | **Empirical recipe** |
 | 7 | **State estimator, F = controller** | Jev estimates named probabilities; deterministic policy with hysteresis acts | foreman (progress/stuck/complete → continue/stop/retry/verify) | The model never commands; interventions enumerated in code | **Empirical recipe** |
@@ -40,6 +40,34 @@ mappings.md conventions.
   never the model.
 - **→ (implication) / chains**: decompose into gate → act → post-judge;
   never encode multi-hop logic in one question (indirection costs accuracy).
+- **Named combinators** ([jev-combinators](https://github.com/voidning/jev-combinators);
+  renamed from decision-combinators, same repo):
+  Then / Gate / Vote / Cascade / Weighted plus **extended**
+  Router / Loop / Retry / Fallback / Memory are
+  **control-plane** wiring, not a license to treat parallel
+  Nouls as independent. Digital-design slogan (transistors /
+  logic gates / chip) is a *metaphor* for soft classifiers;
+  ∧/∨ aggregation still follows the rule above. Fallback is
+  the fail-closed node;   Memory gates what to remember. No
+  measurements. `notes.md` §66, §69.
+- **Decider ≠ executor** ([jeffrey](https://github.com/thomasbrueggemann/jeffrey)):
+  position 4 (Selector of next F) stays Jev; arg fill is
+  generation, not a Jev position. The loop is Jev→tool→Jev.
+  Pick ≠ fill. Mapping §9 still rejects the fused
+  planner-writer. `notes.md` §70.
+- **Selector of the next word** ([jev-gpt](https://github.com/florian-hoenicke/jev-gpt)):
+  position 4 applied to generation itself. Each token is a
+  Choice over a closed lexicon; the model never
+  free-generates. Architecture demo (~400 calls / 75 s /
+  2¢ *theirs*). Distinct from jeffrey (selector of next
+  *tool*). `notes.md` §71.
+- **Hand no-text steps** ([jev-use](https://github.com/shitianfang/jev-use)):
+  selector + gate; writing stays generation. Vercel drops
+  confidence so margin is a different statistic.
+  `notes.md` §71.
+- **TLA+ kernel around votes** ([jev-labs](https://github.com/copyleftdev/jev-labs)):
+  aggregation (quorum, stability, escalate) is the spec,
+  not a multiplied joint of five Nouls. `notes.md` §67.
 
 ## Rules that hold across every position
 
@@ -143,10 +171,22 @@ Reusable shapes when generating applications:
    meeting action items ~150 ms after each utterance.
 7. **Formula embedding**: JUDGE/SCORE/CHOOSE as first-class spreadsheet formulas.
 8. **Pixel-free computer use**: accessibility tree → compact actionable-JSON → one
-   batched question set per step → execute via AX actions.
+   batched question set per step → execute via AX actions. Encoder-backend
+   cousin: gliner2-ultrafast scores observed a11y/DOM controls with
+   local GLiNER2; code clicks; `DONE` ≠ success (`notes.md` §52).
+   Specialist-form cousin: Cua-S1 option-attention (fill/check/click/skip);
+   not TypeSafe Jev; plan ≠ execute; source-only (`notes.md` §54).
 9. **Shadow-mode harness** (jev-harness): policy + confidence gate + shadow mode +
    offline eval CLI replaying fixtures, asserting on actions; 24-row filter 48.9 s
-   (Claude CLI) vs 1.3 s Jev at concurrency 8.
+   (Claude CLI) vs 1.3 s Jev at concurrency 8. Compaction rollout:
+   gliner25-compaction public default `shadowMode: true` (log proposed
+   reduction; do not replace history) (`notes.md` §50).
+   Stdout-prune cousin: [jev-pruner](https://github.com/tamaratran/jev-pruner)
+   archives full stdout before scoring; fail-safe keep original
+   (`notes.md` §53). Marketplace id still `fast-jev-output`.
+   Recovery cousin: [jevons](https://github.com/LilDojd/jevons) default
+   recovery **shadow** (record, do not interrupt); steering never
+   generates commands (`notes.md` §51).
 
 Calibration warning (calibre): routing thresholds and ROI do **not** transfer across
 datasets — every gate is a per-dataset measurement (see validation.md).
@@ -155,10 +195,124 @@ datasets — every gate is a per-dataset measurement (see validation.md).
     Jev as gate/selector/verifier *around* a generator, never instead of one.
     Cost-sensitive prefilter (drop chunks/lines/hunks before the LLM);
     tool/skill routing (Choice + fits-Noul, code dispatches); preference lint
-    (project-defined rules as criteria). Fail-open vs fail-closed is per
+    (project-defined *soft* rules as criteria; linter owns hard rules;
+    Abide is the productized path, `notes.md` §47). Fail-open vs fail-closed is per
     action — LlamaIndex Jev rerank fails open (keep retrieval order), select
     fails closed. Full card: `references/mixed-architecture.md`.
-11. **Structural prove ∩ remainder judge** (jevgate, doc-router): code
-    (allowlist, text layer) decides the easy cases; typed questions only
+11. **Structural prove ∩ remainder judge** (jevgate, doc-router, Abide): code
+    (allowlist, text layer, linter) decides the easy cases; typed questions only
     on leftovers; fail-open unless a real sandbox sits under. Full card:
     `mappings.md` §18.
+12. **1-token selector / tree of Choices** (chakuho, jev-gpt): the
+    generator is reduced to a next-label or next-word Choice. Softmax
+    over declared labels is not a Noul. Numeric rules and writing stay
+    exact. Full cards: `judgment-class.md`, `mixed-architecture.md`.
+13. **Productized System One HTTP** (classifier-dev): the public
+    contract is label + calibrated confidence, not a paragraph. Batch
+    state, escalate-under-threshold, and a `FALLBACK` marker are
+    *code*. Full cards: `mixed-architecture.md`, `validation.md`.
+14. **Evidence-synthesis pointer** (choxos/jev-reviewer, ≠ egma-ai):
+    Jev picks line ids; code copies verbatim; a second absolute Noul
+    checks "does this line itself answer?"; *Not found* is an answer;
+    the human tick is the product. Full cards: `applied-mappings.md`
+    §2, `mixed-architecture.md`.
+15. **Prompted-JSON wire** (githubnext/localjev, ≠ kunchenguid/local-jev):
+    the TypeSafe SDK talks to a local `/v1/systemone`; the model
+    *writes* probabilities rather than exposing logits. Entropy
+    confidence is computed in code from that vector. Full cards:
+    `judgment-class.md`, `mixed-architecture.md`, `validation.md`.
+16. **Script-before-p router** (NandhaKishorM/laya packaging of Hub
+    Laya): pick the checkpoint from script/lang/task *before* the
+    forward pass, because confidence will not drop on OOD (Khmer
+    0.000@0.952). Post-T ECE is not raw ECE; 0.85 is still soft.
+    Full cards: `judgment-class.md`, `mixed-architecture.md`,
+    `faq.md`, `validation.md`.
+17. **External class census** (@airesearch12 / Benchmark
+    Heaven): a named list is not a rank; GLiNER2 and
+    routers on the list are a class-boundary, not
+    identity; incompleteness is lag. Watch the board
+    URL; do not paste live scores into the census card.
+    Full cards: `mixed-architecture.md`, `faq.md`,
+    `validation.md`, `toolbox-mapping.md`.
+18. **Geometric-mean product** (JevBench v1.2): four
+    axes at 25% each; a weak axis cannot be bought
+    back; weighting is a product design, not a law;
+    calibration on the rank is a choice (v1.1 kept it
+    off); instruction models in the same table as NAR
+    rebuilds; ×2 latency and est. costs are assumptions
+    to name. Full cards: `mixed-architecture.md`,
+    `validation.md`, `faq.md`, `mental-models.md`.
+19. **Already-folded class as a recipe** (hourly 0842):
+    when the named HIGHs are already on the branch,
+    extract how-to-apply instead of re-carding —
+    wire-compat ≠ logit-equiv; productize label+p and
+    mark `FALLBACK`; packaging ≠ new species / script-
+    before-p; pointer-not-generator (two-pass; *Not
+    found*; human tick); external census ≠ scored
+    bake-off / geo-mean weights are a design. Skip
+    thin noise. Hard-gating a Noul as a PR/quality
+    gate is soundness theater. Full cards:
+    `mixed-architecture.md`, `faq.md`,
+    `mental-models.md`, `validation.md`.
+20. **S1 keeps flying / S2 one-use** (khordoo/jev-reflex-autonomy-lab
+    delta of §46): position 4 (Selector of next
+    *action*) stays on the reflex every tick;
+    position 7 (state estimator) is the optional
+    planner — advice, not a command. Escalate-
+    under-threshold **without stalling**. Log
+    consumption, not arrival. Local rule-based vs
+    Live API is an A/B of backends, not a scored
+    bake-off; Local controller **≠** githubnext/localjev.
+    Seed = geometry ≠ async replay. No pixels.
+    Confidence ≠ selected probability. 20% still
+    soft. S2 never grants. Full cards:
+    `mixed-architecture.md`, `faq.md`,
+    `mental-models.md`, `agent-self-assessment.md`,
+    `validation.md`.
+21. **OCR+AX observe→score→act** (awlevin/typesafe-computer-use):
+    position 10 (Discretizer: screen → numbered items)
+    then position 4 (Selector of next action). Writer
+    is generation, not a Jev position. Split kind/item/site
+    is width-is-cheap. Overlap is concentration theater.
+    Perception in code rebuilds pixel-free reasoning.
+    Decision never ships screenshots; the answer reader
+    may. 155× is one screenshot *theirs*. Full cards:
+    `mixed-architecture.md`, `faq.md`,
+    `applied-mappings.md` §9, `validation.md`.
+22. **ASR observe→score→act** (moritzkremb/jev-voice-browser):
+    position 10 (Discretizer: waveform → transcript +
+    numbered elements) then position 4 (Selector).
+    Width-is-cheap: 9–11 questions on one request.
+    Partial-speech wait is VOI (closed-set vs free-text).
+    Spoken confirm is not a grant. Overlay numbers are
+    exact, not a second model. Compose with item 21
+    (OCR). Full cards: `mixed-architecture.md`, `faq.md`,
+    `applied-mappings.md` §9, `validation.md`.
+23. **Wrap-as-execution ALLOW/ASK/DENY** (reddpy/AgentGhost):
+    position 3 (constraint on the actuator path) then
+    position 4 (Selector on leftovers). Rules prove;
+    Jev remainder; ASK is an error not a log line.
+    Fail-closed on judge error. Distinct from actiongate
+    (evidence ≠ authority) and jev-use (fail-open).
+    Full cards: `mixed-architecture.md`, `faq.md`,
+    `applied-mappings.md` §7, `mappings.md` §8/§18.
+24. **Application genre atlas** (@studio_yebisu):
+    position 11 (catalog of holes, not a score).
+    Stars are research-time. Same discipline as item 17
+    (class census ≠ bake-off). Full cards:
+    `mixed-architecture.md`, `faq.md`, `validation.md`.
+25. **External pedagogy / how-to-apply** (@akshay_pachaar):
+    position 11 (explainer of already-owned placements,
+    not a new construct). LLM hammer; code owns
+    branches; schema-safe ≠ correct; shadow +
+    questions-as-code. 200×/400× are TypeSafe ceiling.
+    Full cards: `mixed-architecture.md`, `faq.md`,
+    `mental-models.md`.
+26. **Boolean composition of soft Nouls** (uehaj/jev-semgrep):
+    position 5 (Comparator over lines) then code ∧/∨/¬
+    on *thresholded* bits — never multiply parallel p
+    (the logical-operator caveat above). Proposition ≠
+    embedding (contrast-set refund). Not a Gate (position
+    3). **≠** jev-combinators digital-design metaphor
+    **≠** semgrep.dev. Full cards: `mixed-architecture.md`,
+    `faq.md`, `applied-mappings.md` §4, `mappings.md` §4.