File size: 17,014 Bytes
6d07351 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 | ⸻
⸻
VERANTYX — Research Notes & Design Rationale
Why the System Is Built This Way (Unified Edition)
⸻
1. What VERANTYX Is (and Is Not)
VERANTYX is a logic-centered inference system designed to control, constrain, and audit reasoning, rather than to generate language.
It is not an LLM, and it is not an LLM wrapper by design.
At the current stage, VERANTYX relies on conventional large language models only because its internal rule database is incomplete.
This reliance is a temporary engineering choice, not a foundational dependency.
The system is explicitly designed so that as the database grows, external language models become less important — and may eventually be unnecessary for many tasks.
⸻
2. Motivation: Beyond Probabilistic Language Generation
Modern AI research largely equates intelligence with probabilistic language generation.
This approach has produced impressive results, but also systemic limitations:
• Reasoning processes are opaque
• Errors are indistinguishable from creativity
• Correct answers may be produced for incorrect reasons
• Failure modes are difficult to audit
VERANTYX exists to explore an alternative:
explicit, inspectable reasoning built from rules, counterexamples, and necessity checks.
⸻
3. Core Philosophical Commitments
VERANTYX is built on several non-negotiable principles: 1. Reasoning must be inspectable
Every conclusion must be traceable to explicit rules or counterexamples. 2. Correctness is conditional
The system distinguishes clearly between:
• proven answers
• refuted answers
• insufficient evidence
• provisional (mined) answers 3. Language models are scaffolding, not foundations
They exist only to compensate for missing formalized knowledge. 4. Failure is a valid outcome
“Cannot decide” is preferable to unjustified confidence.
⸻
4. Architectural Overview
VERANTYX operates in explicit, auditable phases:
Phase 1 — Database-Only Reasoning
• Uses only rules.json and registered definitions
• No external model invocation
• Deterministic and noise-free
Phase 2 — Mining Fallback (Optional)
• Triggered only when Phase 1 yields insufficient evidence
• External models are queried with strict JSON-only prompts
• Results are sandboxed as provisional knowledge
Phase 3 — Audited Resolution
• Proven rules override mined rules
• Refutations override proofs
• Provisional answers are explicitly labeled
This structure ensures that expanding coverage does not degrade reliability.
⸻
5. Noise Control as a First-Class Concern
LLM-based mining is treated as a high-risk operation.
VERANTYX enforces:
• Strict JSON-only outputs
• No natural language explanations
• No overwriting of core rules
• Explicit provenance tracking
• Clear separation between verified and provisional knowledge
Mining exists to fill gaps — not to redefine truth.
⸻
6. Database-Centric Growth Model
VERANTYX improves by writing rules, not by retraining models.
Knowledge growth is:
• incremental
• inspectable
• reversible
• version-controlled
This makes reasoning evolution slower, but auditable.
⸻
7. The Role of Hardware in Contemporary AI
Modern AI systems implicitly assume:
• Large GPUs
• High VRAM availability
• Continuous training and retraining
• Infrastructure often inaccessible to individuals
This coupling between intelligence and hardware scale is rarely questioned.
VERANTYX explicitly questions it.
⸻
8. Hardware Efficiency as a Structural Consequence
VERANTYX does not aim to eliminate hardware usage.
Instead, it avoids mandatory dependence on large-scale compute.
By design, VERANTYX:
• avoids gradient-based training loops
• avoids persistent high-dimensional latent states
• relies on discrete rules and deterministic simulation
As a result:
• the system can be developed and experimented with in local environments
• GPU usage is optional rather than structurally required
• resource usage grows with database size, not model scale
This is not an optimization trick — it is a consequence of the reasoning model.
⸻
9. Why Individuals Historically Could Not Build AI
For many years, “building AI” effectively meant:
• training large neural networks
• acquiring massive datasets
• accessing expensive compute infrastructure
This created a structural barrier:
• individuals could not meaningfully experiment
• research directions followed compute availability
• architectural diversity collapsed
Many ideas were never tested — not because they were wrong, but because they were impractical.
⸻
10. A Concrete Trigger: Specialization Should Not Require Massive Training
A personal motivation behind VERANTYX is that specialization is often treated as a large-scale training problem, even when the core need is reliable, auditable decision-making.
In Japan, there have been real cases where local governments faced serious wildlife issues — including bear-related incidents — and responded by training image-based AI systems for bear identification.
Those efforts can be valuable, but they often implicitly require:
• large curated datasets,
• training pipelines,
• substantial compute,
• long operational cycles.
That pattern highlights a barrier:
specialized AI is often inaccessible unless you can run large training projects.
VERANTYX was partly motivated by the belief that specialization should also be possible through:
• curated rules,
• structured definitions,
• verified exception handling,
• and auditable failure modes,
so that individuals and small teams can build specialized systems locally — without needing “industrial-scale training” as the default prerequisite.
⸻
11. VERANTYX as a Counter-Design
VERANTYX was motivated by a simple question:
What if reasoning systems were constrained by ideas, not hardware?
By prioritizing:
• explicit rules over learned weights
• database growth over retraining
• auditability over benchmark performance
VERANTYX enables:
• local experimentation
• individual-driven system evolution
• architectural exploration without large infrastructure
⸻
12. Domain Safety Motivation: Finance and Healthcare
Another motivation is the need for systems that can operate under high-accuracy expectations, such as:
• finance,
• medicine,
• safety-critical decision support,
• compliance-heavy environments.
These areas often demand:
• traceability,
• clear reasoning boundaries,
• explicit assumption management,
• and an acceptable ability to say “cannot decide.”
VERANTYX aims to support this style of use by construction:
not by claiming perfect correctness, but by making errors and uncertainty legible and auditable.
⸻
13. Slowness as an Accepted Tradeoff
VERANTYX accepts:
• slower execution
• narrower domain coverage
• frequent “insufficient evidence” outcomes
in exchange for:
• transparency
• controllability
• independence from constant retraining
This is a deliberate value choice.
⸻
14. Relationship to LLMs (Clarified)
LLMs are used in VERANTYX because:
• they are widely available
• they are convenient gap-fillers
• the database is still incomplete
They are not used because they are ideal, controllable, or foundational.
As the database grows, reliance on external models is expected to decrease.
⸻
15. Reframing AI Development
In VERANTYX, AI development means:
• curating knowledge
• articulating rules
• discovering counterexamples
• managing assumptions
not:
• scaling parameters
• minimizing loss functions
• maximizing compute usage
⸻
16. Participation and Accessibility
A system that can be:
• understood
• modified
• broken
• repaired
by an individual is fundamentally different from one that cannot.
Lowering structural barriers changes:
• who can participate
• which ideas can be tested
• what kinds of failures are acceptable
⸻
17. Why VERANTYX v1 Was Closed
VERANTYX v1 was not open source because:
• core abstractions were unstable
• misuse would have produced misleading conclusions
• the system was not defensible as a public artifact
Opening it prematurely would have created noise, not progress.
⸻
18. Why VERANTYX Is Open Now
VERANTYX is open now because it is unfinished.
The system is stable enough to be:
• modified
• criticized
• extended
• broken
without collapsing its core philosophy.
Forks, incompatible extensions, and failed experiments are expected.
⸻
19. Expectations for Users
VERANTYX is not a product.
You are not expected to:
• trust it
• agree with it
• use it unchanged
You are encouraged to:
• inspect every rule
• question every assumption
• extend or replace the database
• change or remove components
Users are participants, not consumers.
⸻
20. Research Status
• Status: Experimental / Research Preview
• Stability: Medium
• Intended Use: Research, system design exploration
• Not intended for: Production or safety-critical systems
⸻
21. Long-Term Direction
The long-term direction is clear:
• reduced reliance on external models
• stronger internal databases
• fully auditable reasoning paths
• increased feasibility of specialized systems built locally
Whether this can scale remains an open question.
⸻
22. Closing Note
VERANTYX does not promise better answers.
It promises clearer reasons for failure.
This repository exists to explore whether that tradeoff is worthwhile — in public.
⸻
⸻
⸻
VERANTYX 研究ノート / 設計思想(統合版)
⸻
1. VERANTYX とは何か(何ではないか)
VERANTYX は、文章生成ではなく 推論そのものを制御・拘束・監査する ことを目的とした論理推論システムです。
LLM そのものでもなく、LLM ラッパーでもありません(設計上そうならないように作られています)。
現時点では内部 DB が未完成であるため、LLM を補助的に利用していますが、
これは 暫定的な工学的判断であり、設計思想の中心ではありません。
DB が成長するにつれ、外部モデルの重要性は低下し、
多くのタスクで不要になる可能性があるように設計されています。
⸻
2. 動機:確率的言語生成の限界
主流の AI は確率的生成に強く依存していますが、そこには:
• 推論過程の不可視性
• 誤りと創造性の混同
• 正答でも理由が誤っている可能性
• 監査不能な失敗
という構造的問題があります。
VERANTYX は、ルール・反例・必要条件検証による
明示的で検査可能な推論という別ルートを検証する試みです。
⸻
3. 基本思想
VERANTYX の原則: 1. 推論は検査可能であるべき 2. 正しさは条件付きである(証明・反証・証拠不足・仮回答を分ける) 3. LLM は足場であり基盤ではない 4. 判断不能は失敗ではない(不当な確信より優先)
⸻
4. アーキテクチャ概要
• Phase 1:DB のみ(ノイズゼロ/決定論)
• Phase 2:採掘(不足時のみ/JSON 限定)
• Phase 3:監査付き決定(証明>仮、反証>証明)
⸻
5. ノイズ制御
採掘は高リスク操作として扱い、
• JSON 限定
• 説明文禁止
• 上書き禁止
• 出所管理
• 検証済みと仮知識の分離
を徹底します。
採掘は「補完」であって「真理の定義」ではありません。
⸻
6. DB 中心の成長モデル
VERANTYX は 学習ではなく記述によって成長します。
知識の増加は:
• 段階的
• 検査可能
• 差し戻し可能
• バージョン管理可能
その代償として成長は遅くなりますが、監査性が守られます。
⸻
7. 現代 AI とハードウェア前提
巨大 GPU・高 VRAM・継続的な学習基盤が前提となった AI 開発は、
個人にとって参入障壁が高すぎました。
この「知性=計算資源」という結びつきは、あまり疑われてきませんでした。
VERANTYX はそこを疑います。
⸻
8. ハードウェア依存を前提にしない設計
VERANTYX は GPU を否定しませんが、
必須ともしません。
設計上、
• 勾配学習ループを必須としない
• 高次元潜在状態を恒常的に持たない
• 離散ルールと決定論シミュレーションを使う
ため、
• ローカルで開発・実験可能
• GPU は「あると便利」止まり
• リソース増加はモデルサイズではなく DB サイズに依存
という形になります。
これは最適化ではなく、推論モデルの帰結です。
⸻
9. なぜ個人は AI を作れなかったのか
長らく AI 開発は:
• 大規模学習
• 大量データ
• 高価な計算資源
を前提としてきました。
その結果、
個人はアイデアの検証すら難しい状況が続いていました。
⸻
10. 具体的な問題意識:特化は「大規模学習」だけである必要はない
VERANTYX の動機の一つは、
「特化型 AI を作る=大規模学習プロジェクト」になりがちな現実への違和感です。
日本では、熊(クマ)による問題が深刻な地域があり、
自治体レベルで熊識別のために熊画像を学習させる AI を構築したようなケースがありました。
それは価値のある取り組みですが、多くの場合そこには:
• データ収集
• 学習パイプライン
• 計算資源
• 運用コスト
といった「大規模前提」が付きまといます。
ここでの問題意識は:
特化はもっと簡単に、個人でも到達できるべきではないか?
VERANTYX は、特化を
• ルールの整備
• 定義の整備
• 例外処理(反例)の蓄積
• 監査可能な失敗設計
によって実現できる道を用意し、
必要なら外部モデルを補助にしながらでも、
個人がローカルで特化型システムを作れる状態を目指します。
⸻
11. VERANTYX の対抗設計
問いは単純です:
制約はハードウェアではなく、アイデアと知識構造であるべきでは?
VERANTYX は:
• 重みよりルール
• 学習より DB
• ベンチマークより監査性
を優先することで、
個人でも検証できる AI を取り戻そうとします。
⸻
12. 金融・医療など「正確性が要求される領域」への動機
もう一つの動機は、金融・医療など
誤りのコストが高い領域で使えるものを作りたいという意図です。
こうした領域では:
• 根拠追跡
• 仮定管理
• 境界の明確化
• 判断不能の許容
が必要になります。
VERANTYX は「完璧な正しさ」を主張するのではなく、
誤りや不確実性が見える形で残ることを重要視します。
⸻
13. 遅さを受け入れる理由
VERANTYX は:
• 遅さ
• 狭さ
• 判断不能の頻発
を受け入れます。
その代わりに:
• 透明性
• 制御可能性
• 学習依存からの独立
を得ます。
これは価値選択です。
⸻
14. LLM との関係
LLM は:
• 便利
• 入手可能
• 不足部分の穴埋めができる
ため使っています。
しかし理想的・制御可能・基盤的だからではありません。
DB が成長するほど、外部モデル依存は減らす前提です。
⸻
15. AI 開発の再定義
VERANTYX における AI 開発とは:
• ルールを書く
• 反例を集める
• 仮定を管理する
• 失敗を設計する
ことです。
パラメータ拡大や損失最小化ではありません。
⸻
16. 参加可能性
個人が
• 理解できる
• 壊せる
• 直せる
システムは、それだけで意味があります。
障壁が下がると、
試せるアイデアの種類が増えます。
⸻
17. v1 非公開の理由
抽象が未成熟で、誤解とノイズを生むだけだったからです。
⸻
18. 今オープンな理由
未完成だからこそ、フォークと失敗を前提に公開します。
⸻
19. ユーザへの期待
疑ってください。壊してください。拡張してください。
ユーザは消費者ではなく参加者です。
⸻
20. 研究ステータス
実験的です。安全重要用途には未対応です。
⸻
21. 将来像
DB が主役になります。
外部モデル依存は減ります。
そして特化型システムを個人がローカルで作れる方向へ寄せます。
⸻
22. 終わりに
VERANTYX は正解を約束しません。
失敗の理由を約束します。
このトレードオフが価値を持つかを、公開の場で検証します。
⸻
⸻
|