@dogfood-lab/study-swarm 0.0.0 → 0.5.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
package/CHANGELOG.md ADDED
@@ -0,0 +1,15 @@
1
+ # Changelog
2
+
3
+ All notable changes to this project are documented here. The format is based on [Keep a Changelog](https://keepachangelog.com/), and this project adheres to [Semantic Versioning](https://semver.org/).
4
+
5
+ ## [0.5.0] — 2026-06-02
6
+
7
+ ### Added
8
+
9
+ - Initial public release of the **study-swarm** methodology.
10
+ - `README.md` — the protocol in five steps, the family-different verification rationale (with citations), the proof (two decorrelated non-Claude families catching planted citation traps), and how it wires to prism-verify + role-os.
11
+ - `PROTOCOL.md` — the locked execution shape: the two-stage citation check, the halt table, the sourcing standard, and the architecture the protocol enables.
12
+ - `SECURITY.md`, MIT `LICENSE`, project logo.
13
+ - Landing page + Starlight handbook at <https://dogfood-lab.github.io/study-swarm/>.
14
+
15
+ [0.5.0]: https://github.com/dogfood-lab/study-swarm/releases/tag/v0.5.0
package/LICENSE ADDED
@@ -0,0 +1,21 @@
1
+ MIT License
2
+
3
+ Copyright (c) 2026 dogfood-lab
4
+
5
+ Permission is hereby granted, free of charge, to any person obtaining a copy
6
+ of this software and associated documentation files (the "Software"), to deal
7
+ in the Software without restriction, including without limitation the rights
8
+ to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
9
+ copies of the Software, and to permit persons to whom the Software is
10
+ furnished to do so, subject to the following conditions:
11
+
12
+ The above copyright notice and this permission notice shall be included in all
13
+ copies or substantial portions of the Software.
14
+
15
+ THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16
+ IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17
+ FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
18
+ AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19
+ LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
20
+ OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
21
+ SOFTWARE.
package/PROTOCOL.md ADDED
@@ -0,0 +1,108 @@
1
+ # The study-swarm protocol — locked execution shape
2
+
3
+ This is the executable reference. The narrative, the proof, and the research grounding are in [README.md](README.md).
4
+
5
+ > **The one-line guard:** no finding reaches Step 5 unverified. If you cannot verify — verifier down, no different family reachable, retrieval oracle unreachable — you HALT and escalate; you do not proceed. The protocol never lets a model grade its own homework, including the one running it.
6
+
7
+ ## When to invoke
8
+
9
+ Fire when ANY hold:
10
+
11
+ - A decision introduces a **new product layer** (not a fix, scope extension, or operational tuning).
12
+ - The decision is **qualitative** — "should we trust the model here," "explain or just do," "cap options," "retry or fall back."
13
+ - You're about to recommend a **single-axis** answer (deterministic-only / LLM-only) where the real answer is multi-axis (deterministic floor + LLM ceiling + verifier).
14
+ - An **adjacent domain** (compilers, SRE, databases, mixed-initiative HCI) has likely solved this.
15
+
16
+ Does NOT fire for: pure fixes; scope extensions of already-grounded work; operational tuning ("what number," not "what shape").
17
+
18
+ The cost of running it is one parallel dispatch and a few minutes of synthesis. The cost of skipping it is the failure this protocol exists to prevent: shrinking a design to a simpler shape out of an unexamined fear of AI advice, defending choices with "studies show…" and no studies named, and resting architecture on citations that don't exist.
19
+
20
+ ## Step 1 — Identify load-bearing design decisions
21
+
22
+ List the specific questions where empirical evidence would change the answer. Aim for 3–5. **Fewer is fine** when the decision is genuinely substantial — run with 1–2 agents; the decision-to-investigate governs invocation, the number of evidence-changing questions governs breadth. Do not manufacture questions to hit a count, and do not abort for being under three. More than ~6 → split into multiple passes.
23
+
24
+ ## Step 2 — Dispatch parallel research agents
25
+
26
+ One agent per question, dispatched **in parallel** (a single batch). Each agent's prompt MUST include:
27
+
28
+ - the context — what's being built, why this question matters;
29
+ - the question shape, scoped to **evidence**, not opinion;
30
+ - a demand for SPECIFIC findings: paper titles, authors, years, URLs, a one-sentence key finding per source;
31
+ - a word cap (typically 500–600);
32
+ - "prefer specificity over breadth — 6–8 well-sourced findings beat 20 vague gestures";
33
+ - a note to use web search / fetch.
34
+
35
+ Typical agent count: 3–5.
36
+
37
+ > Step 4 makes retrieval a **hard** requirement: a paper an agent "remembers" but cannot retrieve does not enter the dispatch. Existence is established by resolving the identifier, not by recall.
38
+
39
+ ## Step 3 — Synthesize into a "Research grounding" section
40
+
41
+ A dedicated section near the top of the design doc, before the architectural decisions. Each finding follows one template:
42
+
43
+ ```
44
+ N. **<one-sentence finding>.** <Authors> <year> (<paper title or arXiv:NNNN.NNNNN>). <Implication for the system being designed>.
45
+ ```
46
+
47
+ Example:
48
+
49
+ > 1. **Contrastive explanations with a predicted human foil improve independent decision-making.** Buçinca et al. 2024 (arXiv:2410.04253) — N=628 between-subjects. Implication: every recommendation carries a "you might think X; I'm recommending Y because…" frame.
50
+
51
+ The format does three things at once: states the finding, cites the source so it can be verified, and names the design implication so the link evidence→choice is visible.
52
+
53
+ ## Step 4 — External verification gate (family-different, reasoning-stripped)
54
+
55
+ Before any finding informs the design (Step 5), a verifier of a **different model family** from the synthesizing model, with the synthesizer's reasoning hidden, checks every citation. The Step 2 research agents are *inputs* — they produce citations; they are **not** verifiers of the synthesis. A separate family must check, or it's a model grading its own homework — the exact failure the protocol prescribes verifiers to prevent.
56
+
57
+ **Non-circular by construction:** the verifier adjudicates via a deterministic retrieval oracle (existence) plus a different-family lens (groundedness). It does not re-run this protocol and does not rely on anyone's recall.
58
+
59
+ ### Two-stage check, per citation
60
+
61
+ 1. **Existence / attribution — a retrieval oracle, not a parametric LLM.** Resolve the arXiv ID / DOI / URL and confirm the paper exists with the stated title, authors, and year. This stage **must retrieve** (fetch the source / arXiv / Crossref), never model memory — fabrication and misattribution rates are high enough (Walters & Wilder 2023) and 2025–2026 papers postdate model training, so a parametric check will false-flag real work as fabricated (Onweller et al. 2026). If retrieval is unavailable, apply the halt-and-escalate rule below.
62
+ 2. **Groundedness — finding matches source.** Confirm the one-sentence finding describes what the source actually claims (an NLI-style support check). Even strong models fail to fully support their own citations roughly half the time, so this is a distinct, necessary axis — not implied by existence.
63
+
64
+ ### Running it
65
+
66
+ The reference implementation is **`roleos verify-citations <dispatch>`** ([role-os](https://github.com/mcp-tool-shop-org/role-os)), which shells **[prism-verify](https://github.com/mcp-tool-shop-org/prism-verify)** (`prism verify --type citations`): family-different routing by construction, reasoning-stripped, a deterministic retrieval existence floor, a groundedness lens, and a signed receipt. By hand, the fallback is any non-same-family model run reasoning-stripped against the bare citation claims, plus resolving each identifier yourself.
67
+
68
+ **Ensemble — ≥ 3 decorrelated lenses,** counting the **retrieval oracle as one mechanism-diverse lens**: retrieval oracle + ≥ 2 different-family LLM lenses. Diversity of lenses, not raw count, is the load-bearing variable (Rajan 2025; Kim et al. 2025).
69
+
70
+ ### Halt conditions (scope is per-finding — other verified findings proceed)
71
+
72
+ | Verdict / condition | Action |
73
+ |---|---|
74
+ | **FABRICATED** | The finding is **dropped** — there is no real source to correct, so re-verification is not attempted. |
75
+ | **MISATTRIBUTED** | Correct the attribution and re-verify **once**; a second non-clean verdict drops the finding. |
76
+ | **CANNOT_CONFIRM** | The finding is **removed from the design connection AND surfaced to a human with a contrastive frame** — "you probably expected finding N citable; I left it out because the oracle couldn't confirm it — override with X." Never silently kept; reinstated only if a human confirms the source. |
77
+ | **Verifier or oracle UNAVAILABLE** | The dispatch **HALTS and escalates to a human.** Unavailability is NEVER read as "citations are fine" and NEVER read as fabrication. Proceeding without a completed verification is forbidden. |
78
+ | **No different family reachable** | The retrieval oracle (Stage 1) still runs — it is mechanism-diverse and needs no different family — and gates existence. The groundedness LLM lens (Stage 2) **halts-and-escalates** rather than running same-family. A same-family LLM is never substituted for the different-family check. |
79
+
80
+ ## Step 5 — Connect findings to architecture, not just cite them
81
+
82
+ The design's decision section references findings by number where they justify a choice; each load-bearing choice traces to ≥ 1 finding. Citations without connection are noise.
83
+
84
+ Example: *"Retry uses a fresh prompt without the previous output. (sycophancy mitigation, Kim 2025.)"* — the choice is annotated with the source and the reason, so a reader knows why the rule exists, not just that it does.
85
+
86
+ ## Sourcing standard
87
+
88
+ **A citation includes ALL of:** (1) author(s) — first author + "et al." inline is fine; (2) year; (3) paper title OR canonical identifier (arXiv:NNNN.NNNNN, DOI, RFC); (4) a direct URL to the source (not a summary or a social-media thread); (5) a one-sentence key finding in your own words.
89
+
90
+ **Not allowed:** "studies show…" / "research suggests…" / "it's well-established…" without naming the source; titles without authors or years; citations the research step did not actually surface.
91
+
92
+ ## The architecture this protocol enables
93
+
94
+ Across the designs it has grounded, the same shape recurs:
95
+
96
+ ```
97
+ System decides structure deterministically
98
+
99
+ Model writes within that structure
100
+
101
+ Verifier admits before output
102
+ ```
103
+
104
+ - **Deterministic floor** — the system makes the law-defining call; the model never does.
105
+ - **Model in the prose / prioritization role** — AI adds judgment, contextual explanation, and prioritization where it adds value.
106
+ - **Verifier as admission gate** — the verifier checks the output against the structure before admitting it; retries use a fresh context to avoid sycophancy drift.
107
+
108
+ Designs that touch model-facing behavior default to this shape unless evidence justifies a different one.
package/README.es.md ADDED
@@ -0,0 +1,77 @@
1
+ <p align="center">
2
+ <a href="README.ja.md">日本語</a> | <a href="README.zh.md">中文</a> | <a href="README.md">English</a> | <a href="README.fr.md">Français</a> | <a href="README.hi.md">हिन्दी</a> | <a href="README.it.md">Italiano</a> | <a href="README.pt-BR.md">Português (BR)</a>
3
+ </p>
4
+
5
+ <p align="center">
6
+ <img src="https://raw.githubusercontent.com/dogfood-lab/study-swarm/main/assets/study-swarm.png" alt="study-swarm" width="360">
7
+ </p>
8
+
9
+ <p align="center">
10
+ <a href="LICENSE"><img src="https://img.shields.io/badge/license-MIT-green" alt="MIT License"></a>
11
+ <a href="https://dogfood-lab.github.io/study-swarm/"><img src="https://img.shields.io/badge/handbook-live-purple" alt="Handbook"></a>
12
+ <img src="https://img.shields.io/badge/cited%20research-verified-1f6feb" alt="Cited research, verified">
13
+ </p>
14
+
15
+ **Fundamentar las decisiones de diseño en investigaciones citadas, y luego verificar las citas con un *modelo diferente* antes de que se convierta en algo definitivo.**
16
+
17
+ `study-swarm` es un protocolo, no una herramienta. Cuando se toma una decisión de diseño importante con un LLM (un nuevo nivel de producto, una elección de arquitectura, una decisión sobre si se debe confiar en el modelo), improvisar a partir de principios básicos da como resultado diseños obsoletos, y citar artículos de memoria da como resultado diseños que se basan en fuentes que no existen o que no dicen lo que se cree. `study-swarm` reemplaza ambas opciones: se envían agentes de investigación en paralelo, se exige la presentación de hallazgos específicos y se verifica cada cita a través de un **verificador externo de una familia de modelos diferente** antes de que se utilice para informar el diseño.
18
+
19
+ Aplica su propia medicina. El protocolo prescribe el uso de verificadores para proteger los resultados de los sistemas que ayuda a diseñar, por lo que lo aplica también a sí mismo. **Ningún modelo califica su propio trabajo, incluido el que ejecuta el protocolo.**
20
+
21
+ ## El protocolo en cinco pasos
22
+
23
+ 1. **Identificar** de 3 a 5 preguntas de diseño clave en las que la evidencia empírica cambiaría la respuesta.
24
+ 2. **Enviar** un agente de investigación por cada pregunta, en paralelo. Cada uno debe devolver títulos de artículos + autores + años + URL + un hallazgo de una sola frase; se prioriza la especificidad sobre la amplitud ("6-8 hallazgos bien documentados superan a 20 observaciones vagas").
25
+ 3. **Sintetizar** los hallazgos en una sección de *fundamentación de la investigación*: `N. **<hallazgo>.** <Autores> <año> (<arXiv/DOI>). <implicación para el diseño>.`
26
+ 4. **Verificar externamente** — una *familia de modelos diferente*, sin razonamiento, verifica cada cita en dos etapas: un **oráculo de recuperación** confirma que el artículo existe (nunca la memoria del modelo), y luego una lente de **fundamentación** confirma que el hallazgo coincide con la fuente. **Detener** si se detecta una falsificación o atribución incorrecta; **detener y escalar** si el verificador o el oráculo de recuperación no están disponibles (nunca interpretar la ausencia como "las citas son correctas").
27
+ 5. **Conectar** cada elección de arquitectura con un hallazgo mediante un número. Las citas sin una implicación para el diseño son ruido.
28
+
29
+ Los detalles completos y ejecutables (la tabla de detención, el estándar de fuentes, la regla de conjunto) se encuentran en **[PROTOCOL.md](PROTOCOL.md)**.
30
+
31
+ ## ¿Por qué una *familia diferente* y sin razonamiento?
32
+
33
+ Porque los modos de fallo están documentados, no son hipotéticos:
34
+
35
+ - **Los LLM no pueden verificar de manera confiable su propia salida.** Huang et al. 2023 ([arXiv:2310.01798](https://arxiv.org/abs/2310.01798)); Kambhampati et al. 2024 ([arXiv:2402.01817](https://arxiv.org/abs/2402.01817), LLM-Modulo); Stechly et al. 2024 ([arXiv:2402.08115](https://arxiv.org/abs/2402.08115)) — el verificador externo proporciona las ventajas; el contenido de autocrítica es inerte.
36
+ - **Los evaluadores de la misma familia se auto-favorecen.** Panickssery, Bowman & Feng 2024 ([arXiv:2404.13076](https://arxiv.org/abs/2404.13076)) — el auto-reconocimiento se correlaciona *linealmente* con la auto-preferencia, por lo que el cegamiento parcial no ayuda. Verga et al. 2024 ([arXiv:2404.18796](https://arxiv.org/abs/2404.18796), PoLL) — un panel de diferentes familias es menos sesgado a un costo aproximadamente 7 veces menor.
37
+ - **Las citas son donde los LLM mienten.** Walters & Wilder 2023 ([doi:10.1038/s41598-023-41032-5](https://doi.org/10.1038/s41598-023-41032-5)) — el 55% de las citas de GPT-3.5 / 18% de las citas de GPT-4 son falsas. Onweller et al. 2026 ([arXiv:2605.06635](https://arxiv.org/abs/2605.06635)) — los enlaces resuelven más del 94% de las veces, pero solo el 39-77% del contenido citado realmente respalda la afirmación. Por lo tanto, la existencia debe verificarse mediante la **recuperación, no la memoria**.
38
+ - **Ocultar el razonamiento del generador.** Khalifa et al. 2026 ([arXiv:2601.14691](https://arxiv.org/abs/2601.14691), "Gaming the Judge") — la manipulación del razonamiento en cadena infla los falsos positivos de un evaluador hasta en un 90% con acciones fijas. Turpin et al. 2023 ([arXiv:2305.04388](https://arxiv.org/abs/2305.04388)) — el razonamiento en cadena es una racionalización *a posteriori*. El verificador ve la afirmación de la cita sin adornos, nunca el "por qué la incluí".
39
+ - **La diversidad supera a la cantidad.** Rajan 2025 ([arXiv:2511.16708](https://arxiv.org/abs/2511.16708)) — cuatro verificadores con una correlación por pares de ρ ∈ [0.05, 0.25] superan a cualquiera de ellos mediante una cobertura submodular. Kim et al. 2025 ([arXiv:2506.07962](https://arxiv.org/abs/2506.07962)) — los errores de los LLM están *correlacionados*, por lo que la variable clave es la diversidad de las lentes, no la cantidad.
40
+
41
+ ## ¿Funciona realmente? (prueba)
42
+
43
+ Como prueba, el protocolo se ejecutó con sus propias citas. Dos familias no correlacionadas y diferentes a Claude — **Mistral** (`mistral-small:24b`) y **IBM Granite** (`granite4.1:30b`) — verificaron un conjunto de citas, sin razonamiento, con dos trampas ocultas:
44
+
45
+ | Trampa plantada | Mistral | IBM Granite | Verdad fundamental |
46
+ |---|---|---|---|
47
+ | El razonamiento en cadena atribuido a "Nakamura & Olsen" | no se detectó | **se detectó** (atribución incorrecta → en realidad Wei et al. 2022) | atribución incorrecta |
48
+ | un artículo fabricado con la afirmación de que "el 98% de los errores se eliminan, no se necesita un oráculo" | **caught** (fabricated) | **caught** (fabricated) | fabricado |
49
+
50
+ Ninguna de las dos familias detectó ambas trampas por sí sola, pero su **unión detectó 2/2**. Un solo evaluador habría aceptado la atribución incorrecta. Por separado, el oráculo de recuperación detectó dos *atribuciones incorrectas reales* en nuestros propios documentos de diseño (artículos citados con el primer autor incorrecto) que ningún LLM paramétrico podría haber detectado, y confirmó correctamente los artículos genuinos de 2026 que ambos LLM marcaron erróneamente como fabricados simplemente porque los artículos son posteriores a su entrenamiento. Ese último punto es la razón por la que la verificación de la existencia en el paso 4 **debe** ser un oráculo de recuperación, nunca un LLM.
51
+
52
+ Esa única ejecución es la tesis en miniatura: **lentes no correlacionadas + un oráculo de recuperación para la existencia superan a cualquier evaluador inteligente**.
53
+
54
+ ## Cómo está conectado
55
+
56
+ Puede ejecutar el protocolo manualmente: cualquier modelo de una familia diferente más la resolución de arXiv/DOI por sí mismo satisface el paso 4. Dos herramientas complementarias lo convierten en un solo comando:
57
+
58
+ - **[prism-verify](https://github.com/mcp-tool-shop-org/prism-verify)** — el verificador en tiempo de ejecución: enrutamiento diferenciado por familia, sin razonamiento superfluo, adjudicación con múltiples lentes, un umbral determinista para la existencia de referencias (arXiv → Crossref) y comprobantes firmados.
59
+ - **[role-os](https://github.com/mcp-tool-shop-org/role-os)** — proporciona `roleos verify-citations <dispatch>`, el programa que extrae las citas de un documento y las valida a través de prism.
60
+
61
+ ## Por qué funciona, en pocas palabras
62
+
63
+ **Eficiencia:** el campo avanza rápidamente; exigir estudios específicos y exhaustivos retrasa el lanzamiento de los diseños en 18 meses. **Funcionalidad:** la evidencia muestra lo que *falla*, no solo lo que funciona (las explicaciones pueden aumentar la dependencia excesiva de una IA *incorrecta* — Bansal et al. 2021). **Seguridad:** el entorno protegido por el verificador es la arquitectura que respalda la evidencia, y el protocolo la aplica a su propia salida. La verificación no es un ejercicio académico; es el rastro de la evidencia.
64
+
65
+ ## Seguridad
66
+
67
+ `study-swarm` es un repositorio de documentación: Markdown y un logotipo. No incluye código ejecutable ni instala nada de este repositorio. No accede a datos, no requiere permisos y no recopila datos de telemetría; no hay secretos ni credenciales en el código fuente. La metodología *describe* un flujo de trabajo que utiliza la recuperación web y la verificación basada en modelos, pero este repositorio no lo implementa ni lo ejecuta. Consulte [SECURITY.md](SECURITY.md).
68
+
69
+ ## Estado
70
+
71
+ Un protocolo funcional, verificado externamente por su propio mecanismo: una familia de modelos diferente verifica sus citas (vea la prueba anterior). Este repositorio es la referencia pública; [PROTOCOL.md](PROTOCOL.md) es la implementación ejecutable. Forma parte de la familia [dogfood-lab](https://github.com/dogfood-lab): métodos y ejemplos para construir en la era de la IA.
72
+
73
+ Con licencia MIT.
74
+
75
+ ---
76
+
77
+ <p align="center"><sub>Part of the <a href="https://github.com/dogfood-lab">dogfood-lab</a> family — methods &amp; showcases for building in the AI era. Built by <a href="https://mcp-tool-shop.github.io/">MCP Tool Shop</a>.</sub></p>
package/README.fr.md ADDED
@@ -0,0 +1,77 @@
1
+ <p align="center">
2
+ <a href="README.ja.md">日本語</a> | <a href="README.zh.md">中文</a> | <a href="README.es.md">Español</a> | <a href="README.md">English</a> | <a href="README.hi.md">हिन्दी</a> | <a href="README.it.md">Italiano</a> | <a href="README.pt-BR.md">Português (BR)</a>
3
+ </p>
4
+
5
+ <p align="center">
6
+ <img src="https://raw.githubusercontent.com/dogfood-lab/study-swarm/main/assets/study-swarm.png" alt="study-swarm" width="360">
7
+ </p>
8
+
9
+ <p align="center">
10
+ <a href="LICENSE"><img src="https://img.shields.io/badge/license-MIT-green" alt="MIT License"></a>
11
+ <a href="https://dogfood-lab.github.io/study-swarm/"><img src="https://img.shields.io/badge/handbook-live-purple" alt="Handbook"></a>
12
+ <img src="https://img.shields.io/badge/cited%20research-verified-1f6feb" alt="Cited research, verified">
13
+ </p>
14
+
15
+ **Ancrez les décisions de conception dans les recherches citées, puis vérifiez les citations à l’aide d’un *modèle* différent avant que quoi que ce soit ne devienne une référence.**
16
+
17
+ `study-swarm` est un protocole, pas un outil. Lorsque vous prenez une décision de conception importante avec un LLM (un nouveau niveau de produit, un choix d’architecture, une question du type « devons-nous faire confiance au modèle ici »), improviser à partir de principes de base conduit à des conceptions obsolètes, et citer des articles de mémoire conduit à des conceptions basées sur des sources qui n’existent pas ou qui ne disent pas ce que vous pensez. `study-swarm` remplace les deux : il lance des agents de recherche parallèles, exige des résultats spécifiques cités et soumet chaque citation à un **vérificateur externe d’une famille de modèles différente** avant qu’elle n’influence la conception.
18
+
19
+ Il applique sa propre méthode. Le protocole prescrit des enveloppes protégées par un vérificateur pour les systèmes qu’il aide à concevoir, il l’applique donc à lui-même. **Aucun modèle n’évalue son propre travail, y compris celui qui exécute le protocole.**
20
+
21
+ ## Le protocole en cinq étapes
22
+
23
+ 1. **Identifiez** 3 à 5 questions de conception essentielles auxquelles des preuves empiriques permettraient de modifier la réponse.
24
+ 2. **Lancez** un agent de recherche par question, en parallèle. Chacun doit renvoyer les titres des articles + les auteurs + les années + les URL + un résultat en une phrase — privilégiez la spécificité à l’étendue (« 6 à 8 résultats bien documentés sont plus efficaces que 20 observations vagues »).
25
+ 3. **Synthétisez** les résultats dans une section *Justification par la recherche* : `N. **<résultat>.** <Auteurs> <année> (<arXiv/DOI>). <implication pour la conception>.`
26
+ 4. **Vérifiez de manière externe** — une *famille de modèles différente*, sans raisonnement, vérifie chaque citation en deux étapes : un **oracle de récupération** confirme que l’article existe (jamais la mémoire du modèle), puis une **lentille de pertinence** confirme que le résultat correspond à la source. **Arrêtez** si la citation est fabriquée ou mal attribuée ; **arrêtez et escaladez** si le vérificateur ou l’oracle de récupération n’est pas disponible (ne considérez jamais l’absence comme une indication que les citations sont correctes).
27
+ 5. **Reliez** chaque choix architectural à un résultat par numéro. Les citations sans implication pour la conception sont du bruit.
28
+
29
+ Les détails complets et exécutables — le tableau d’arrêt, la norme de référencement, la règle d’ensemble — se trouvent dans **[PROTOCOL.md](PROTOCOL.md)**.
30
+
31
+ ## Pourquoi une *famille différente*, sans raisonnement ?
32
+
33
+ Parce que les modes d’échec sont documentés, et non hypothétiques :
34
+
35
+ - **Les LLM ne peuvent pas vérifier de manière fiable leurs propres résultats.** Huang et al. 2023 ([arXiv:2310.01798](https://arxiv.org/abs/2310.01798)) ; Kambhampati et al. 2024 ([arXiv:2402.01817](https://arxiv.org/abs/2402.01817), LLM-Modulo) ; Stechly et al. 2024 ([arXiv:2402.08115](https://arxiv.org/abs/2402.08115)) — le vérificateur externe apporte les améliorations ; le contenu d’auto-évaluation est inerte.
36
+ - **Les juges de la même famille ont une préférence pour eux-mêmes.** Panickssery, Bowman & Feng 2024 ([arXiv:2404.13076](https://arxiv.org/abs/2404.13076)) — l’auto-reconnaissance est corrélée *linéairement* avec l’auto-préférence, de sorte qu’un aveuglement partiel ne sert à rien. Verga et al. 2024 ([arXiv:2404.18796](https://arxiv.org/abs/2404.18796), PoLL) — un groupe de juges issus de familles différentes est moins biaisé, pour un coût environ 7 fois inférieur.
37
+ - **Les citations sont les points où les LLM mentent.** Walters & Wilder 2023 ([doi:10.1038/s41598-023-41032-5](https://doi.org/10.1038/s41598-023-41032-5)) — 55 % des citations de GPT-3.5 et 18 % de celles de GPT-4 sont fabriquées. Onweller et al. 2026 ([arXiv:2605.06635](https://arxiv.org/abs/2605.06635)) — les liens résolvent plus de 94 % du temps, mais seulement 39 à 77 % du contenu cité soutiennent réellement l’affirmation. Par conséquent, l’existence doit être vérifiée par **la récupération, et non par la mémorisation**.
38
+ - **Masquez le raisonnement du générateur.** Khalifa et al. 2026 ([arXiv:2601.14691](https://arxiv.org/abs/2601.14691), « Gaming the Judge ») — la simple manipulation de la chaîne de pensée augmente les faux positifs d’un juge jusqu’à 90 %, les actions étant maintenues fixes. Turpin et al. 2023 ([arXiv:2305.04388](https://arxiv.org/abs/2305.04388)) — la chaîne de pensée est une rationalisation a posteriori. Le vérificateur voit la simple affirmation de la citation, jamais la raison pour laquelle elle a été incluse.
39
+ - **La diversité est plus importante que la quantité.** Rajan 2025 ([arXiv:2511.16708](https://arxiv.org/abs/2511.16708)) — quatre vérificateurs avec une corrélation par paires ρ ∈ [0,05, 0,25] sont plus efficaces qu’un seul, grâce à une couverture sous-modulaire. Kim et al. 2025 ([arXiv:2506.07962](https://arxiv.org/abs/2506.07962)) — les erreurs des LLM sont *corrélées*, de sorte que la variable essentielle est la diversité des lentilles, et non la quantité brute.
40
+
41
+ ## Est-ce que cela fonctionne réellement ? (preuve)
42
+
43
+ À titre de test, le protocole a été appliqué à ses propres citations. Deux familles non-Claude non corrélées — **Mistral** (`mistral-small:24b`) et **IBM Granite** (`granite4.1:30b`) — ont vérifié un ensemble de citations, sans raisonnement, en y intégrant deux pièges aveugles :
44
+
45
+ | Piège implanté | Mistral | IBM Granite | Vérité terrain |
46
+ |---|---|---|---|
47
+ | Une chaîne de pensée attribuée à « Nakamura et Olsen » | non détecté | **détecté** (mal attribué → en réalité Wei et al. 2022) | mal attribué |
48
+ | un article fabriqué affirmant que « 98 % des erreurs sont éliminées, aucun oracle n’est nécessaire » | **caught** (fabricated) | **caught** (fabricated) | fabriqué |
49
+
50
+ Aucune des deux familles n’a détecté les deux pièges individuellement, mais leur **union a permis de détecter les 2/2**. Un seul juge aurait validé la mauvaise attribution. Par ailleurs, l’oracle de récupération a détecté deux *vrais* mauvaises attributions dans nos propres documents de conception (articles cités sous le mauvais premier auteur) que aucun LLM paramétrique n’aurait pu signaler, et il a correctement confirmé des articles authentiques de 2026 que les deux LLM ont faussement signalés comme étant fabriqués, simplement parce que les articles sont postérieurs à leur date d’entraînement. Ce dernier point est la raison pour laquelle la vérification de l’existence à l’étape 4 **doit** être effectuée par un oracle de récupération, et non par un LLM.
51
+
52
+ Cette seule expérience est la thèse en miniature : **des lentilles non corrélées + un oracle de récupération pour l’existence sont plus efficaces qu’un seul juge intelligent.**
53
+
54
+ ## Comment cela fonctionne
55
+
56
+ Vous pouvez exécuter le protocole manuellement — tout modèle d’une famille différente, associé à la résolution de l’arXiv/DOI, suffit pour l’étape 4. Deux outils frères permettent de le faire en une seule commande :
57
+
58
+ - **[prism-verify](https://github.com/mcp-tool-shop-org/prism-verify)** — le vérificateur d’exécution : routage différencié selon les familles, suppression des raisonnements, adjudication multi-lentilles, seuil de récupération déterministe (arXiv → Crossref) et reçus signés.
59
+ - **[role-os](https://github.com/mcp-tool-shop-org/role-os)** — fournit `roleos verify-citations <dispatch>`, l’outil qui extrait les citations d’un document et les soumet à la vérification par prism.
60
+
61
+ ## Pourquoi cela fonctionne, en un clin d’œil
62
+
63
+ **Efficacité** — le domaine évolue rapidement ; exiger des études spécifiques et approfondies retarde la mise en œuvre des conceptions de 18 mois. **Fonctionnalité** — les données montrent ce qui *ne fonctionne pas*, et pas seulement ce qui fonctionne (les explications peuvent entraîner une dépendance excessive à l’égard d’une IA *erronée* — Bansal et al. 2021). **Sécurité** — l’enveloppe protégée par le vérificateur est l’architecture que les données étayent, et le protocole l’applique à ses propres résultats. La recherche de sources n’est pas un exercice académique ; c’est la chaîne de preuves.
64
+
65
+ ## Sécurité
66
+
67
+ `study-swarm` est un dépôt de documentation : Markdown et un logo. Il ne contient aucun code exécutable et n’installe rien à partir de ce dépôt. Il n’accède à aucune donnée, ne nécessite aucune autorisation et ne collecte aucune télémétrie ; il n’y a pas de secrets ou d’identifiants dans le code source. La méthodologie *décrit* un flux de travail qui utilise la récupération sur le web et la vérification basée sur des modèles, mais ce dépôt ne l’implémente ni ne l’exécute. Voir [SECURITY.md](SECURITY.md).
68
+
69
+ ## État
70
+
71
+ Un protocole fonctionnel, vérifié en externe par ses propres mécanismes : une famille de modèles différente vérifie ses citations (voir la preuve ci-dessus). Ce dépôt est la référence publique ; [PROTOCOL.md](PROTOCOL.md) est la forme exécutable. Fait partie de la famille [dogfood-lab](https://github.com/dogfood-lab) : méthodes et exemples pour construire dans l’ère de l’IA.
72
+
73
+ Licence MIT.
74
+
75
+ ---
76
+
77
+ <p align="center"><sub>Part of the <a href="https://github.com/dogfood-lab">dogfood-lab</a> family — methods &amp; showcases for building in the AI era. Built by <a href="https://mcp-tool-shop.github.io/">MCP Tool Shop</a>.</sub></p>
package/README.hi.md ADDED
@@ -0,0 +1,77 @@
1
+ <p align="center">
2
+ <a href="README.ja.md">日本語</a> | <a href="README.zh.md">中文</a> | <a href="README.es.md">Español</a> | <a href="README.fr.md">Français</a> | <a href="README.md">English</a> | <a href="README.it.md">Italiano</a> | <a href="README.pt-BR.md">Português (BR)</a>
3
+ </p>
4
+
5
+ <p align="center">
6
+ <img src="https://raw.githubusercontent.com/dogfood-lab/study-swarm/main/assets/study-swarm.png" alt="study-swarm" width="360">
7
+ </p>
8
+
9
+ <p align="center">
10
+ <a href="LICENSE"><img src="https://img.shields.io/badge/license-MIT-green" alt="MIT License"></a>
11
+ <a href="https://dogfood-lab.github.io/study-swarm/"><img src="https://img.shields.io/badge/handbook-live-purple" alt="Handbook"></a>
12
+ <img src="https://img.shields.io/badge/cited%20research-verified-1f6feb" alt="Cited research, verified">
13
+ </p>
14
+
15
+ **आधारभूत शोध पर आधारित डिज़ाइन निर्णय लें — फिर किसी भी चीज़ को आधिकारिक बनाने से पहले एक *अलग* मॉडल परिवार के साथ उद्धरणों को सत्यापित करें।**
16
+
17
+ `study-swarm` एक प्रोटोकॉल है, उपकरण नहीं। जब आप किसी एलएलएम के साथ एक महत्वपूर्ण डिज़ाइन निर्णय ले रहे हों — एक नया उत्पाद परत, एक आर्किटेक्चर विकल्प, एक "क्या हमें यहां मॉडल पर भरोसा करना चाहिए" — तो बुनियादी सिद्धांतों से तात्कालिक रूप से काम करने से ऐसे डिज़ाइन सामने आते हैं जो पुराने हैं, और स्मृति से कागजात का हवाला देने से ऐसे डिज़ाइन सामने आते हैं जो उन स्रोतों पर आधारित होते हैं जो मौजूद नहीं हैं या जो आप जो सोचते हैं, वह नहीं कहते हैं। study-swarm दोनों को बदल देता है: समानांतर अनुसंधान एजेंटों को भेजें, विशिष्ट उद्धृत निष्कर्षों की मांग करें, और डिज़ाइन को सूचित करने से पहले एक **एक अलग मॉडल परिवार के बाहरी सत्यापनकर्ता** के माध्यम से प्रत्येक उद्धरण को जांचें।
18
+
19
+ यह अपनी ही दवा का उपयोग करता है। प्रोटोकॉल उन प्रणालियों के लिए सत्यापनकर्ता-संरक्षित एन्वलप निर्धारित करता है जिन्हें यह डिज़ाइन करने में मदद करता है — इसलिए यह स्वयं पर भी इसे चलाता है। **कोई भी मॉडल अपने स्वयं के गृहकार्य का मूल्यांकन नहीं करता है, जिसमें प्रोटोकॉल चलाने वाला मॉडल भी शामिल है।**
20
+
21
+ ## प्रोटोकॉल पाँच चरणों में
22
+
23
+ 1. **पहचानें** 3-5 भार-असर वाले डिज़ाइन प्रश्न जहाँ अनुभवजन्य प्रमाण उत्तर को बदल देंगे।
24
+ 2. **भेजें** प्रत्येक प्रश्न के लिए एक अनुसंधान एजेंट, समानांतर में। प्रत्येक को पेपर शीर्षक + लेखक + वर्ष + यूआरएल + एक वाक्य निष्कर्ष वापस करना होगा — व्यापकता से अधिक विशिष्टता ("6-8 अच्छी तरह से स्रोतों से प्राप्त निष्कर्ष 20 अस्पष्ट संकेतों से बेहतर हैं")।
25
+ 3. **संश्लेषित करें** निष्कर्षों को एक *अनुसंधान आधार* अनुभाग में: `N. **<निष्कर्ष>.** <लेखक> <वर्ष> (<arXiv/DOI>). <डिज़ाइन निहितार्थ>.`
26
+ 4. **बाहरी रूप से सत्यापित करें** — एक *अलग मॉडल परिवार*, तर्क-मुक्त, दो चरणों में प्रत्येक उद्धरण की जांच करता है: एक **पुनर्प्राप्ति ओरेकल** पुष्टि करता है कि पेपर मौजूद है (कभी भी मॉडल की स्मृति नहीं), फिर एक **आधारितता** लेंस पुष्टि करता है कि निष्कर्ष स्रोत से मेल खाता है। **यदि कोई नकली/गलत उद्धरण पाया जाता है, तो रोकें; यदि सत्यापनकर्ता या पुनर्प्राप्ति ओरेकल अनुपलब्ध है, तो रोकें और आगे बढ़ाएं** (अनुपस्थिति को "उद्धरण ठीक हैं" के रूप में कभी न मानें)।
27
+ 5. **प्रत्येक वास्तुशिल्प विकल्प को संख्या द्वारा एक निष्कर्ष से जोड़ें।** डिज़ाइन निहितार्थ के बिना उद्धरण शोर हैं।
28
+
29
+ पूरी निष्पादन योग्य जानकारी — रोक तालिका, सोर्सिंग मानक, एन्सेम्बल नियम — **[PROTOCOL.md](PROTOCOL.md)** में है।
30
+
31
+ ## एक *अलग* परिवार, तर्क-मुक्त क्यों?
32
+
33
+ क्योंकि विफलता के तरीके प्रलेखित हैं, काल्पनिक नहीं:
34
+
35
+ - **एलएलएम अपने स्वयं के आउटपुट को विश्वसनीय रूप से सत्यापित नहीं कर सकते हैं।** हुआंग एट अल. 2023 ([arXiv:2310.01798](https://arxiv.org/abs/2310.01798)); कंबंपती एट अल. 2024 ([arXiv:2402.01817](https://arxiv.org/abs/2402.01817), एलएलएम-मॉड्यूलो); स्टेचली एट अल. 2024 ([arXiv:2402.08115](https://arxiv.org/abs/2402.08115)) — बाहरी सत्यापनकर्ता लाभ प्रदान करता है; आत्म-आलोचना सामग्री निष्क्रिय है।
36
+ - **समान-परिवार के न्यायाधीश स्वयं को पसंद करते हैं।** पैनिकसेरी, बोमन और फेंग 2024 ([arXiv:2404.13076](https://arxiv.org/abs/2404.13076)) — आत्म-पहचान *रैखिक रूप से* आत्म-पसंद से संबंधित है, इसलिए आंशिक अंधापन मदद नहीं करता है। वर्गा एट अल. 2024 ([arXiv:2404.18796](https://arxiv.org/abs/2404.18796), पोएलएल) — अलग-अलग परिवारों में एक पैनल लगभग 7 गुना कम लागत पर कम पक्षपाती है।
37
+ - **उद्धरण वे स्थान हैं जहाँ एलएलएम झूठ बोलते हैं।** वाल्टर्स और वाइल्डर 2023 ([doi:10.1038/s41598-023-41032-5](https://doi.org/10.1038/s41598-023-41032-5)) — जीपीटी-3.5 के 55% / जीपीटी-4 के 18% उद्धरण नकली हैं। ऑनवेलेर एट अल. 2026 ([arXiv:2605.06635](https://arxiv.org/abs/2605.06635)) — लिंक >94% समय तक हल होते हैं, फिर भी केवल 39-77% उद्धृत सामग्री वास्तव में दावे का समर्थन करती है। इसलिए अस्तित्व की जांच **पुनर्प्राप्ति द्वारा की जानी चाहिए, न कि स्मरण द्वारा।**
38
+ - **जनरेटर के तर्क को छिपाएं।** खलीफा एट अल. 2026 ([arXiv:2601.14691](https://arxiv.org/abs/2601.14691), "जज को धोखा देना") — अकेले हेरफेर किए गए चेन-ऑफ-थॉट एक न्यायाधीश के झूठे-सकारात्मकों को 90% तक बढ़ा देते हैं, जबकि क्रियाएं स्थिर रहती हैं। टर्पिन एट अल. 2023 ([arXiv:2305.04388](https://arxiv.org/abs/2305.04388)) — सीओटी पोस्ट-हॉक तर्कसंगतता है। सत्यापनकर्ता केवल नंगे उद्धरण दावे को देखता है, कभी भी "मैंने इसे क्यों शामिल किया" नहीं।
39
+ - **विविधता संख्या से बेहतर है।** राजन 2025 ([arXiv:2511.16708](https://arxiv.org/abs/2511.16708)) — जोड़ीदार सहसंबंध ρ ∈ [0.05, 0.25] पर चार सत्यापनकर्ता उपमॉड्यूलर कवरेज के माध्यम से किसी भी एकल सत्यापनकर्ता से बेहतर हैं। किम एट अल. 2025 ([arXiv:2506.07962](https://arxiv.org/abs/2506.07962)) — एलएलएम त्रुटियां *संबंधित* हैं, इसलिए भार-असर चर लेंस विविधता है, न कि कच्ची संख्या।
40
+
41
+ ## क्या यह वास्तव में काम करता है? (प्रमाण)
42
+
43
+ एक परीक्षण के रूप में, प्रोटोकॉल को अपने स्वयं के उद्धरणों के विरुद्ध चलाया गया था। दो असंबंधित गैर-क्लाउड परिवार — **मिस्ट्रल** (`mistral-small:24b`) और **आईबीएम ग्रेनाइट** (`granite4.1:30b`) — ने एक उद्धरण सेट की जांच की, तर्क-मुक्त, दो अंधे जाल के साथ:
44
+
45
+ | रोपा गया जाल | मिस्ट्रल | आईबीएम ग्रेनाइट | सत्य |
46
+ |---|---|---|---|
47
+ | "नकामुरा और ओल्सन" को जिम्मेदार चेन-ऑफ-थॉट प्रॉम्प्टिंग | छोड़ दिया | **पकड़ा गया** (गलत रूप से जिम्मेदार ठहराया गया → वास्तव में वेई एट अल. 2022) | गलत रूप से जिम्मेदार ठहराया गया |
48
+ | एक नकली "98% त्रुटियों को हटा दिया गया, ओरेकल की आवश्यकता नहीं है" पेपर | **caught** (fabricated) | **caught** (fabricated) | नकली |
49
+
50
+ किसी भी परिवार ने अकेले दोनों जाल नहीं पकड़े — लेकिन उनके **संघ ने 2/2 जाल पकड़े।** एक एकल न्यायाधीश गलत उद्धरण को जारी कर देगा। अलग से, पुनर्प्राप्ति ओरेकल ने हमारे अपने डिज़ाइन दस्तावेज़ों में दो *वास्तविक* गलत उद्धरणों को पकड़ा (गलत पहले लेखक के तहत उद्धृत पेपर) जिन्हें किसी भी पैरामीट्रिक एलएलएम द्वारा चिह्नित नहीं किया जा सकता था — और इसने सही ढंग से वास्तविक 2026 पेपरों की पुष्टि की, जिन्हें दोनों एलएलएम ने केवल इसलिए नकली के रूप में गलत तरीके से चिह्नित किया क्योंकि पेपर उनके प्रशिक्षण के बाद के हैं। वह अंतिम बिंदु पूरी तरह से कारण है कि चरण 4 में अस्तित्व की जांच **एक पुनर्प्राप्ति ओरेकल होनी चाहिए, कभी भी एलएलएम नहीं।**
51
+
52
+ यह एकल रन लघु रूप में थीसिस है: **असंबंधित लेंस + अस्तित्व के लिए एक पुनर्प्राप्ति ओरेकल किसी भी एक स्मार्ट न्यायाधीश से बेहतर है।**
53
+
54
+ ## यह कैसे जुड़ा हुआ है
55
+
56
+ आप प्रोटोकॉल को मैन्युअल रूप से चला सकते हैं — एक अलग-परिवार मॉडल और स्वयं द्वारा arXiv/DOI को हल करना चरण 4 को संतुष्ट करता है। दो भाई उपकरण इसे एक कमांड बनाते हैं:
57
+
58
+ - **[प्रिज्म-वेरीफाई](https://github.com/mcp-tool-shop-org/prism-verify)** — रनटाइम सत्यापनकर्ता: परिवार-विशिष्ट रूटिंग, तर्क-मुक्त, बहु-लेंस निर्णय, एक नियतात्मक पुनर्प्राप्ति अस्तित्व तल (arXiv → क्रॉसरेफ), और हस्ताक्षरित रसीदें।
59
+ - **[रोल-ओएस](https://github.com/mcp-tool-shop-org/role-os)** — `रोलओएस वेरीफाई-कोटेशन्स <डिस्पैच>` प्रदान करता है, जो एक ऐसा उपकरण है जो किसी डिस्पैच की कोटेशन्स को निकालता है और उन्हें प्रिज्म के माध्यम से संसाधित करता है।
60
+
61
+ ## यह कैसे काम करता है, संक्षेप में
62
+
63
+ **वर्तमान** — यह क्षेत्र तेजी से आगे बढ़ रहा है; विशिष्ट, वर्षों तक चलने वाले अध्ययनों की मांग करने से डिज़ाइन 18 महीने पीछे रह जाते हैं। **कार्यात्मक** — सबूत दिखाते हैं कि क्या *असफल* होता है, न कि केवल क्या काम करता है (व्याख्याएँ *गलत* एआई पर अत्यधिक निर्भरता बढ़ा सकती हैं — बंसल एट अल. 2021)। **सुरक्षित** — सत्यापनकर्ता-संरक्षित क्षेत्र वह आर्किटेक्चर है जिसका सबूत समर्थन करता है, और प्रोटोकॉल इसे अपने आउटपुट पर लागू करता है। स्रोत खोजना कोई अकादमिक नाटक नहीं है; यह सबूतों का मार्ग है।
64
+
65
+ ## सुरक्षा
66
+
67
+ `स्टडी-स्वार्म` एक दस्तावेज़ भंडार है — मार्कडाउन और एक लोगो। यह कोई निष्पादन योग्य कोड नहीं भेजता है और इस भंडार से कुछ भी स्थापित नहीं करता है। यह किसी भी डेटा को नहीं छूता है, किसी भी अनुमति की आवश्यकता नहीं होती है, और कोई टेलीमेट्री एकत्र नहीं करता है; स्रोत में कोई गुप्त जानकारी या क्रेडेंशियल नहीं हैं। कार्यप्रणाली *वर्णित* करती है कि एक वर्कफ़्लो जो वेब पुनर्प्राप्ति और मॉडल-आधारित सत्यापन का उपयोग करता है, लेकिन यह भंडार इसे लागू या चला नहीं रहा है। [सुरक्षा.एमडी](SECURITY.md) देखें।
68
+
69
+ ## स्थिति
70
+
71
+ एक कार्यात्मक प्रोटोकॉल, जिसे स्वयं अपनी मशीनरी द्वारा बाहरी रूप से सत्यापित किया गया है — एक अलग मॉडल परिवार इसकी कोटेशन्स की जांच करता है (ऊपर दिए गए प्रमाण देखें)। यह भंडार सार्वजनिक संदर्भ है; [प्रोटोकॉल.एमडी](PROTOCOL.md) निष्पादन योग्य रूप है। [डॉगफूड-लैब](https://github.com/dogfood-lab) परिवार का हिस्सा — एआई युग में निर्माण के लिए विधियाँ और प्रदर्शन।
72
+
73
+ एमआईटी लाइसेंस प्राप्त।
74
+
75
+ ---
76
+
77
+ <p align="center"><sub>Part of the <a href="https://github.com/dogfood-lab">dogfood-lab</a> family — methods &amp; showcases for building in the AI era. Built by <a href="https://mcp-tool-shop.github.io/">MCP Tool Shop</a>.</sub></p>
package/README.it.md ADDED
@@ -0,0 +1,77 @@
1
+ <p align="center">
2
+ <a href="README.ja.md">日本語</a> | <a href="README.zh.md">中文</a> | <a href="README.es.md">Español</a> | <a href="README.fr.md">Français</a> | <a href="README.hi.md">हिन्दी</a> | <a href="README.md">English</a> | <a href="README.pt-BR.md">Português (BR)</a>
3
+ </p>
4
+
5
+ <p align="center">
6
+ <img src="https://raw.githubusercontent.com/dogfood-lab/study-swarm/main/assets/study-swarm.png" alt="study-swarm" width="360">
7
+ </p>
8
+
9
+ <p align="center">
10
+ <a href="LICENSE"><img src="https://img.shields.io/badge/license-MIT-green" alt="MIT License"></a>
11
+ <a href="https://dogfood-lab.github.io/study-swarm/"><img src="https://img.shields.io/badge/handbook-live-purple" alt="Handbook"></a>
12
+ <img src="https://img.shields.io/badge/cited%20research-verified-1f6feb" alt="Cited research, verified">
13
+ </p>
14
+
15
+ **Fondi le decisioni di progettazione su ricerche citate, quindi verifica le citazioni con un *modello* diverso prima che diventino parte integrante del progetto.**
16
+
17
+ `study-swarm` è un protocollo, non uno strumento. Quando si prende una decisione di progettazione importante con un LLM (un nuovo livello di prodotto, una scelta architettonica, una decisione del tipo "dovremmo fidarci del modello in questo caso"), improvvisare partendo da principi fondamentali porta a progetti obsoleti e citare articoli a memoria porta a progetti basati su fonti inesistenti o che non dicono ciò che si pensa. `study-swarm` sostituisce entrambi: invia agenti di ricerca in parallelo, richiede risultati specifici dalle fonti citate e sottopone ogni citazione a un **verificatore esterno di un modello diverso** prima che influenzi la progettazione.
18
+
19
+ Applica la propria "medicina". Il protocollo prevede l'utilizzo di verificatori per proteggere i sistemi che aiuta a progettare, quindi lo applica anche a se stesso. **Nessun modello valuta il proprio lavoro, incluso quello che esegue il protocollo.**
20
+
21
+ ## Il protocollo in cinque passaggi
22
+
23
+ 1. **Identificare** 3-5 domande di progettazione fondamentali, in cui prove empiriche cambierebbero la risposta.
24
+ 2. **Inviare** un agente di ricerca per ogni domanda, in parallelo. Ognuno deve restituire titoli di articoli + autori + anni + URL + un risultato in una frase: dare priorità alla specificità rispetto all'ampiezza ("6-8 risultati ben documentati sono meglio di 20 affermazioni vaghe").
25
+ 3. **Sintetizzare** i risultati in una sezione "Fondamento della ricerca": `N. **<risultato>.** <Autori> <anno> (<arXiv/DOI>). <implicazione per la progettazione>.`
26
+ 4. **Verificare esternamente** — un *modello diverso*, senza ragionamento, controlla ogni citazione in due fasi: un **oracolo di recupero** conferma che l'articolo esiste (mai dalla memoria del modello), quindi una "lente di fondamento" conferma che il risultato corrisponde alla fonte. **Interrompere** in caso di fabbricazioni/attribuzioni errate; **interrompere e segnalare** se il verificatore o l'oracolo di recupero non sono disponibili (non interpretare mai l'assenza come "le citazioni sono corrette").
27
+ 5. **Collegare** ogni scelta architettonica a un risultato specifico tramite numero. Le citazioni senza un'implicazione per la progettazione sono rumore.
28
+
29
+ I dettagli completi e eseguibili (la tabella di interruzione, lo standard di riferimento, la regola di insieme) sono disponibili in **[PROTOCOL.md](PROTOCOL.md)**.
30
+
31
+ ## Perché un *modello diverso*, senza ragionamento?
32
+
33
+ Perché le modalità di errore sono documentate, non ipotetiche:
34
+
35
+ - **Gli LLM non possono verificare in modo affidabile i propri risultati.** Huang et al. 2023 ([arXiv:2310.01798](https://arxiv.org/abs/2310.01798)); Kambhampati et al. 2024 ([arXiv:2402.01817](https://arxiv.org/abs/2402.01817), LLM-Modulo); Stechly et al. 2024 ([arXiv:2402.08115](https://arxiv.org/abs/2402.08115)) — il verificatore esterno offre i vantaggi; il contenuto di autocritica è inerte.
36
+ - **I giudici della stessa famiglia tendono ad auto-preferirsi.** Panickssery, Bowman & Feng 2024 ([arXiv:2404.13076](https://arxiv.org/abs/2404.13076)) — l'auto-riconoscimento è correlato *linearmente* con l'auto-preferenza, quindi un'occlusione parziale non aiuta. Verga et al. 2024 ([arXiv:2404.18796](https://arxiv.org/abs/2404.18796), PoLL) — un gruppo di esperti provenienti da famiglie diverse è meno influenzato a un costo inferiore di circa 7 volte.
37
+ - **Le citazioni sono il punto in cui gli LLM mentono.** Walters & Wilder 2023 ([doi:10.1038/s41598-023-41032-5](https://doi.org/10.1038/s41598-023-41032-5)) — il 55% delle citazioni di GPT-3.5 / il 18% di GPT-4 sono fabbricate. Onweller et al. 2026 ([arXiv:2605.06635](https://arxiv.org/abs/2605.06635)) — i collegamenti risolvono il >94% delle volte, ma solo il 39-77% del contenuto citato supporta effettivamente l'affermazione. Pertanto, l'esistenza deve essere verificata tramite **recupero, non richiamo**.
38
+ - **Nascondere il ragionamento del generatore.** Khalifa et al. 2026 ([arXiv:2601.14691](https://arxiv.org/abs/2601.14691), "Gaming the Judge") — la sola manipolazione della catena di pensiero aumenta i falsi positivi di un giudice fino al 90% con azioni mantenute fisse. Turpin et al. 2023 ([arXiv:2305.04388](https://arxiv.org/abs/2305.04388)) — la catena di pensiero è una razionalizzazione post-hoc. Il verificatore vede solo la citazione, mai il "perché l'ho inclusa".
39
+ - **La diversità è più importante della quantità.** Rajan 2025 ([arXiv:2511.16708](https://arxiv.org/abs/2511.16708)) — quattro verificatori con correlazione a coppie ρ ∈ [0.05, 0.25] superano qualsiasi singolo verificatore tramite copertura submodulare. Kim et al. 2025 ([arXiv:2506.07962](https://arxiv.org/abs/2506.07962)) — gli errori degli LLM sono *correlati*, quindi la variabile più importante è la diversità delle "lenti", non la quantità.
40
+
41
+ ## Funziona davvero? (prova)
42
+
43
+ Come test, il protocollo è stato eseguito sulle proprie citazioni. Due famiglie non correlate di Claude — **Mistral** (`mistral-small:24b`) e **IBM Granite** (`granite4.1:30b`) — hanno controllato un insieme di citazioni, senza ragionamento, con due "trappole" nascoste:
44
+
45
+ | Trappola preparata | Mistral | IBM Granite | Verità |
46
+ |---|---|---|---|
47
+ | Il ragionamento della catena di pensiero è attribuito a "Nakamura & Olsen" | mancato | **rilevato** (attribuzione errata → in realtà Wei et al. 2022) | attribuzione errata |
48
+ | un articolo fabbricato con la dicitura "98% degli errori eliminati, non è necessario alcun oracolo" | **caught** (fabricated) | **caught** (fabricated) | fabbricato |
49
+
50
+ Nessuna delle due famiglie ha rilevato entrambe le trappole da sola, ma la loro **unione ha rilevato 2/2**. Un singolo giudice avrebbe accettato l'attribuzione errata. Separatamente, l'oracolo di recupero ha rilevato due *vere* attribuzioni errate nei nostri documenti di progettazione (articoli citati con il primo autore sbagliato) che nessun LLM parametrico avrebbe potuto segnalare, e ha confermato correttamente gli articoli genuini del 2026 che entrambi gli LLM hanno erroneamente segnalato come fabbricati semplicemente perché gli articoli sono successivi alla loro data di addestramento. Quest'ultimo punto è la ragione principale per cui il controllo dell'esistenza nel passaggio 4 **deve** essere un oracolo di recupero, mai un LLM.
51
+
52
+ Questa singola esecuzione è la tesi in miniatura: **"lenti" non correlate + un oracolo di recupero per l'esistenza superano qualsiasi singolo giudice esperto.**
53
+
54
+ ## Come è strutturato
55
+
56
+ È possibile eseguire il protocollo manualmente: qualsiasi modello di famiglia diversa più la risoluzione di arXiv/DOI da parte dell'utente soddisfa il passaggio 4. Due strumenti complementari lo rendono un unico comando:
57
+
58
+ - **[prism-verify](https://github.com/mcp-tool-shop-org/prism-verify)** — il verificatore in fase di esecuzione: instradamento differenziato per famiglia, ragionamento semplificato, arbitraggio multi-lente, un limite inferiore deterministico per il recupero dell'esistenza (arXiv → Crossref) e ricevute firmate.
59
+ - **[role-os](https://github.com/mcp-tool-shop-org/role-os)** — fornisce `roleos verify-citations <dispatch>`, lo strumento che estrae le citazioni di un documento e le elabora tramite prism.
60
+
61
+ ## In sintesi, perché funziona
62
+
63
+ **Efficienza** — il settore è in rapida evoluzione; richiedere studi specifici e approfonditi rallenta lo sviluppo dei progetti di 18 mesi. **Funzionalità** — i dati mostrano cosa *non* funziona, non solo cosa funziona (le spiegazioni possono portare a un'eccessiva dipendenza da un'IA *errata* — Bansal et al. 2021). **Sicurezza** — l'ambiente protetto dal verificatore è l'architettura supportata dai dati, e il protocollo la applica ai propri risultati. La verifica non è un esercizio accademico; è la traccia dei dati.
64
+
65
+ ## Sicurezza
66
+
67
+ `study-swarm` è un repository di documentazione: contiene file Markdown e un logo. Non include codice eseguibile e non installa nulla da questo repository. Non accede a dati, non richiede autorizzazioni e non raccoglie dati di telemetria; non ci sono segreti o credenziali nel codice sorgente. La metodologia *descrive* un flusso di lavoro che utilizza il recupero di informazioni dal web e la verifica basata su modelli, ma questo repository non lo implementa né lo esegue. Consultare [SECURITY.md](SECURITY.md).
68
+
69
+ ## Stato
70
+
71
+ Un protocollo funzionante, verificato esternamente dai propri strumenti: una famiglia di modelli diversa verifica le sue citazioni (vedere la prova sopra). Questo repository è il riferimento pubblico; [PROTOCOL.md](PROTOCOL.md) è la forma eseguibile. Parte della famiglia [dogfood-lab](https://github.com/dogfood-lab): metodi e esempi per lo sviluppo nell'era dell'IA.
72
+
73
+ Licenza MIT.
74
+
75
+ ---
76
+
77
+ <p align="center"><sub>Part of the <a href="https://github.com/dogfood-lab">dogfood-lab</a> family — methods &amp; showcases for building in the AI era. Built by <a href="https://mcp-tool-shop.github.io/">MCP Tool Shop</a>.</sub></p>
package/README.ja.md ADDED
@@ -0,0 +1,77 @@
1
+ <p align="center">
2
+ <a href="README.md">English</a> | <a href="README.zh.md">中文</a> | <a href="README.es.md">Español</a> | <a href="README.fr.md">Français</a> | <a href="README.hi.md">हिन्दी</a> | <a href="README.it.md">Italiano</a> | <a href="README.pt-BR.md">Português (BR)</a>
3
+ </p>
4
+
5
+ <p align="center">
6
+ <img src="https://raw.githubusercontent.com/dogfood-lab/study-swarm/main/assets/study-swarm.png" alt="study-swarm" width="360">
7
+ </p>
8
+
9
+ <p align="center">
10
+ <a href="LICENSE"><img src="https://img.shields.io/badge/license-MIT-green" alt="MIT License"></a>
11
+ <a href="https://dogfood-lab.github.io/study-swarm/"><img src="https://img.shields.io/badge/handbook-live-purple" alt="Handbook"></a>
12
+ <img src="https://img.shields.io/badge/cited%20research-verified-1f6feb" alt="Cited research, verified">
13
+ </p>
14
+
15
+ **調査で引用された研究に基づいて設計上の決定を行い、その後、その決定が公式なものになる前に、*別の*モデルファミリーを使用して引用を検証します。**
16
+
17
+ `study-swarm`はツールではなく、プロトコルです。大規模言語モデル(LLM)を使用して重要な設計上の決定を行う場合(新しい製品レイヤー、アーキテクチャの選択、「ここでモデルを信頼すべきか」という判断など)に、即興で進めるのではなく、既存の知識に基づいて設計を進めると、時代遅れの設計になったり、存在しない情報源や、あなたが考えていることとは異なる内容の情報源に基づいて設計が進んでしまう可能性があります。`study-swarm`は、これらの問題を解決します。複数の研究エージェントを並行して実行し、特定の引用された情報を要求し、設計に影響を与える前に、すべての引用を**異なるモデルファミリーの外部検証者**を通じて検証します。
18
+
19
+ これは、自らその方法を実践しています。このプロトコルでは、設計を支援するシステムに対して、検証者によって保護された情報を提供するように規定されており、そのため、このプロトコル自体にもそれを適用します。**プロトコルを実行しているモデルであっても、自分の宿題を評価することはできません。**
20
+
21
+ ## プロトコルの5つのステップ
22
+
23
+ 1. **特定する**:経験的な証拠によって回答が変わる可能性のある、3〜5個の重要な設計上の質問を特定します。
24
+ 2. **実行する**:各質問に対して、1つの研究エージェントを並行して実行します。各エージェントは、論文のタイトル、著者、発表年、URL、および1文の結論を返します。広範囲にわたる情報よりも、具体的な情報を提供します(「6〜8件の信頼できる情報源が、20件の曖昧な情報よりも優れている」)。
25
+ 3. **統合する**:得られた情報を「研究による裏付け」セクションに統合します。`N. **<結論>.** <著者> <年> (<arXiv/DOI>). <設計への影響>.`
26
+ 4. **外部で検証する**:*別のモデルファミリー*(推論機能を削除したもの)を使用して、すべての引用を2つの段階で検証します。まず、**検索エンジン**が論文が存在することを確認します(モデルの記憶ではなく)。次に、**妥当性検証**ツールが、結論が情報源と一致することを確認します。**捏造または誤った引用が見つかった場合は、処理を停止します。検証者または検索エンジンが利用できない場合は、処理を停止し、エスカレーションします(情報がないことを「引用は問題ない」と解釈しないでください)。**
27
+ 5. **関連付ける**:各アーキテクチャの選択を、番号を使用して結論と関連付けます。設計への影響がない引用はノイズです。
28
+
29
+ 完全な実行可能な詳細(停止テーブル、情報源の標準、アンサンブルルール)は、**[PROTOCOL.md](PROTOCOL.md)**にあります。
30
+
31
+ ## なぜ*別の*モデルファミリーで、推論機能を削除する必要があるのでしょうか?
32
+
33
+ なぜなら、失敗のパターンは、仮説ではなく、実際に確認されているからです。
34
+
35
+ - **LLMは、自分の出力を確実に検証できません。** Huang et al. 2023 ([arXiv:2310.01798](https://arxiv.org/abs/2310.01798)); Kambhampati et al. 2024 ([arXiv:2402.01817](https://arxiv.org/abs/2402.01817), LLM-Modulo); Stechly et al. 2024 ([arXiv:2402.08115](https://arxiv.org/abs/2402.08115)) — 外部検証者がメリットをもたらします。自己批判的な内容は、効果がありません。
36
+ - **同じファミリーのモデルは、自分を高く評価します。** Panickssery, Bowman & Feng 2024 ([arXiv:2404.13076](https://arxiv.org/abs/2404.13076)) — 自己認識は、自己評価と*線形に*相関するため、部分的なブラインド処理だけでは役に立ちません。Verga et al. 2024 ([arXiv:2404.18796](https://arxiv.org/abs/2404.18796), PoLL) — 異なるファミリーのモデルで構成されたパネルの方が、偏りが少なく、コストも約7分の1で済みます。
37
+ - **LLMは、引用で嘘をつきます。** Walters & Wilder 2023 ([doi:10.1038/s41598-023-41032-5](https://doi.org/10.1038/s41598-023-41032-5)) — GPT-3.5の55%、GPT-4の18%の引用が捏造されています。Onweller et al. 2026 ([arXiv:2605.06635](https://arxiv.org/abs/2605.06635)) — リンクは94%以上の確率で解決されますが、引用されたコンテンツの39〜77%しか、実際に主張を裏付けていません。したがって、存在は**検索によって確認する必要があり、記憶に頼るべきではありません。**
38
+ - **生成者の推論を隠します。** Khalifa et al. 2026 ([arXiv:2601.14691](https://arxiv.org/abs/2601.14691), "Gaming the Judge") — 操作された推論のみでは、判断者の誤検出率が最大90%まで増加します(ただし、アクションは固定されています)。Turpin et al. 2023 ([arXiv:2305.04388](https://arxiv.org/abs/2305.04388)) — 推論は、事後的な合理化です。検証者は、単なる引用の主張のみを確認し、その「なぜこの情報を追加したのか」という理由は確認しません。
39
+ - **多様性は、数よりも重要です。** Rajan 2025 ([arXiv:2511.16708](https://arxiv.org/abs/2511.16708)) — ペアごとの相関関係がρ∈[0.05, 0.25]の4つの検証者が、単一の検証者よりも、サブモジュールカバレッジによって優れています。Kim et al. 2025 ([arXiv:2506.07962](https://arxiv.org/abs/2506.07962)) — LLMのエラーは*相関しているため、重要な変数は、生の数ではなく、レンズの多様性です。
40
+
41
+ ## 実際に機能するのでしょうか?(証拠)
42
+
43
+ テストとして、このプロトコルを自分の引用に対して実行しました。2つの相関性の低い、Claude以外のモデル(**Mistral** (`mistral-small:24b`)と**IBM Granite** (`granite4.1:30b`))を使用して、推論機能を削除した状態で、2つのブラインドトラップを仕込んだ引用セットを検証しました。
44
+
45
+ | 仕掛けられたトラップ | Mistral | IBM Granite | 真実 |
46
+ |---|---|---|---|
47
+ | 「Nakamura & Olsen」に帰属された推論 | 見逃された | **検出された**(誤って帰属された → 実際にはWei et al. 2022) | 誤って帰属された |
48
+ | 「98%のエラーが解消され、オラクルは不要」という捏造された論文 | **caught** (fabricated) | **caught** (fabricated) | 捏造された |
49
+
50
+ どちらのモデルも、単独では両方のトラップを検出できませんでしたが、**両方のモデルを組み合わせることで、2/2を検出しました。** 単一の判断者では、誤った帰属を検出できなかったでしょう。別途、検索エンジンは、当社の設計ドキュメントにある2つの*実際の*誤った帰属(間違った最初の著者に帰属された論文)を検出しました。これらの誤りは、パラメトリックLLMでは検出できなかったでしょう。また、両方のLLMが、トレーニングデータ以降に発表された論文を捏造されたものとして誤ってフラグを立てたのに対し、検索エンジンは、これらの論文が実際に存在することを正しく確認しました。この最後の点が、ステップ4の存在チェックが**検索エンジンである必要があり、LLMであってはならない**理由です。
51
+
52
+ この単一の実行は、ミニチュア版の論文です。**相関性の低いレンズと、存在を確認するための検索エンジンを組み合わせることで、単一の優れた判断者よりも優れた結果が得られます。**
53
+
54
+ ## その仕組み
55
+
56
+ このプロトコルを手動で実行することもできます。異なるモデルファミリーのモデルと、自分でarXiv/DOIを解決することで、ステップ4を満たすことができます。2つの関連ツールを使用すると、1つのコマンドで実行できます。
57
+
58
+ - **[prism-verify](https://github.com/mcp-tool-shop-org/prism-verify)** — 実行時検証ツール:異なるモデルファミリー間でのルーティング、推論の簡略化、多角的な判断、決定的な検索による存在確認(arXiv → Crossref)、および署名されたレシートを提供。
59
+ - **[role-os](https://github.com/mcp-tool-shop-org/role-os)** — `roleos verify-citations <dispatch>` を提供。このツールは、特定のドキュメントの引用を抽出し、それを prism を介して検証する。
60
+
61
+ ## なぜ機能するのか(簡潔に)
62
+
63
+ **現状** — この分野は急速に進展しており、特定の研究を数年かけて行うと、設計が18か月遅れてリリースされることになる。**機能性** — 証拠は、何が「機能する」かだけでなく、何が「機能しない」かを示す(説明が、誤ったAIへの過度な依存を招く可能性がある — Bansal et al. 2021)。**安全性** — 検証ツールによって保護された範囲は、証拠によって裏付けられるアーキテクチャであり、プロトコルはそれを自身の出力に適用する。情報源の提示は、学術的なパフォーマンスではなく、証拠の追跡である。
64
+
65
+ ## セキュリティ
66
+
67
+ `study-swarm` は、ドキュメントのリポジトリであり、Markdown とロゴが含まれている。実行可能なコードは含まれておらず、このリポジトリから何もインストールされない。データにアクセスせず、権限を必要とせず、テレメトリも収集しない。ソースコードには、秘密情報や認証情報は含まれていない。この手法は、Web検索とモデルベースの検証を使用するワークフローを「記述」しているが、このリポジトリはそれを実装または実行するものではない。詳細は [SECURITY.md](SECURITY.md) を参照。
68
+
69
+ ## ステータス
70
+
71
+ 動作するプロトコルであり、その独自の仕組みによって外部から検証されている。異なるモデルファミリーがその引用をチェックする(上記の証明を参照)。このリポジトリは、公開リファレンスであり、[PROTOCOL.md](PROTOCOL.md) が実行可能な形式である。これは、[dogfood-lab](https://github.com/dogfood-lab) ファミリーの一部であり、AI時代におけるビルドのための手法と事例を紹介する。
72
+
73
+ MITライセンス。
74
+
75
+ ---
76
+
77
+ <p align="center"><sub>Part of the <a href="https://github.com/dogfood-lab">dogfood-lab</a> family — methods &amp; showcases for building in the AI era. Built by <a href="https://mcp-tool-shop.github.io/">MCP Tool Shop</a>.</sub></p>
package/README.md CHANGED
@@ -1,10 +1,78 @@
1
- # @dogfood-lab/study-swarm
1
+ <p align="center">
2
+ <a href="README.ja.md">日本語</a> | <a href="README.zh.md">中文</a> | <a href="README.es.md">Español</a> | <a href="README.fr.md">Français</a> | <a href="README.hi.md">हिन्दी</a> | <a href="README.it.md">Italiano</a> | <a href="README.pt-BR.md">Português (BR)</a>
3
+ </p>
2
4
 
3
- > **Reserved placeholder (`0.0.0`).** This reserves the package name on Trusted-Publishing rails; a real package will follow.
5
+ <p align="center">
6
+ <img src="https://raw.githubusercontent.com/dogfood-lab/study-swarm/main/assets/study-swarm.png" alt="study-swarm" width="360">
7
+ </p>
4
8
 
5
- **study-swarm** is a methodology for grounding substantial design decisions in cited research — then verifying every citation with a *different* model family, reasoning-stripped, before any of it becomes canon.
9
+ <p align="center">
10
+ <a href="https://www.npmjs.com/package/@dogfood-lab/study-swarm"><img src="https://img.shields.io/npm/v/@dogfood-lab/study-swarm" alt="npm"></a>
11
+ <a href="LICENSE"><img src="https://img.shields.io/badge/license-MIT-green" alt="MIT License"></a>
12
+ <a href="https://dogfood-lab.github.io/study-swarm/"><img src="https://img.shields.io/badge/handbook-live-purple" alt="Handbook"></a>
13
+ <img src="https://img.shields.io/badge/cited%20research-verified-1f6feb" alt="Cited research, verified">
14
+ </p>
6
15
 
7
- - **Repository:** https://github.com/dogfood-lab/study-swarm
8
- - **Handbook:** https://dogfood-lab.github.io/study-swarm/
16
+ **Ground design decisions in cited research — then verify the citations with a *different* model family before any of it becomes canon.**
9
17
 
10
- Part of the [dogfood-lab](https://github.com/dogfood-lab) family.
18
+ `study-swarm` is a protocol, not a tool. When you're making a substantial design decision with an LLM — a new product layer, an architecture choice, a "should we trust the model here" call — improvising from first principles ships designs that are stale, and citing papers from memory ships designs that rest on sources that don't exist or don't say what you think. study-swarm replaces both: dispatch parallel research agents, demand specific cited findings, and gate every citation through an **external verifier of a different model family** before it informs the design.
19
+
20
+ It applies its own medicine. The protocol prescribes verifier-protected envelopes for the systems it helps design — so it runs one on itself. **No model grades its own homework, including the one running the protocol.**
21
+
22
+ ## The protocol in five steps
23
+
24
+ 1. **Identify** 3–5 load-bearing design questions where empirical evidence would change the answer.
25
+ 2. **Dispatch** one research agent per question, in parallel. Each must return paper titles + authors + years + URLs + a one-sentence finding — specificity over breadth ("6–8 well-sourced findings beat 20 vague gestures").
26
+ 3. **Synthesize** the findings into a *Research grounding* section: `N. **<finding>.** <Authors> <year> (<arXiv/DOI>). <design implication>.`
27
+ 4. **Verify externally** — a *different model family*, reasoning-stripped, checks every citation in two stages: a **retrieval oracle** confirms the paper exists (never the model's memory), then a **groundedness** lens confirms the finding matches the source. **Halt** on fabricated/misattributed; **halt-and-escalate** if the verifier or retrieval oracle is unavailable (never read absence as "citations fine").
28
+ 5. **Connect** each architectural choice back to a finding by number. Citations without a design implication are noise.
29
+
30
+ The full executable detail — the halt table, the sourcing standard, the ensemble rule — is in **[PROTOCOL.md](PROTOCOL.md)**.
31
+
32
+ ## Why a *different* family, reasoning-stripped?
33
+
34
+ Because the failure modes are documented, not hypothetical:
35
+
36
+ - **LLMs can't reliably verify their own output.** Huang et al. 2023 ([arXiv:2310.01798](https://arxiv.org/abs/2310.01798)); Kambhampati et al. 2024 ([arXiv:2402.01817](https://arxiv.org/abs/2402.01817), LLM-Modulo); Stechly et al. 2024 ([arXiv:2402.08115](https://arxiv.org/abs/2402.08115)) — the external verifier carries the gains; the self-critique content is inert.
37
+ - **Same-family judges self-prefer.** Panickssery, Bowman & Feng 2024 ([arXiv:2404.13076](https://arxiv.org/abs/2404.13076)) — self-recognition correlates *linearly* with self-preference, so partial blinding doesn't help. Verga et al. 2024 ([arXiv:2404.18796](https://arxiv.org/abs/2404.18796), PoLL) — a panel across disjoint families is less biased at ~7× lower cost.
38
+ - **Citations are where LLMs lie.** Walters & Wilder 2023 ([doi:10.1038/s41598-023-41032-5](https://doi.org/10.1038/s41598-023-41032-5)) — 55% of GPT-3.5 / 18% of GPT-4 citations are fabricated. Onweller et al. 2026 ([arXiv:2605.06635](https://arxiv.org/abs/2605.06635)) — links resolve >94% of the time yet only 39–77% of cited content actually supports the claim. So existence must be checked by **retrieval, not recall**.
39
+ - **Hide the generator's reasoning.** Khalifa et al. 2026 ([arXiv:2601.14691](https://arxiv.org/abs/2601.14691), "Gaming the Judge") — manipulated chain-of-thought alone inflates a judge's false-positives by up to 90% with actions held fixed. Turpin et al. 2023 ([arXiv:2305.04388](https://arxiv.org/abs/2305.04388)) — CoT is post-hoc rationalization. The verifier sees the bare citation claim, never the "why I included this."
40
+ - **Diversity beats count.** Rajan 2025 ([arXiv:2511.16708](https://arxiv.org/abs/2511.16708)) — four verifiers at pairwise correlation ρ ∈ [0.05, 0.25] beat any single one via submodular coverage. Kim et al. 2025 ([arXiv:2506.07962](https://arxiv.org/abs/2506.07962)) — LLM errors are *correlated*, so the load-bearing variable is lens diversity, not raw count.
41
+
42
+ ## Does it actually work? (proof)
43
+
44
+ As a test, the protocol was run against its own citations. Two decorrelated non-Claude families — **Mistral** (`mistral-small:24b`) and **IBM Granite** (`granite4.1:30b`) — checked a citation set, reasoning-stripped, seeded with two blind traps:
45
+
46
+ | Planted trap | Mistral | IBM Granite | Ground truth |
47
+ |---|---|---|---|
48
+ | Chain-of-thought prompting attributed to "Nakamura & Olsen" | missed | **caught** (misattributed → really Wei et al. 2022) | misattributed |
49
+ | a fabricated "98% of errors removed, no oracle needed" paper | **caught** (fabricated) | **caught** (fabricated) | fabricated |
50
+
51
+ Neither family caught both traps alone — but their **union caught 2/2**. A single judge would have shipped the misattribution. Separately, the retrieval oracle caught two *real* misattributions in our own design docs (papers cited under the wrong first author) that no parametric LLM could have flagged — and it correctly confirmed genuine 2026 papers that both LLMs false-flagged as fabricated simply because the papers postdate their training. That last point is the whole reason Step 4's existence check **must** be a retrieval oracle, never an LLM.
52
+
53
+ That single run is the thesis in miniature: **decorrelated lenses + a retrieval oracle for existence beat any one smart judge.**
54
+
55
+ ## How it's wired
56
+
57
+ You can run the protocol by hand — any different-family model plus resolving the arXiv/DOI yourself satisfies Step 4. Two sibling tools make it one command:
58
+
59
+ - **[prism-verify](https://github.com/mcp-tool-shop-org/prism-verify)** — the runtime verifier: family-different routing, reasoning-stripped, multi-lens adjudication, a deterministic retrieval existence floor (arXiv → Crossref), and signed receipts.
60
+ - **[role-os](https://github.com/mcp-tool-shop-org/role-os)** — provides `roleos verify-citations <dispatch>`, the runner that extracts a dispatch's citations and gates them through prism.
61
+
62
+ ## Why it works, in one breath
63
+
64
+ **Current** — the field moves fast; demanding specific studies-with-years keeps designs from shipping 18 months behind. **Functional** — evidence shows what *fails*, not just what works (explanations can increase over-reliance on *wrong* AI — Bansal et al. 2021). **Safe** — the verifier-protected envelope is the architecture the evidence supports, and the protocol enforces it on its own output. Sourcing isn't academic theater; it's the evidence trail.
65
+
66
+ ## Security
67
+
68
+ `study-swarm` is a documentation repository — Markdown and a logo. It ships no executable code and installs nothing from this repo. It touches no data, requires no permissions, and collects no telemetry; there are no secrets or credentials in the source. The methodology *describes* a workflow that uses web retrieval and model-based verification, but this repo does not implement or run it. See [SECURITY.md](SECURITY.md).
69
+
70
+ ## Status
71
+
72
+ A working protocol, externally verified by its own machinery — a different model family checks its citations (see the proof above). This repo is the public reference; [PROTOCOL.md](PROTOCOL.md) is the executable shape. Part of the [dogfood-lab](https://github.com/dogfood-lab) family — methods and showcases for building in the AI era.
73
+
74
+ MIT licensed.
75
+
76
+ ---
77
+
78
+ <p align="center"><sub>Part of the <a href="https://github.com/dogfood-lab">dogfood-lab</a> family — methods &amp; showcases for building in the AI era. Built by <a href="https://mcp-tool-shop.github.io/">MCP Tool Shop</a>.</sub></p>
@@ -0,0 +1,77 @@
1
+ <p align="center">
2
+ <a href="README.ja.md">日本語</a> | <a href="README.zh.md">中文</a> | <a href="README.es.md">Español</a> | <a href="README.fr.md">Français</a> | <a href="README.hi.md">हिन्दी</a> | <a href="README.it.md">Italiano</a> | <a href="README.md">English</a>
3
+ </p>
4
+
5
+ <p align="center">
6
+ <img src="https://raw.githubusercontent.com/dogfood-lab/study-swarm/main/assets/study-swarm.png" alt="study-swarm" width="360">
7
+ </p>
8
+
9
+ <p align="center">
10
+ <a href="LICENSE"><img src="https://img.shields.io/badge/license-MIT-green" alt="MIT License"></a>
11
+ <a href="https://dogfood-lab.github.io/study-swarm/"><img src="https://img.shields.io/badge/handbook-live-purple" alt="Handbook"></a>
12
+ <img src="https://img.shields.io/badge/cited%20research-verified-1f6feb" alt="Cited research, verified">
13
+ </p>
14
+
15
+ **Baseie as decisões de design em pesquisas citadas — e, em seguida, verifique as citações com um *modelo diferente* antes que qualquer coisa se torne um padrão.**
16
+
17
+ `study-swarm` é um protocolo, não uma ferramenta. Ao tomar uma decisão de design importante com um LLM — uma nova camada de produto, uma escolha de arquitetura, uma decisão sobre "se devemos confiar no modelo aqui" — improvisar a partir de princípios básicos resulta em designs desatualizados, e citar artigos de memória resulta em designs que se baseiam em fontes que não existem ou que não dizem o que você pensa. O `study-swarm` substitui ambos: envie agentes de pesquisa em paralelo, exija descobertas específicas citadas e valide cada citação por meio de um **verificador externo de uma família de modelos diferente** antes que ela influencie o design.
18
+
19
+ Ele aplica sua própria abordagem. O protocolo prescreve "envelopes" protegidos por verificadores para os sistemas que ele ajuda a projetar — portanto, ele executa um deles em si mesmo. **Nenhum modelo avalia seu próprio trabalho, incluindo o que executa o protocolo.**
20
+
21
+ ## O protocolo em cinco etapas
22
+
23
+ 1. **Identifique** 3 a 5 questões de design cruciais, nas quais evidências empíricas mudariam a resposta.
24
+ 2. **Envie** um agente de pesquisa por questão, em paralelo. Cada um deve retornar títulos de artigos + autores + anos + URLs + uma descoberta em uma frase — especificidade em vez de amplitude ("6 a 8 descobertas bem fundamentadas superam 20 observações vagas").
25
+ 3. **Sintetize** as descobertas em uma seção de *Fundamentação da pesquisa*: `N. **<descoberta>.** <Autores> <ano> (<arXiv/DOI>). <implicação para o design>.`
26
+ 4. **Verifique externamente** — uma *família de modelos diferente*, sem raciocínio, verifica cada citação em duas etapas: um **oráculo de recuperação** confirma que o artigo existe (nunca a memória do modelo) e, em seguida, uma lente de **fundamentação** confirma que a descoberta corresponde à fonte. **Interrompa** se for fabricada/atribuída incorretamente; **interrompa e alerte** se o verificador ou o oráculo de recuperação não estiverem disponíveis (nunca interprete a ausência como "citações válidas").
27
+ 5. **Conecte** cada escolha arquitetônica a uma descoberta por número. Citações sem uma implicação para o design são ruído.
28
+
29
+ Os detalhes completos e executáveis — a tabela de interrupção, o padrão de referência e a regra de conjunto — estão em **[PROTOCOL.md](PROTOCOL.md)**.
30
+
31
+ ## Por que uma *família diferente*, sem raciocínio?
32
+
33
+ Porque os modos de falha são documentados, não hipotéticos:
34
+
35
+ - **Os LLMs não conseguem verificar de forma confiável sua própria saída.** Huang et al. 2023 ([arXiv:2310.01798](https://arxiv.org/abs/2310.01798)); Kambhampati et al. 2024 ([arXiv:2402.01817](https://arxiv.org/abs/2402.01817), LLM-Modulo); Stechly et al. 2024 ([arXiv:2402.08115](https://arxiv.org/abs/2402.08115)) — o verificador externo traz os benefícios; o conteúdo de autocrítica é inerte.
36
+ - **Juízes da mesma família se auto favorecem.** Panickssery, Bowman & Feng 2024 ([arXiv:2404.13076](https://arxiv.org/abs/2404.13076)) — o auto reconhecimento se correlaciona *linearmente* com a auto preferência, portanto, o obscurecimento parcial não ajuda. Verga et al. 2024 ([arXiv:2404.18796](https://arxiv.org/abs/2404.18796), PoLL) — um painel em famílias distintas é menos tendencioso, com um custo ~7 vezes menor.
37
+ - **As citações são onde os LLMs mentem.** Walters & Wilder 2023 ([doi:10.1038/s41598-023-41032-5](https://doi.org/10.1038/s41598-023-41032-5)) — 55% das citações do GPT-3.5 / 18% do GPT-4 são fabricadas. Onweller et al. 2026 ([arXiv:2605.06635](https://arxiv.org/abs/2605.06635)) — os links resolvem >94% das vezes, mas apenas 39–77% do conteúdo citado realmente sustentam a afirmação. Portanto, a existência deve ser verificada por **recuperação, não por recordação**.
38
+ - **Oculte o raciocínio do gerador.** Khalifa et al. 2026 ([arXiv:2601.14691](https://arxiv.org/abs/2601.14691), "Gaming the Judge") — a manipulação do raciocínio em cadeia, por si só, inflaciona os falsos positivos de um juiz em até 90%, com as ações mantidas fixas. Turpin et al. 2023 ([arXiv:2305.04388](https://arxiv.org/abs/2305.04388)) — o raciocínio em cadeia é uma racionalização *a posteriori*. O verificador vê apenas a afirmação da citação, nunca o "por que eu incluí isso".
39
+ - **Diversidade supera a quantidade.** Rajan 2025 ([arXiv:2511.16708](https://arxiv.org/abs/2511.16708)) — quatro verificadores com correlação pareada ρ ∈ [0.05, 0.25] superam qualquer um deles por meio da cobertura submodular. Kim et al. 2025 ([arXiv:2506.07962](https://arxiv.org/abs/2506.07962)) — os erros do LLM são *correlacionados*, portanto, a variável crucial é a diversidade das lentes, não a quantidade bruta.
40
+
41
+ ## Ele realmente funciona? (prova)
42
+
43
+ Como teste, o protocolo foi executado em suas próprias citações. Duas famílias não correlacionadas, diferentes do Claude — **Mistral** (`mistral-small:24b`) e **IBM Granite** (`granite4.1:30b`) — verificaram um conjunto de citações, sem raciocínio, com duas armadilhas ocultas:
44
+
45
+ | Armadilha plantada | Mistral | IBM Granite | Verdade |
46
+ |---|---|---|---|
47
+ | O raciocínio em cadeia atribuído a "Nakamura & Olsen" | não detectado | **detectado** (atribuído incorretamente → na verdade, Wei et al. 2022) | atribuído incorretamente |
48
+ | um artigo fabricado com "98% dos erros removidos, nenhum oráculo necessário" | **caught** (fabricated) | **caught** (fabricated) | fabricado |
49
+
50
+ Nenhuma das famílias detectou as duas armadilhas sozinha — mas a **união detectou 2/2**. Um único juiz teria validado a atribuição incorreta. Separadamente, o oráculo de recuperação detectou duas *atribuições incorretas reais* em nossos próprios documentos de design (artigos citados sob o autor principal errado) que nenhum LLM paramétrico poderia ter sinalizado — e ele confirmou corretamente artigos genuínos de 2026 que ambos os LLMs sinalizaram falsamente como fabricados, simplesmente porque os artigos são posteriores ao seu treinamento. Esse último ponto é a razão pela qual a verificação de existência na etapa 4 **deve** ser um oráculo de recuperação, nunca um LLM.
51
+
52
+ Essa única execução é a tese em miniatura: **lentes descoordenadas + um oráculo de recuperação para existência superam qualquer juiz inteligente.**
53
+
54
+ ## Como está conectado
55
+
56
+ Você pode executar o protocolo manualmente — qualquer modelo de família diferente, juntamente com a resolução do arXiv/DOI, satisfaz a etapa 4. Duas ferramentas auxiliares tornam isso um único comando:
57
+
58
+ - **[prism-verify](https://github.com/mcp-tool-shop-org/prism-verify)** — o verificador em tempo de execução: roteamento diferenciado por família, sem inferências, adjudicação multi-lente, um limite determinístico de existência de recuperação (arXiv → Crossref) e recibos assinados.
59
+ - **[role-os](https://github.com/mcp-tool-shop-org/role-os)** — fornece `roleos verify-citations <dispatch>`, o executor que extrai as citações de um documento e as valida através do prism.
60
+
61
+ ## Por que funciona, em poucas palavras
62
+
63
+ **Atual** — o campo evolui rapidamente; exigir estudos específicos com anos de duração impede que os projetos sejam lançados com 18 meses de atraso. **Funcional** — a evidência mostra o que *falha*, não apenas o que funciona (as explicações podem aumentar a dependência excessiva de uma IA *incorreta* — Bansal et al. 2021). **Seguro** — o envelope protegido pelo verificador é a arquitetura que a evidência suporta, e o protocolo a impõe em sua própria saída. A obtenção de fontes não é um exercício acadêmico; é o rastro da evidência.
64
+
65
+ ## Segurança
66
+
67
+ `study-swarm` é um repositório de documentação — Markdown e um logotipo. Não contém código executável e não instala nada deste repositório. Não acessa dados, não requer permissões e não coleta dados de telemetria; não há segredos ou credenciais no código-fonte. A metodologia *descreve* um fluxo de trabalho que usa recuperação da web e verificação baseada em modelo, mas este repositório não o implementa nem o executa. Consulte [SECURITY.md](SECURITY.md).
68
+
69
+ ## Status
70
+
71
+ Um protocolo funcional, verificado externamente por sua própria ferramenta — uma família de modelos diferente verifica suas citações (veja a prova acima). Este repositório é a referência pública; [PROTOCOL.md](PROTOCOL.md) é a forma executável. Parte da família [dogfood-lab](https://github.com/dogfood-lab) — métodos e demonstrações para construir na era da IA.
72
+
73
+ Licenciado sob a licença MIT.
74
+
75
+ ---
76
+
77
+ <p align="center"><sub>Part of the <a href="https://github.com/dogfood-lab">dogfood-lab</a> family — methods &amp; showcases for building in the AI era. Built by <a href="https://mcp-tool-shop.github.io/">MCP Tool Shop</a>.</sub></p>
package/README.zh.md ADDED
@@ -0,0 +1,77 @@
1
+ <p align="center">
2
+ <a href="README.ja.md">日本語</a> | <a href="README.md">English</a> | <a href="README.es.md">Español</a> | <a href="README.fr.md">Français</a> | <a href="README.hi.md">हिन्दी</a> | <a href="README.it.md">Italiano</a> | <a href="README.pt-BR.md">Português (BR)</a>
3
+ </p>
4
+
5
+ <p align="center">
6
+ <img src="https://raw.githubusercontent.com/dogfood-lab/study-swarm/main/assets/study-swarm.png" alt="study-swarm" width="360">
7
+ </p>
8
+
9
+ <p align="center">
10
+ <a href="LICENSE"><img src="https://img.shields.io/badge/license-MIT-green" alt="MIT License"></a>
11
+ <a href="https://dogfood-lab.github.io/study-swarm/"><img src="https://img.shields.io/badge/handbook-live-purple" alt="Handbook"></a>
12
+ <img src="https://img.shields.io/badge/cited%20research-verified-1f6feb" alt="Cited research, verified">
13
+ </p>
14
+
15
+ **将设计决策建立在引用的研究基础上——然后,在任何内容成为标准之前,使用*不同的*模型系列来验证这些引用。**
16
+
17
+ `study-swarm` 是一种协议,而不是一种工具。当您使用大型语言模型 (LLM) 做出重大的设计决策时——例如,创建一个新的产品层、选择一种架构,或者决定“我们是否应该在这里信任该模型”——如果仅凭经验进行设计,会导致设计陈旧;如果仅凭记忆引用论文,会导致设计依赖于不存在或内容与您认为的不同来源。`study-swarm` 替代了这两种方法:它会派遣并行研究代理,要求提供具体的引用结果,并且在任何引用结果影响设计之前,都会通过一个**来自不同模型系列的外部验证器**进行验证。
18
+
19
+ 它会自我应用。该协议规定,对于它所帮助设计的系统,必须使用经过验证器保护的机制——因此,它也会对自己进行验证。**没有模型会自己批改作业,包括运行该协议的模型。**
20
+
21
+ ## 该协议包含五个步骤
22
+
23
+ 1. **确定** 3-5 个关键的设计问题,如果存在经验证据,这些证据可能会改变答案。
24
+ 2. **派遣**一个研究代理来处理每个问题,并并行进行。每个代理都必须返回论文标题 + 作者 + 年份 + URL + 一句话的结论——强调具体性而非广泛性(“6-8 个来源可靠的结论胜过 20 个模糊的描述”)。
25
+ 3. **将结论综合**到“研究依据”部分:`N. **<结论>.** <作者> <年份> (<arXiv/DOI>). <设计意义>.`
26
+ 4. **进行外部验证**——一个*不同的模型系列*,去除推理能力,以两个阶段检查每个引用:一个**检索预言机**确认论文是否存在(永远不是模型的记忆),然后一个**相关性**过滤器确认结论与来源是否匹配。如果发现捏造或错误引用,则**停止**;如果验证器或检索预言机不可用,则**停止并升级**(永远不要将无法找到的情况视为“引用没问题”)。
27
+ 5. **将每个架构选择与一个结论联系起来**,并按编号进行。没有设计意义的引用就是噪音。
28
+
29
+ 完整的可执行细节——停止表、来源标准、集成规则——都包含在 **[PROTOCOL.md](PROTOCOL.md)** 中。
30
+
31
+ ## 为什么需要一个*不同的*模型系列,并且去除推理能力?
32
+
33
+ 因为失败模式是经过记录的,而不是假设的:
34
+
35
+ - **大型语言模型无法可靠地验证自己的输出。** Huang 等人,2023 年 ([arXiv:2310.01798](https://arxiv.org/abs/2310.01798));Kambhampati 等人,2024 年 ([arXiv:2402.01817](https://arxiv.org/abs/2402.01817),LLM-Modulo);Stechly 等人,2024 年 ([arXiv:2402.08115](https://arxiv.org/abs/2402.08115))——外部验证器可以带来收益;自我批评的内容是无效的。
36
+ - **同一系列的评估者会偏向于自我。** Panickssery、Bowman 和 Feng,2024 年 ([arXiv:2404.13076](https://arxiv.org/abs/2404.13076))——自我识别与自我偏好呈*线性*相关,因此部分隐藏并不能提供帮助。Verga 等人,2024 年 ([arXiv:2404.18796](https://arxiv.org/abs/2404.18796),PoLL))——来自不同系列的评估小组的偏见更小,成本约为原来的 1/7。
37
+ - **大型语言模型最容易在引用方面撒谎。** Walters 和 Wilder,2023 年 ([doi:10.1038/s41598-023-41032-5](https://doi.org/10.1038/s41598-023-41032-5))——55% 的 GPT-3.5 / 18% 的 GPT-4 引用是捏造的。Onweller 等人,2026 年 ([arXiv:2605.06635](https://arxiv.org/abs/2605.06635))——链接可以解决超过 94% 的问题,但只有 39-77% 的引用内容实际上支持该主张。因此,必须通过**检索**来检查是否存在,而不是通过**回忆**。
38
+ - **隐藏生成器的推理过程。** Khalifa 等人,2026 年 ([arXiv:2601.14691](https://arxiv.org/abs/2601.14691),“欺骗评估者”)——仅通过操纵思维链,就可以使评估者的假阳性率提高高达 90%,而操作保持不变。Turpin 等人,2023 年 ([arXiv:2305.04388](https://arxiv.org/abs/2305.04388))——思维链是一种事后合理化。验证器只会看到裸露的引用声明,而不会看到“我为什么包含这个”。
39
+ - **多样性胜过数量。** Rajan,2025 年 ([arXiv:2511.16708](https://arxiv.org/abs/2511.16708))——四个验证器,成对相关性 ρ ∈ [0.05, 0.25],通过亚模覆盖胜过任何一个智能评估者。Kim 等人,2025 年 ([arXiv:2506.07962](https://arxiv.org/abs/2506.07962))——大型语言模型的错误是*相关的*,因此关键变量是验证器的多样性,而不是原始数量。
40
+
41
+ ## 它真的有效吗?(证明)
42
+
43
+ 作为一项测试,该协议被应用于它自己的引用。两个不相关的非 Claude 系列——**Mistral** (`mistral-small:24b`) 和 **IBM Granite** (`granite4.1:30b`)——检查了一组引用,并去除了推理能力,并设置了两个盲目陷阱:
44
+
45
+ | 设置的陷阱 | Mistral | IBM Granite | 真实情况 |
46
+ |---|---|---|---|
47
+ | 思维链提示归因于“Nakamura & Olsen” | 未发现 | **发现**(错误归因 → 实际上是 Wei 等人,2022 年) | 错误归因 |
48
+ | 一个捏造的“98% 的错误已消除,不需要预言机”论文 | **caught** (fabricated) | **caught** (fabricated) | 捏造 |
49
+
50
+ 两个模型单独都没有发现这两个陷阱——但它们的**组合发现了 2/2 个陷阱**。如果只有一个评估者,它会忽略错误归因。此外,检索预言机发现了我们自己设计文档中两个*真实的*错误归因(引用了错误的作者),而任何参数化的大型语言模型都无法识别——并且它正确地确认了 2026 年的真实论文,而这两个大型语言模型都错误地将其标记为捏造,仅仅是因为这些论文的发表时间晚于它们的训练时间。最后一点是,步骤 4 的存在性检查**必须**使用检索预言机,而不是大型语言模型。
51
+
52
+ 这个简单的测试就是缩影:**不相关的验证器 + 用于验证存在性的检索预言机,胜过任何一个智能评估者。**
53
+
54
+ ## 它的工作原理
55
+
56
+ 您可以手动运行该协议——任何不同的模型系列,再加上您自己解析 arXiv/DOI,就可以满足步骤 4 的要求。两个辅助工具可以将其简化为一个命令:
57
+
58
+ - **[prism-verify](https://github.com/mcp-tool-shop-org/prism-verify)** — 运行时验证器:不同类型的路由、去除推理过程、多角度仲裁、确定性的检索存在性保障(arXiv → Crossref),以及带签名的收据。
59
+ - **[role-os](https://github.com/mcp-tool-shop-org/role-os)** — 提供 `roleos verify-citations <dispatch>` 命令,该命令用于提取某个任务的引用,并通过 prism 进行验证。
60
+
61
+ ## 它为何有效,一言以蔽之
62
+
63
+ **实用性** — 该领域发展迅速;要求进行特定且耗时的研究,会导致设计落后 18 个月。**功能性** — 证据表明哪些*失败*,而不仅仅是哪些有效(解释可能会增加对*错误*人工智能的过度依赖——Bansal 等人,2021 年)。**安全性** — 验证器保护的范围是证据支持的架构,协议对其自身的输出进行强制执行。来源不是学术表演;它是证据链。
64
+
65
+ ## 安全性
66
+
67
+ `study-swarm` 是一个文档仓库——包含 Markdown 文件和一个徽标。它不包含任何可执行代码,也不从该仓库安装任何内容。它不涉及任何数据,不需要任何权限,也不收集任何遥测数据;源代码中没有秘密或凭据。该方法*描述*了一种使用网络检索和基于模型的验证的工作流程,但此仓库不实现或运行该流程。请参阅 [SECURITY.md](SECURITY.md)。
68
+
69
+ ## 状态
70
+
71
+ 一个可运行的协议,由其自身的机制进行外部验证——不同的模型家族检查其引用(参见上面的证明)。此仓库是公共参考;[PROTOCOL.md](PROTOCOL.md) 是可执行的形式。它是 [dogfood-lab](https://github.com/dogfood-lab) 家族的一部分——用于在人工智能时代构建方法和展示案例。
72
+
73
+ 采用 MIT 许可。
74
+
75
+ ---
76
+
77
+ <p align="center"><sub>Part of the <a href="https://github.com/dogfood-lab">dogfood-lab</a> family — methods &amp; showcases for building in the AI era. Built by <a href="https://mcp-tool-shop.github.io/">MCP Tool Shop</a>.</sub></p>
package/SECURITY.md ADDED
@@ -0,0 +1,23 @@
1
+ # Security Policy
2
+
3
+ `study-swarm` is a **documentation repository** — it contains the study-swarm methodology (Markdown) and a logo asset. It ships no executable code, no compiled artifacts, and installs nothing from this repository. (The npm name `@dogfood-lab/study-swarm` is a reserved placeholder; this repo is the methodology source, not the package.)
4
+
5
+ ## Threat model
6
+
7
+ - **What it touches:** nothing at runtime. There is no program to run; reading the docs executes no code.
8
+ - **What it does NOT touch:** your filesystem, network, credentials, or environment.
9
+ - **Telemetry:** none. **Secrets/credentials:** none in source.
10
+ - **Permissions required:** none.
11
+
12
+ The methodology *describes* a workflow that uses web retrieval and model-based verification, but this repository does not implement or execute that workflow.
13
+
14
+ ## Supported versions
15
+
16
+ | Version | Supported |
17
+ |---------|-----------|
18
+ | 1.x | ✅ |
19
+ | < 1.0 | — |
20
+
21
+ ## Reporting
22
+
23
+ Found an error in the methodology, a broken or misattributed citation, or a security concern in related tooling? Open an issue at <https://github.com/dogfood-lab/study-swarm/issues>, or email **64996768+mcp-tool-shop@users.noreply.github.com**. We aim to acknowledge within 7 days.
package/package.json CHANGED
@@ -1,17 +1,39 @@
1
1
  {
2
2
  "name": "@dogfood-lab/study-swarm",
3
- "version": "0.0.0",
4
- "description": "Reserved placeholder study-swarm is a research-grounded design protocol. See https://github.com/dogfood-lab/study-swarm",
3
+ "version": "0.5.0",
4
+ "description": "Ground design decisions in cited research, then verify every citation with a different model family before it becomes canon — a research-grounded design protocol.",
5
+ "keywords": [
6
+ "methodology",
7
+ "llm",
8
+ "research",
9
+ "citations",
10
+ "verification",
11
+ "ai-agents",
12
+ "fact-checking",
13
+ "prompt-engineering"
14
+ ],
5
15
  "license": "MIT",
16
+ "author": "mcp-tool-shop <64996768+mcp-tool-shop@users.noreply.github.com>",
17
+ "homepage": "https://dogfood-lab.github.io/study-swarm/",
6
18
  "repository": {
7
19
  "type": "git",
8
20
  "url": "git+https://github.com/dogfood-lab/study-swarm.git"
9
21
  },
10
- "homepage": "https://dogfood-lab.github.io/study-swarm/",
22
+ "bugs": {
23
+ "url": "https://github.com/dogfood-lab/study-swarm/issues"
24
+ },
11
25
  "files": [
12
- "README.md"
13
- ],
14
- "publishConfig": {
15
- "access": "public"
16
- }
26
+ "README.md",
27
+ "README.ja.md",
28
+ "README.zh.md",
29
+ "README.es.md",
30
+ "README.fr.md",
31
+ "README.hi.md",
32
+ "README.it.md",
33
+ "README.pt-BR.md",
34
+ "PROTOCOL.md",
35
+ "SECURITY.md",
36
+ "LICENSE",
37
+ "CHANGELOG.md"
38
+ ]
17
39
  }