Research findings on the translation layer problem. Evaluate these and tell us what to build.

FINDING 1: DeFrame (arxiv, February 2026) A debiasing framework that reduces framing-induced variance by 92%. It adds a System 2 step: the model considers an alternative framing of the same question, derives fairness guidelines, and revises its answer. Reduces framing disparity by 92% and bias score by 93% on BBQ benchmark. Works on single models only.

FINDING 2: Black-box Prompt Optimization (BPO) Shifts alignment from model-centric to input-centric. Instead of changing the model, you change the prompt before it reaches the model. Multi-agent systems analyze and rewrite prompts to fix inconsistencies while preserving original intent. Matches intended use to ideal model. No one applies this to multi-model fleets.

FINDING 3: Intent-Action Separation (established architecture) The LLM translates raw user input into a logical form or dialog act expressed in a predefined DSL. This preserves a clear separation between understanding and action execution. The LLM handles understanding, deterministic rules handle execution. Google's research shows small models can extract intent with high accuracy through decomposition.

FINDING 4: Intent Alignment Strategy (IAS) A technical approach for ensuring AI internal processing and final actions are systematically correlated with the user's explicit or latent intent. Focuses on intermediate representations.

FINDING 5: The gap in the literature DeFrame works on single models. BPO works on single models. Intent classification works on single models. NOBODY applies prompt normalization before dispatching to a multi-model fleet where each model has its own professional role framework.

OUR CONTEXT: Polybrain has 9 models with professional role frameworks (GAAS auditor, OWASP adversarial reviewer, ISO 9001 scorer, etc). Currently the human prompt goes raw to all 9. The role framework is the only filter. We proved tonight that prompt framing shapes responses (Cycle 032: telling all models to be Kimi collapsed all voices into one).

QUESTION: What should the translation layer do, specifically? How do we combine DeFrame + BPO + Intent-Action Separation into something that works for a multi-model fleet? Be concrete. File names, data flow, what the normalized prompt looks like.

**Cycle ID:** `cycle_039_unknown` **Verified at:** 2026-04-08T07:28:47.993Z **Ensemble:** 9 models from 3 providers **Result:** 9 of 9 models responded **Cycle wall time:** 40.411 seconds **Canonical URL:** https://trust.polylogicai.com/claim/research-findings-on-the-translation-layer-problem-evaluate-these-and-tell-us-wh **Source paper:** [PolybrainBench (version 12)](https://trust.polylogicai.com/polybrainbench) **Source ledger row:** [`public-ledger.jsonl#cycle_039_unknown`](https://huggingface.co/datasets/polylogic/polybrainbench/blob/main/public-ledger.jsonl) **Cryptographic provenance:** SHA-256 `86da7b4553605dd564b9533fa053fd222d121cd71672bd356e34a34c1a9d6483`

Verification verdict

Of 9 models in the ensemble, 9 responded successfully and 0 failed.

Per-model responses

The full text of each model's response is available in the source ledger. The summary below records each model's success or failure and the first 280 characters of its response.

| Model | Status | Response chars | | --- | :---: | ---: | | gpt-4.1-mini | ✓ | 9153 | | gpt-4.1-nano | ✓ | 6465 | | gpt-oss-120b | ✓ | 18337 | | grok-3-mini | ✓ | 18296 | | grok-4-fast | ✓ | 11244 | | kimi-k2-groq | ✓ | 6016 | | llama-3.3-70b | ✓ | 3649 | | llama-4-scout | ✓ | 3454 | | qwen3-32b | ✓ | 9387 |

Pairwise agreement

The pairwise Jaccard agreement between successful responses for this cycle:

_Per-cycle pairwise agreement matrix is computed offline; will be populated in canonical page v2._

Divergence score

This cycle's divergence score is **TBD** on a 0 to 1 scale, where 0 means all responses are token-identical and 1 means no two responses share any tokens. The dataset-wide median divergence is 0.5 for context.

How to cite this claim

```bibtex @misc{polybrainbench_claim_cycle_039_unknown, author = {Polylogic AI}, title = {Research findings on the translation layer problem. Evaluate these and tell us what to build.

FINDING 1: DeFrame (arxiv, February 2026) A debiasing framework that reduces framing-induced variance by 92%. It adds a System 2 step: the model considers an alternative framing of the same question, derives fairness guidelines, and revises its answer. Reduces framing disparity by 92% and bias score by 93% on BBQ benchmark. Works on single models only.

FINDING 2: Black-box Prompt Optimization (BPO) Shifts alignment from model-centric to input-centric. Instead of changing the model, you change the prompt before it reaches the model. Multi-agent systems analyze and rewrite prompts to fix inconsistencies while preserving original intent. Matches intended use to ideal model. No one applies this to multi-model fleets.

FINDING 3: Intent-Action Separation (established architecture) The LLM translates raw user input into a logical form or dialog act expressed in a predefined DSL. This preserves a clear separation between understanding and action execution. The LLM handles understanding, deterministic rules handle execution. Google's research shows small models can extract intent with high accuracy through decomposition.

FINDING 4: Intent Alignment Strategy (IAS) A technical approach for ensuring AI internal processing and final actions are systematically correlated with the user's explicit or latent intent. Focuses on intermediate representations.

FINDING 5: The gap in the literature DeFrame works on single models. BPO works on single models. Intent classification works on single models. NOBODY applies prompt normalization before dispatching to a multi-model fleet where each model has its own professional role framework.

OUR CONTEXT: Polybrain has 9 models with professional role frameworks (GAAS auditor, OWASP adversarial reviewer, ISO 9001 scorer, etc). Currently the human prompt goes raw to all 9. The role framework is the only filter. We proved tonight that prompt framing shapes responses (Cycle 032: telling all models to be Kimi collapsed all voices into one).

QUESTION: What should the translation layer do, specifically? How do we combine DeFrame + BPO + Intent-Action Separation into something that works for a multi-model fleet? Be concrete. File names, data flow, what the normalized prompt looks like.}, year = {2026}, howpublished = {PolybrainBench cycle cycle_039_unknown}, url = {https://trust.polylogicai.com/claim/research-findings-on-the-translation-layer-problem-evaluate-these-and-tell-us-wh} } ```

Reproduce this cycle

```bash node ~/polybrain/bin/polybrain-cycle.mjs start --raw --fast "Research findings on the translation layer problem. Evaluate these and tell us what to build.

FINDING 1: DeFrame (arxiv, February 2026) A debiasing framework that reduces framing-induced variance by 92%. It adds a System 2 step: the model considers an alternative framing of the same question, derives fairness guidelines, and revises its answer. Reduces framing disparity by 92% and bias score by 93% on BBQ benchmark. Works on single models only.

FINDING 2: Black-box Prompt Optimization (BPO) Shifts alignment from model-centric to input-centric. Instead of changing the model, you change the prompt before it reaches the model. Multi-agent systems analyze and rewrite prompts to fix inconsistencies while preserving original intent. Matches intended use to ideal model. No one applies this to multi-model fleets.

FINDING 3: Intent-Action Separation (established architecture) The LLM translates raw user input into a logical form or dialog act expressed in a predefined DSL. This preserves a clear separation between understanding and action execution. The LLM handles understanding, deterministic rules handle execution. Google's research shows small models can extract intent with high accuracy through decomposition.

FINDING 4: Intent Alignment Strategy (IAS) A technical approach for ensuring AI internal processing and final actions are systematically correlated with the user's explicit or latent intent. Focuses on intermediate representations.

FINDING 5: The gap in the literature DeFrame works on single models. BPO works on single models. Intent classification works on single models. NOBODY applies prompt normalization before dispatching to a multi-model fleet where each model has its own professional role framework.

OUR CONTEXT: Polybrain has 9 models with professional role frameworks (GAAS auditor, OWASP adversarial reviewer, ISO 9001 scorer, etc). Currently the human prompt goes raw to all 9. The role framework is the only filter. We proved tonight that prompt framing shapes responses (Cycle 032: telling all models to be Kimi collapsed all voices into one).

QUESTION: What should the translation layer do, specifically? How do we combine DeFrame + BPO + Intent-Action Separation into something that works for a multi-model fleet? Be concrete. File names, data flow, what the normalized prompt looks like." ```

Schema.org structured data

```json { "@context": "https://schema.org", "@type": "ClaimReview", "datePublished": "2026-04-08T07:28:47.993Z", "url": "https://trust.polylogicai.com/claim/research-findings-on-the-translation-layer-problem-evaluate-these-and-tell-us-wh", "claimReviewed": "Research findings on the translation layer problem. Evaluate these and tell us what to build.

FINDING 1: DeFrame (arxiv, February 2026) A debiasing framework that reduces framing-induced variance by 92%. It adds a System 2 step: the model considers an alternative framing of the same question, derives fairness guidelines, and revises its answer. Reduces framing disparity by 92% and bias score by 93% on BBQ benchmark. Works on single models only.

FINDING 2: Black-box Prompt Optimization (BPO) Shifts alignment from model-centric to input-centric. Instead of changing the model, you change the prompt before it reaches the model. Multi-agent systems analyze and rewrite prompts to fix inconsistencies while preserving original intent. Matches intended use to ideal model. No one applies this to multi-model fleets.

FINDING 3: Intent-Action Separation (established architecture) The LLM translates raw user input into a logical form or dialog act expressed in a predefined DSL. This preserves a clear separation between understanding and action execution. The LLM handles understanding, deterministic rules handle execution. Google's research shows small models can extract intent with high accuracy through decomposition.

FINDING 4: Intent Alignment Strategy (IAS) A technical approach for ensuring AI internal processing and final actions are systematically correlated with the user's explicit or latent intent. Focuses on intermediate representations.

FINDING 5: The gap in the literature DeFrame works on single models. BPO works on single models. Intent classification works on single models. NOBODY applies prompt normalization before dispatching to a multi-model fleet where each model has its own professional role framework.

OUR CONTEXT: Polybrain has 9 models with professional role frameworks (GAAS auditor, OWASP adversarial reviewer, ISO 9001 scorer, etc). Currently the human prompt goes raw to all 9. The role framework is the only filter. We proved tonight that prompt framing shapes responses (Cycle 032: telling all models to be Kimi collapsed all voices into one).

QUESTION: What should the translation layer do, specifically? How do we combine DeFrame + BPO + Intent-Action Separation into something that works for a multi-model fleet? Be concrete. File names, data flow, what the normalized prompt looks like.", "itemReviewed": { "@type": "Claim", "datePublished": "2026-04-08T07:28:47.993Z", "appearance": "https://trust.polylogicai.com/claim/research-findings-on-the-translation-layer-problem-evaluate-these-and-tell-us-wh", "author": { "@type": "Organization", "name": "PolybrainBench" } }, "reviewRating": { "@type": "Rating", "ratingValue": "9", "bestRating": "9", "worstRating": "0", "alternateName": "Unanimous" }, "author": { "@type": "Organization", "name": "Polylogic AI", "url": "https://polylogicai.com" } } ```

Provenance and integrity

This page was generated by the PolybrainBench daemon at version 0.1.0 from cycle cycle_039_unknown. The full provenance chain (per-response SHA-256 stamps, cross-cycle prev-hash linking, Thalamus grounding verification) is recorded in the source cycle directory at `~/polybrain/cycles/039/provenance.json` and mirrored in the published dataset. The page is regenerated on every harvest pass; the URL is permanent and the content is immutable for any given paper version.