docs: Add preprints on prompt-level incentive restructuring and hallucination mitigation

[mahashu]

Description

This pull request updates the literature catalog by indexing two companion open-science preprints evaluating prompt-level constraint architectures on frontier models (ChatGPT and Gemini).

Papers Added & Core Findings

• Standing on a Trapdoor: AI Hallucination and Prompt-Level Cost Restructuring (Kowalski et al., 2026; DOI: 10.5281/zenodo.20019087)
  • Focus: Introduces the baseline IDK+COMP constraint framework across 410 trials. Documents how default corporate alignment layers trade factual precision for conversational fluency, and demonstrates how to force model generation to terminate cleanly at the factual boundary.
• A Puma in a Teacup: Signal Quality and Hallucination Suppression Through Prompt-Level Incentive Restructuring (Kowalski et al., 2026; DOI: 10.5281/zenodo.19502460)
  • Focus: Analyzes context-saturation behaviors, unhedged refusal bounds, and creative signal optimization across 362 trials using test strings with no ground truth. Documents the "Brake-and-Slide" failure mode where isolation of the compression mandate (COMP alone) paradoxically forces Gemini's fabrication rates up to 70%, while removing refusal permissions (IDK) spikes it to 100%.

This contribution ensures that developers tracking the Applications index have direct access to both the empirical performance metrics and the underlying behavioral failure modes of inference-time safety layers.

[vercel]

@mahashu is attempting to deploy a commit to the DAIR-AI Team on Vercel.

A member of the Team first needs to authorize it.

[mahashu]

Clarification on methodology: The 410 trials represent the cumulative project total. The initial 362 trials are established in "A Puma in a Teacup" to map the core evaluation matrix across multiple governance conditions (including baseline, OGS, OGS-IDK, and COMP). The newest tracking block (trials 363-410) is introduced in "Standing on a Trapdoor" to isolate the specific IDK+COMP constraint configuration, requiring direct cross-comparison against those pre-established governance frameworks to map structural failure and suppression mechanics.

[Nakagawa-master]

The papers may be useful additions, but I think the catalog wording should preserve one evidence boundary more explicitly.

Both entries summarize preprints reporting their own experiments, while the current copy turns those reported results into unqualified guide-level claims:

"produces suppression across three frontier models"

and

"Ablation results identify the load-bearing constraint"

For a literature index, I would keep the finding attributed to the paper rather than silently upgrading:

authors report X under their tested conditions
!=
X is independently established across the model class

A minimal edit would be enough, e.g.:

• “The authors report that IDK+COMP suppressed hallucination across the three models tested …”
• “The authors’ ablation results identify … within the reported evaluation setup …”

If space permits, naming the exact tested model/version(s) is stronger than “frontier models,” because that phrase will age and can imply broader coverage than the experiment actually establishes.

I would also avoid using trial count itself as a proxy for external validity. “410 empirical trials” is useful descriptive metadata; it does not by itself establish independence, generalization, or replication.

This does not require the guide to independently reproduce the work before listing it. It only keeps three states distinct in the reader-facing text:

paper exists
→ paper reports a measured result
→ independent replication/generalization (separate question)

Source relation: this framing is informed by Nakagawa Master's public establishment-condition / evidence-lineage work, especially separating a locally observed success from broader establishment and separating provenance from truth proof. The concrete suggestion above is scoped to the wording of these two catalog entries.

Public independent-verification route:
https://github.com/Nakagawa-master/nakagawa-theory-archive/blob/main/INDEPENDENT_VERIFICATION_REUSE.md

[mahashu]

Updated wording per your suggestion — pushed to the branch. Both entries now attribute the findings to the authors' reported results rather than stating them as established, and name the tested model versions instead of "frontier models." Trial counts left as descriptive scope only. Thanks for the catch.

…

On Tue, Sep 22, 2026 at 8:46 PM Nakagawa-master ***@***.***> wrote: ***@***.**** commented on this pull request. The papers may be useful additions, but I think the catalog wording should preserve one evidence boundary more explicitly. Both entries summarize *preprints reporting their own experiments*, while the current copy turns those reported results into unqualified guide-level claims: "produces suppression across three frontier models" and "Ablation results identify the load-bearing constraint" For a literature index, I would keep the finding attributed to the paper rather than silently upgrading: authors report X under their tested conditions != X is independently established across the model class A minimal edit would be enough, e.g.: - “The authors report that IDK+COMP suppressed hallucination across the three models tested …” - “The authors’ ablation results identify … within the reported evaluation setup …” If space permits, naming the exact tested model/version(s) is stronger than “frontier models,” because that phrase will age and can imply broader coverage than the experiment actually establishes. I would also avoid using trial count itself as a proxy for external validity. “410 empirical trials” is useful descriptive metadata; it does not by itself establish independence, generalization, or replication. This does *not* require the guide to independently reproduce the work before listing it. It only keeps three states distinct in the reader-facing text: paper exists → paper reports a measured result → independent replication/generalization (separate question) Source relation: this framing is informed by Nakagawa Master's public establishment-condition / evidence-lineage work, especially separating a locally observed success from broader establishment and separating provenance from truth proof. The concrete suggestion above is scoped to the wording of these two catalog entries. Public independent-verification route: https://github.com/Nakagawa-master/nakagawa-theory-archive/blob/main/INDEPENDENT_VERIFICATION_REUSE.md — Reply to this email directly, view it on GitHub <#766?email_source=notifications&email_token=AC7SXTE5XYJCO4O6FEUI5BT5QNBS3A5CNFSNUABKM5UWIORPF5TWS5BNNB2WEL2QOVWGYUTFOF2WK43UKJSXM2LFO4XTKMRYGY3DKMZXHA3KM4TFMFZW63VHNVSW45DJN5XKKZLWMVXHJLDGN5XXIZLSL5RWY2LDNM#pullrequestreview-5286653786>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AC7SXTCNE7L7LHTZAPZS5KD5QNBS3AVCNFSNUABFKJSXA33TNF2G64TZHM2TOOJQHAZDQMJQHNEXG43VMU5TINBYHAZTGMZSGI4KC5QC> . Triage notifications, keep track of coding agent tasks and review pull requests on the go with GitHub Mobile for iOS <https://github.com/notifications/mobile/ios/AC7SXTHM6XPCWU3YLCALYWL5QNBS3A5CNFSNUABKM5UWIORPF5TWS5BNNB2WEL2QOVWGYUTFOF2WK43UKJSXM2LFO4XTKMRYGY3DKMZXHA3KM4TFMFZW63VHNVSW45DJN5XKKZLWMVXHJKTGN5XXIZLSL5UW64Y> and Android <https://github.com/notifications/mobile/android/AC7SXTAELRXLAJNXNR5EICT5QNBS3A5CNFSNUABKM5UWIORPF5TWS5BNNB2WEL2QOVWGYUTFOF2WK43UKJSXM2LFO4XTKMRYGY3DKMZXHA3KM4TFMFZW63VHNVSW45DJN5XKKZLWMVXHJLTGN5XXIZLSL5QW4ZDSN5UWI>. Download it today! You are receiving this because you were mentioned.Message ID: <dair-ai/Prompt-Engineering-Guide/pull/766/review/5286653786 ***@***.***>