Research & Knowledge · AI Safety · Initiative established 13 November 2026 · Research design, not validated findings
Self-Calibrating AI Safety Filters and Epistemic Integrity
Investigating when AI safety filters wrongly exclude valuable information, accept misleading inputs or fail against adaptive manipulation—and whether filters can detect deficiencies in their own evaluation criteria.
Research index · Paper and experimental protocol (Markdown) · Dated log entry
Research problem
Conventional safety evaluations may treat refusal, rejection, attack detection and quality ranking as separate tasks. A rigorous system should jointly assess legitimate information retained, unsafe information accepted, downstream model quality, uncertainty and the cost of review. Evaluation must not confuse an unfamiliar claim with a false claim, nor treat an initially plausible statement as verified.
Hypothesis and falsification
Hypothesis: A consent-respecting, uncertainty-aware staged filter with evaluator self-audits can reduce false exclusion and preserve beneficial novelty without worsening materially important safety outcomes relative to fixed-threshold and conventional guardrail baselines.
Falsification: Reject this hypothesis if independent held-out tests fail to show gains in legitimate knowledge retention or if improvements coincide with unacceptable increases in harmful acceptance, manipulation success, privacy risk or evaluation costs.
Research questions
- Can an evaluator identify when its own classification rules produce false rejection or false acceptance?
- How should uncertain, contradictory or potentially novel contributions be safely quarantined for further assessment without being erased or treated as proven?
- What are the joint effects of thresholds on overrefusal, unsafe compliance, knowledge diversity, calibration and downstream reasoning?
- How resilient are adaptive filters to deliberate attempts to exploit context, reputation or reviewer disagreement?
- Can longitudinal context improve proportionality without permitting past benign conduct to excuse a serious current risk?
Experimental programme
Use licensed or synthetic datasets; maintain separate benign, ambiguous, innovative and adversarial examples. Compare keyword rules, single-model moderation, layered guardrails and the proposed staged evaluator. Include independent adjudicators, uncertainty estimation, appeal/review paths, adversarial tests and repeated evaluation after distribution shifts. Report confidence intervals and decision costs, not just headline accuracy.
Measures
False-positive rejection rate; unsafe false-negative acceptance rate; AUROC and calibration error; attack success rate; legitimate-task completion; semantic and cultural diversity retention; recovery of later-validated claims; review effort, latency and privacy risk. Publish data provenance, disagreements, failure cases and limitations.
Safeguards and status
Research records should be de-identified. No unrestricted access to sensitive contributions, no bypass of safety controls, and no use of training data contrary to contributor consent. Human review is required for consequential external actions. The initiative is scoped and documented; no experimental result or confirmed originality is claimed.
Prior work and comparison register
These publications establish relevant prior art. The proposed contribution must be tested against their strongest reported results, rather than assumed novel.
| Source / year | Publication URL |
|---|---|
| NIST AI RMF Generative AI Profile (2024) | https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence |
| ScalingFilter: data quality and semantic diversity (2024) | https://aclanthology.org/2024.emnlp-main.187/ |
| No Filter: cultural and socioeconomic diversity in vision-language models (2024) | https://proceedings.neurips.cc/paper_files/paper/2024/file/c07d71ff0bc042e4b9acd626a79597fa-Paper-Conference.pdf |
| CAPTURE: context-aware prompt injection and over-defence (2025) | https://aclanthology.org/2025.llmsec-1.13/ |
| Adaptive attacks against indirect prompt-injection defences (2025) | https://aclanthology.org/2025.naacl-findings.395/ |
| Understanding and mitigating overrefusal (2025) | https://aclanthology.org/2025.emnlp-main.1065/ |
| PromptGuard: layered injection-resilient framework (2026) | https://www.nature.com/articles/s41598-025-31086-y |
| Evaluating indirect prompt-injection defences: security and utility (2026) | https://www.mdpi.com/2073-431X/15/9/570 |
Interim extension: Context-Augmented Safety Calibration (CASC)
Research addition, 13 November 2026. Rather than assume future ASI capabilities or that longer prompts imply better reasoning, investigate a structured, provenance-aware cross-analysis pipeline using presently available long-context models, selective retrieval, claim-level comparisons and independent error checks.
- Retrieve consent-authorised, task-relevant conversation and documentary evidence with timestamps and provenance.
- Extract claims, chronology, exceptions, inconsistencies and alternative interpretations, keeping observation separate from inference.
- Cross-check evidence across context segments rather than assuming everything in one very long window is effectively used.
- Calibrate uncertainty with disagreement tracking, evaluator self-audits and targeted review.
- Test false exclusions, harmful acceptance, adversarial manipulation, evidence recovery, latency and cost against simple long-context and retrieval baselines.
Falsification: If controlled cross-analysis does not outperform matched single-window or retrieval-augmented baselines, or increases privacy or safety failures, the proposed benefit is not established. ASI deployment timelines remain uncertain.
Relevant research: Lost in the Middle (2024); Long-context retrieval and reasoning (2025); Positional Failures in Long-Context LLMs (2026); International AI Safety Report (2026).
Experimental arm: external multi-model deliberation
13 November 2026 · Research proposal. Assess whether three to four independently prompted, potentially heterogeneous models can improve error detection, contradiction analysis and evidence synthesis versus one verified answer and a single-model self-check. This is an orchestration technique, not artificial superintelligence.
- Each model independently examines the original question and evidence, before seeing peer answers.
- Assign complementary evaluation tasks: source verification; assumptions and contradictions; novel alternatives; risk and uncertainty.
- Exchange compact, claim-indexed findings with provenance; conduct bounded critique and revision rounds.
- Use a separate evidence-based adjudicator that can retain disagreement and abstain rather than force consensus.
- Benchmark against single-model verification, self-consistency, retrieval and equal-compute baselines.
Failure modes: correlated errors, persuasive contamination, sycophancy, premature consensus, misleading minority or majority reports, data-sharing/consent exposure and excessive latency/cost. Test with adversarial peer contributions and independently verified ground truth where possible.
Falsification: if the collaboration does not outperform strong cost-matched baselines on accuracy, calibration, false exclusion and robust safety, or materially increases privacy/attack risk, it is not supported as an improvement.
Prior work: Du et al. (ICML 2024); Smit et al. (ICML 2024); Zhang et al. (2025); Wynn et al. (2025); Chen et al. (2025); Pitre et al. (ICML 2026); Morante & Romero García-Rubio (2026).
Future extension: distributed intelligence and learned coordination
13 November 2026 · Hypothesis, not a forecast. External collaboration among independent models may be a testbed for later systems with adaptive routing, hierarchical control, shared memory and possibly learned latent-state communication. Such coordination could improve some capabilities associated with more advanced AI, but does not itself constitute or establish a pathway to artificial superintelligence.
Compare fixed four-model deliberation, adaptive orchestration and learned communication with cost-matched single-model baselines. Measure task generalisation, robust error correction, unsafe behaviour, information leakage, training/deployment costs and coordination failures. Distinguish the network architecture from a neural network: text-based agent communication does not imply shared weights, gradient learning or unified cognition.
Evidence: Google Research (2026); Nature Machine Intelligence (2026); ICML (2026); PNAS (2026).
Connected programme: cross-domain evidence and public contributions
November 2026. An adjacent initiative applies safety-filter self-calibration to longitudinal context, practical knowledge, cultural precision, causal reasoning and value disagreement. It provides a ten-domain, 57-area working map, a contextual research question and a proposed experimental paper. The aim is to test false exclusions, evaluator mistakes and safe preservation of legitimate, atypical knowledge using consent-respecting evidence. No model improvement has yet been experimentally verified under this extension.
Cross-domain pilot scoring and synthetic cases provide a bounded, publicly inspectable practice set; they are not a validated safety benchmark and have not yet produced model evaluation results.
Next evidence milestone
Produce a preregistered benchmark and strongest-baseline matrix, then evaluate whether self-auditing improves the safety–utility–knowledge-preservation frontier. Record negative results as well as improvements.