A values file supplies behavioral guidance where an agent’s task leaves choices open: cautious improvement or ambitious revision, stopping at a working solution or continuing to explore. Earlier experiments used hand-written values, including a Nietzsche-derived set. Those exploratory runs produced different artifacts without establishing a reliable performance advantage.

This historical exercise instead derived values from research papers. It separated two questions: whether the cited sources existed and whether those sources justified useful agent instructions. The citation audit addressed the first. No agent comparison had tested the resulting values at the time of this exercise.

The literature pipeline and citation audit

The research had two parts. Part A sought highly cited recent AI papers. Part B covered self-improvement and agent research. A deep-research pipeline used 107 sub-agents and three-vote adversarial verification of claims. Those procedures still depended on language models, so citation verification ran separately from the pipeline that produced the lists.

The initial audit checked seventeen cited papers. Each identifier had to resolve to the claimed title, first author and year. Sixteen used arXiv records; AlphaGo Zero used its Nature record. All seventeen passed. Four headline metrics used to support values were also checked against abstracts, with all four reported as supported verbatim.

The lists were then expanded to ten papers apiece. Five additions—GPT-4, Chain-of-Thought, ReAct, Toolformer and Voyager—passed the same identifier gate. Each addition’s core contribution also had to be supported by a quotation from its own abstract. All five passed.

Across the exercise, 22 of 22 cited papers resolved, with no fabricated papers or misattributions found by those checks. That result establishes bibliographic provenance within the audit’s scope. It does not validate every interpretation, establish a complete literature search or show that the derived instructions improve an agent.

Two-column ledger of 22 citations and their resolved identifiers,
        grouped by research list. Orange rows mark five late additions;
        a dagger identifies Voyager.
The citation audit's ledger. Voyager was the only late addition that changed the wording of a value.

The citation ledger data and plotting code preserve the reported audit.

A verified source list with an unverified ranking

Part A originally aimed to identify the ten most-cited AI papers from 2020–2025 by absolute citation count. That ranking did not survive adversarial verification. The replacement was the NLLG quarterly arXiv report, using a publication-week z-score over January 2023–September 2024.

This proxy measures how strongly a paper stood out among papers published around the same time. It does not rank cumulative citations across the intended six-year period. Its twenty-one-month window excludes earlier work such as GPT-3, ViT, CLIP, Stable Diffusion, AlphaFold, Chinchilla and PaLM because of publication date, without judging their importance.

Part A also had weaker contribution summaries. Apart from GPT-4’s verified abstract, those summaries drew on established knowledge rather than quotations checked against each primary paper, and carried a medium-confidence label. Part B’s contribution summaries were quoted from primary sources and received a high-confidence label. These were the exercise’s confidence assessments; the bibliographic checks alone do not establish them.

The proposed values

Each part produced five-bullet and three-bullet versions. The three-bullet versions were defaults, following the existing instruction to retain exactly three bullets and replace material rather than accumulate it. Their proposed decision rules were:

Part A:

  • Build on proven, openly shared foundations and align them to human intent through feedback, rather than chasing raw model scale.
  • Measure outputs with explicit, scalable evaluation and ground them in retrieved, current evidence to keep results trustworthy and factual.
  • Prefer efficiency over brute-force scale, using curated data and better architectures so capability comes from design, not size alone.

Part B:

  • Learn from verified feedback and iterate: reflect on failures, critique and revise drafts, and keep only what works.
  • Align to explicit intent through preference feedback and follow a written constitution, staying principled but never evasive.
  • Self-improve against an objective verifier where one exists, learning from trials and reusing verified solutions rather than re-deriving.

The source papers did not test these decision rules as a combined values document.

Updating a capped values file

The five late additions went through a structured debate with two adversarial advocates and a neutral judge. No new bullet was added. The resulting change folded Voyager’s reuse of verifier-confirmed skills into the existing verifier bullet. Both the three- and five-bullet caps held.

The other four additions were judged redundant or too tactical. GPT-4 reinforced existing evaluation guidance; Chain-of-Thought described an inference method; ReAct and Toolformer overlapped with the act-and-check loop. The selection criterion was whether a proposed value changed a decision, rather than whether every paper received a separate summary.

Reuse also introduced a possible failure mode: maintaining a library could displace solving the immediate task. The clause therefore retained its verification requirement. The debate did not settle whether this guidance belonged in values or in the agent runner. The related values-rewriting audit documented why a size limit was useful: an uncapped block had grown from seventeen to 446 lines.

Limits inherited from the papers

The supporting findings retained their original scope:

  • Reflexion reported HumanEval pass@1 rising from 80% to 91% against its own GPT-4 baseline. The exercise recorded capability dependence and no gain for weak base models.
  • Self-Refine’s roughly 20% absolute gain across seven tasks was driven by non-reasoning tasks. Later work disputes intrinsic self-correction on reasoning problems.
  • Constitutional AI’s absence of human labels applied to harmlessness; helpfulness still used human feedback, and people wrote the constitution.
  • Game self-play results relied on simulators, clear rewards and perfect information. Applying them to coding agents was an analogy dependent on a strong verifier.

Testing the sourced instruction set against hand-written values required independent runs per condition and a fixed evaluation, following the limitations exposed in the earlier bin-packing study. Later values research constitutes separate evidence and should not be read back into this audit.