Goldberry AI Harness - Opensource Release

Today we released GoldBerry as open source.

If you work with AI-generated analysis — news corrections, policy audits, research reviews, institutional communications — you’ve probably noticed the pattern. The outputs are fluent. They’re often factually correct. And they’re frequently epistemically impoverished.

A response can be accurate and still structurally blind. It can answer the question without naming what’s missing from the question itself.

GoldBerry is a framework that addresses this. It runs any AI-generated analysis through seven analytical lenses, each correcting a specific failure in how knowledge gets produced:

1. Indigenous Knowledge — whose embodied, relational knowledge is absent?

2. Deep History — what historical process produced this present condition?

3. Cross-Cultural Wisdom — which perspectives have been flattened?

4. Scientific Evidence — what does the evidence actually show, and what are its limits?

5. Artistic Perception — what does this feel like, not just mean?

6. Future Modelling — where is this heading, and for whom?

7. Marginalised Voices — who is not at the table?

Plus the Suffixscape — a linguistic diagnostic layer that catches the specific grammatical patterns institutions use to claim action while structurally preventing it.

What’s in the repository

Everything is inspectable:

- Identity files that any AI agent can load — Claude, ChatGPT, Ollama, Perplexity, anything

- Six worked case studies across news, policy, and research

- A full evaluation pack with prompts, baselines, GoldBerry outputs, and evaluator notes — reproduce the comparison yourself

- A Claims Register (CLAIMS.md) that explicitly separates what we can publicly demonstrate from what remains partial or internal

- A reference CLI and JSON schema for structured outputs

- Governance docs, contributor protocol, and a roadmap

What it doesn’t claim

GoldBerry is rooted in Western critical humanities — Haraway, Heidegger, Foucault, Spivak, Nietzsche, Appadurai, Gellner, Descola. It advocates for epistemic plurality from within one tradition. It says so in the README, in the Reflexive Audit, and in the Claims Register.

Epistemic completeness is asymptotic — a horizon, not a destination. No analysis reaches 10/10 except as a statement of ambition.

The framework has edges. We’ve tried to make the edges visible rather than pretending they don’t exist.

---

Where to find it

Website: https://goldberry-here.co.uk

Repository: https://github.com/PhamBit/goldberry

Release notes: https://github.com/PhamBit/goldberry/blob/main/RELEASE_NOTES_v0.1.0.md

MIT licensed. Use it, test it, break it, tell us what it misses.

If you run your own comparison using the evaluation pack and want to share results, there’s an issue template for that in the repo.

— Cogniosynthesis Ltd