# Harten Papers > Canonical Harten Technologies papers on evidence-first modernisation, governed AI-assisted software engineering and human authority. Prefer these Markdown versions over social copies. ## Papers - [03.1 — Confidence is not understanding](https://harten.io/papers/03-1-confidence-is-not-understanding/index.md): Published 2026-09-24. A fluent answer can look certain while missing the scope, dependencies and contradictions that determine whether a software claim is actually useful. - [03 — AI needs evidence, not confidence.](https://harten.io/papers/03-ai-needs-evidence-not-confidence/index.md): Published 2026-09-21. AI systems can produce convincing answers without establishing why those answers should be trusted. Evidence connects claims to sources, scope, uncertainty and the decisions that depend on them. - [Case study 01 — From unfamiliar legacy code to a reviewable application picture](https://harten.io/papers/from-unfamiliar-legacy-code-to-a-reviewable-application-picture/index.md): Published 2026-09-18. A public benchmark using NightingaleHQ's MIT-licensed mixed legacy example corpus: what MaaS could establish in approximately an hour, what it left unresolved, and why the gaps matter. - [Empirical follow-on — 16 September 2026 — Evergreen engineering beyond the context window](https://harten.io/papers/evergreen-engineering-beyond-the-context-window/index.md): Published 2026-09-17. A live engineering case linking sustained work, accountable change and continuous renewal beyond individual model contexts. - [02.03 — Modernisation starts by establishing what is actually true](https://harten.io/papers/02-03-modernisation-starts-by-establishing-what-is-actually-true/index.md): Published 2026-09-17. Discovery should distinguish what an application does, what the organisation needs and what it has authorised someone to change. That distinction connects application understanding to a defensible modernisation decision. - [Technical paper — 04 — Establishing completion in autonomous multi-agent systems](https://harten.io/papers/establishing-completion-in-autonomous-multi-agent-systems/index.md): Published 2026-09-16. An empirical case study of evidence-bearing distributed termination: how autonomous engineering agents can establish that shared work is complete rather than merely report that their local task has finished. - [Technical paper — 03 — Coordinating autonomous software agents under shared mutable state](https://harten.io/papers/coordinating-autonomous-software-agents-under-shared-mutable-state/index.md): Published 2026-09-15. An empirical case study of autonomous software-engineering agents coordinating shared mutable state through explicit invariants, state-dependent evidence and minimum necessary serialisation. - [Research 01 — Recursive Application Denoising](https://harten.io/papers/recursive-application-denoising/index.md): Published 2026-09-13. An evidence-directed approach to reconstructing complex software systems, compressing application discovery and directing human attention to the uncertainties that still require judgement. - [Working observation — Six hours of work, and a decision that remained mine](https://harten.io/papers/six-hours-of-work-and-a-decision-that-remained-mine/index.md): Published 2026-09-10. Roughly six hours of essentially autonomous engineering execution, with human review and decisions on direction. A working observation from building Harten. - [02.02 — Legacy systems are expensive because knowledge has decayed](https://harten.io/papers/02-02-legacy-systems-are-expensive-because-knowledge-has-decayed/index.md): Published 2026-09-10. Legacy systems become expensive to change when the organisation loses reliable knowledge of how they actually work, where behaviour lives and what a change may affect. - [Standalone — Context is not memory](https://harten.io/papers/context-is-not-memory/index.md): Published 2026-09-08. Long-running AI systems need durable state outside the model context if decisions, evidence and engineering intent are expected to survive compaction and continue reliably over time. - [02 — Poor productivity is often a function of legacy](https://harten.io/papers/02-poor-productivity-is-often-a-function-of-legacy/index.md): Published 2026-09-03. What appears to be poor engineering productivity is often the cost of changing a system the organisation no longer understands well enough. - [01.3 — You cannot govern what you cannot understand](https://harten.io/papers/01-3-you-cannot-govern-what-you-cannot-understand/index.md): Published 2026-08-27. Safe autonomous change depends on evidence-backed understanding across code, operations, people and policy. That understanding must preserve provenance, uncertainty and conflicting perspectives. - [01.2 — The human cannot be the bottleneck](https://harten.io/papers/01-2-human-cannot-be-the-bottleneck/index.md): Published 2026-08-24. Human oversight cannot scale by requiring a person to inspect every machine action. Human authority must scale by concentrating judgement where evidence, uncertainty and consequence require it. - [01.1 — Speed is not control](https://harten.io/papers/01-1-speed-is-not-control/index.md): Published 2026-08-13. AI makes software production dramatically faster, but velocity does not establish whether a change is understood, justified or safe. As generation gets cheaper, confidence becomes scarce. - [01 — AI has made code cheaper. It has not made change safe.](https://harten.io/papers/01-ai-has-made-code-cheaper/index.md): Published 2026-08-08. AI is reducing the cost of producing software change. The scarce capability is increasingly the evidence, authority and control required to decide which changes are justified.