falsifiability · research archive

falsifiability research

7 posts tagged falsifiability — essays and experiments on falsifiability, drawn from the working notes behind computational sovereignty.

7 posts tagged falsifiability clear

Aug 9, 2026Hermes Atlas v0.1: A Local-First, Incrementally-Compiled, Falsifiable Knowledge SystemHermes Atlas v0.1 is a local-first, incrementally-compiled, falsifiable knowledge system. This is a full replication guide: every command was run for real, and the failure modes are the findings.sovereign-ailocal-aiknowledge-graphgraphragfalsifiabilitylocal-first
Aug 9, 2026Hermes Atlas v0.2: Claim-Level Coverage, a Determinism Gate, and a Knowledge-Mutation CanaryHermes Atlas v0.2 is a benchmark-and-fix release: claim-level coverage localisation, a determinism gate that proves compile reproducibility, and a knowledge-mutation canary. Every number below was measured, not asserted.sovereign-ailocal-aiknowledge-graphgraphragfalsifiabilitylocal-first
Aug 9, 2026Hermes Atlas v0.3: An Evidence Ledger, a Provable Claim Lifecycle, and an Explain CommandHermes Atlas v0.3 adds an evidence ledger to every claim, a provable claim lifecycle (including DELETE-driven invalidation that retains history), and an `atlas explain` command. Every number below was measured, not asserted.sovereign-ailocal-aiknowledge-graphgraphragfalsifiabilitylocal-first
Aug 9, 2026Hermes Atlas v0.4: Proving the Evidence Ledger & Claim Lifecycle with a Gold-Labeled BenchmarkHermes Atlas v0.4 proves the v0.3 evidence ledger and provable claim lifecycle with a gold-labeled benchmark. Every number was measured on a fixed, model-free corpus; two real bugs were found and fixed in the process. Everything below is reproducible.sovereign-ailocal-aiknowledge-graphgraphragfalsifiabilitylocal-first
Jul 26, 2026Does the Compiler Know When It's Wrong? Semantic Rediscovery and the Limits of Falsification-AwarenessPopper-ORKG ended with two honest failures: temporal rediscovery was near-zero, and the compiler confidently rebuilt Lamarckian inheritance. v4 exists to ask whether those are fixable weaknesses or fundamental limits — by adding measurement, not by tuning the scorer. The answer turned out to be one of each. Mechanism-level semantic matching shows rediscovery was a measurement artifact all along (0.63 vs 0.005 exact-match, flat across temporal gaps). But a falsification-awareness consensus check could not downgrade Lamarckism or caloric: the ORKG index contains no detectable refutation edges, so the near-miss is a genuine limit of the knowledge substrate. Everything in popper/ and orkg/ stayed byte-for-byte frozen, md5-verified. Run locally on a 35B model.ai-agentscompile-time-aiknowledge-compilerlocal-first-aiscientific-discoveryfalsifiability
Jul 25, 2026Does the Hypothesis Compiler Generalize? Running Popper on a Corpus I Did Not BuildPopper beat its controls on five hand-curated benchmark literatures — but I built those corpora, so the result could be an artifact of clean data. This experiment removes that objection: the identical pipeline, byte-for-byte, run on a corpus I did not build — the Open Research Knowledge Graph, 6.3 million triples across 65,689 papers and seven scientific disciplines. Same schema, same rubric, same four-way controls; the only independent variable is the corpus. The compiler still wins (0.85 composite vs 0.56 keyword vs 0.00 ungrounded), an ablation proves the architecture is load-bearing (remove graph reasoning and it collapses to zero), and the grounding audit finds zero hallucinated citations across 265 evidence items. It also fails honestly: exact temporal rediscovery is near-zero, and it confidently rebuilds Lamarckian inheritance. Run locally on a 35B model.ai-agentscompile-time-aiknowledge-compilerlocal-first-aiscientific-discoveryfalsifiability
Jul 24, 2026Popper: Making the Hypothesis Compiler Falsifiable — Rediscovery Benchmarks, Negative Controls, and the Move From Plausible to ProvableThe last experiment proved a knowledge-graph compiler survives scale. This one asks a harder question: are its hypotheses scientifically meaningful, or only linguistically convincing? Popper stops trying to make the compiler smarter and starts making it falsifiable — a strict hypothesis schema, source-level provenance, three negative controls, and a historical rediscovery benchmark where the discovery paper is removed and the system must reconstruct the leap from prior literature alone. Run locally on a 35B model: the compiler scored 0.89–0.97 composite against 0.00 for ungrounded generation, and reconstructed the CTLA-4 checkpoint-blockade and nucleoside-modified-mRNA discoveries from their prior literature — while honestly failing on CRISPR and refusing to confidently rebuild phlogiston.ai-agentscompile-time-aiknowledge-compilerlocal-first-aiscientific-discoveryfalsifiability

Full archive — all 184 posts