research-ledger
An MIT-licensed open-source tool for research discipline with AI assistance, distributed as a folder of SKILL.md files. Through processes such as sieving, translation, preregistration and blind-test rounds, it helps researchers avoid self-deception, and records each idea under one of five verdicts, including DEAD, LANDED and BRIDGE, in an append-only ledger.
Timeline (last 90 days)
research-ledger is an MIT-licensed open-source tool in the form of a SKILL.md folder that turns an AI assistant capable of reading skills into a disciplined research partner.
research-ledger can be obtained via git clone https://github.com/Freeky7819/research-ledger and copied to ~/.claude/skills/ for use.
research-ledger also supports running in Claude.ai (uploading the compressed folder) and Codex (the folder contains agents/openai.yaml).
research-ledger has five verdict types in total: DEAD, LANDED, BRIDGE CANDIDATE, BRIDGE, EXPLANATION.
research-ledger requires that preregistration freeze the predictions, thresholds, control groups, and statistical power, with a separate PREREG commit pushed to Git before the run serving as timestamp proof.
research-ledger stipulates that pilot tests may adjust parameters but can never give a PASS verdict.
In a numerology experiment, about 3700 formulas could fit both the real measured ratios and randomly fabricated fake target ratios, showing that these fits do not constitute valid evidence.
In a continual learning test, the noise control group was algebraically identical to the treatment group, rendering the test statistically powerless.
research-ledger's blind-test rounds are triggered at milestone points, sending the same question seed to models from different labs and different vendors, conducted in clean conversations with no project context, to strip away the author's own framing but unable to eliminate the bias of shared training data among models.
The research-ledger ledger has an append-only structure, retaining each failure record and the applicable conditions of its lesson.
1 more timeline events
All Facts (11)
research-ledger's rules make clear that it can only reduce self-deception and cannot produce scientific discoveries.
50%Unverifiedresearch-ledger can be obtained via git clone https://github.com/Freeky7819/research-ledger and copied to ~/.claude/skills/ for use.
50%Unverifiedresearch-ledger also supports running in Claude.ai (uploading the compressed folder) and Codex (the folder contains agents/openai.yaml).
50%Unverifiedresearch-ledger has five verdict types in total: DEAD, LANDED, BRIDGE CANDIDATE, BRIDGE, EXPLANATION.
50%Unverifiedresearch-ledger requires that preregistration freeze the predictions, thresholds, control groups, and statistical power, with a separate PREREG commit pushed to Git before the run serving as timestamp proof.
50%Unverifiedresearch-ledger stipulates that pilot tests may adjust parameters but can never give a PASS verdict.
50%Unverifiedresearch-ledger's blind-test rounds are triggered at milestone points, sending the same question seed to models from different labs and different vendors, conducted in clean conversations with no project context, to strip away the author's own framing but unable to eliminate the bias of shared training data among models.
50%UnverifiedThe research-ledger ledger has an append-only structure, retaining each failure record and the applicable conditions of its lesson.
50%Unverifiedresearch-ledger is an MIT-licensed open-source tool in the form of a SKILL.md folder that turns an AI assistant capable of reading skills into a disciplined research partner.
50%UnverifiedIn a numerology experiment, about 3700 formulas could fit both the real measured ratios and randomly fabricated fake target ratios, showing that these fits do not constitute valid evidence.
50%UnverifiedIn a continual learning test, the noise control group was algebraically identical to the treatment group, rendering the test statistically powerless.
50%