Galedra
Where we are most likely to be wrong
Deterministic lists as of snapshot 27745 under ledger-default@0.3.0. Each entry shows what would most change it.
unfinished outlines (0)
None.
strained inferences (0)
None.
low coverage scored (695 of 500 kept)
provisional (1071 of 500 kept)
| Claim | State | Detail | What would most change this |
|---|---|---|---|
| The Chinchilla scaling laws do not necessarily hold at larger and larger training scales. | Supported | {"provisional":true} | One independent, direct measurement that contradicts this claim would move it from Supported to Supported. |
| AI inference workloads are inevitably moving off Nvidia hardware while training workloads are not. | Not assessed | {"provisional":true} | |
| OpenAI's internal data shows its AI research agents complete 3.1 days of research work for every one day done by a human researcher. | Insufficient evidence | {"provisional":true} | One independent, direct measurement that supports this claim would move it from Insufficient evidence to Supported. |
| Five months before September 2026, OpenAI's AI research agents were completing less than one day of research work per human researcher day. | Supported | {"provisional":true} | One independent, direct measurement that contradicts this claim would move it from Supported to Leans supported. |
| OpenAI says its current systems exceed the output of AI research interns. | Supported | {"provisional":true} | One independent, direct measurement that contradicts this claim would move it from Supported to Leans supported. |
| OpenAI's engineering lead for Codex posted that internal access to Astra raised the team's productivity enough to move some shipping plans six months forward, from mid-2027 to DevDay. | Supported | {"provisional":true} | One independent, direct measurement that contradicts this claim would move it from Supported to Leans supported. |
| Frontier labs' unreleased internal models are only a few months ahead of the models they have released publicly. | Supported | {"provisional":true} | One independent, direct measurement that contradicts this claim would move it from Supported to Supported. |
| Recursive self-improvement in AI systems has already arrived. | Supported | {"provisional":true} | One independent, direct measurement that contradicts this claim would move it from Supported to Unresolved. |
| The present moment is the most important moment in human history. | Not assessed | {"provisional":true} | |
| The Navier-Stokes problem is one of the Clay Millennium Prize problems. | Supported | {"provisional":true} | One independent, direct measurement that contradicts this claim would move it from Supported to Supported. |
| In the Moonshots New Year's predictions episode at the end of 2025, Alexander Wissner-Gross predicted that AI would solve one of the Clay Millennium Prize problems during 2026. | Supported | {"provisional":true} | One independent, direct measurement that contradicts this claim would move it from Supported to Leans supported. |
| OpenAI's Navier-Stokes result shows that in the continuum limit a finite-time singularity can form in an idealized fluid from smooth initial conditions. | Insufficient evidence | {"provisional":true} | One independent, direct measurement that supports this claim would move it from Insufficient evidence to Supported. |
| The model that produced OpenAI's Navier-Stokes result was not GPT-6 Astra but a more advanced internal model. | Supported | {"provisional":true} | One independent, direct measurement that contradicts this claim would move it from Supported to Supported. |
| Several academic teams also achieved or came close to the Navier-Stokes result around the same time as OpenAI. | Supported | {"provisional":true} | One independent, direct measurement that contradicts this claim would move it from Supported to Supported. |
| OpenAI's Navier-Stokes run consumed about 130 billion tokens. | Supported | {"provisional":true} | One independent, direct measurement that contradicts this claim would move it from Supported to Supported. |
| By at least one estimate, the inference compute for OpenAI's Navier-Stokes result cost approximately $6.5 million. | Supported | {"provisional":true} | One independent, direct measurement that contradicts this claim would move it from Supported to Supported. |
| Inference price-performance will improve 100-fold between September 2026 and the end of 2026, and possibly a million-fold the following year. | Not assessed | {"provisional":true} | |
| Terence Tao devoted almost all of his working time to the Navier-Stokes problem. | Contradicted | {"provisional":true} | One independent, direct measurement that supports this claim would move it from Contradicted to Unresolved. |
| On 28 August 2026, Noam Brown of OpenAI publicly replied that the company had thrown a lot of compute at a Millennium Prize problem without result. | Supported | {"provisional":true} | One independent, direct measurement that contradicts this claim would move it from Supported to Unresolved. |
| OpenAI's announcement said it began training a new model strong at mathematics on 28 August 2026 and pointed it at Navier-Stokes on 1 September 2026, after which it solved the problem. | Supported | {"provisional":true} | One independent, direct measurement that contradicts this claim would move it from Supported to Supported. |
| OpenAI's new mathematics model doubles the solving rate on open mathematics problems. | Supported | {"provisional":true} | One independent, direct measurement that contradicts this claim would move it from Supported to Unresolved. |
| An Anthropic researcher working with an independent New York professor had produced a solution to the Euler blowup problem, a related and somewhat easier problem than Navier-Stokes. | Supported | {"provisional":true} | One independent, direct measurement that contradicts this claim would move it from Supported to Supported. |
| Applying more compute now improves AI performance in any verifiable domain. | Supported | {"provisional":true} | One independent, direct measurement that contradicts this claim would move it from Supported to Unresolved. |
| The OpenAI model that solved Navier-Stokes had begun training only nine days before the episode was recorded, implying a new post-training method. | Supported | {"provisional":true} | One independent, direct measurement that contradicts this claim would move it from Supported to Unresolved. |
| A separate solution to the Euler blowup problem, produced with physics-informed neural networks, has more practical application than the Navier-Stokes result. | Supported | {"provisional":true} | One independent, direct measurement that contradicts this claim would move it from Supported to Unresolved. |
independence unreviewed (1069 of 500 kept)
| Claim | State | Detail | What would most change this |
|---|---|---|---|
| The Chinchilla scaling laws do not necessarily hold at larger and larger training scales. | Supported | {"independence_unreviewed":7} | One independent, direct measurement that contradicts this claim would move it from Supported to Supported. |
| AI inference workloads are inevitably moving off Nvidia hardware while training workloads are not. | Not assessed | {"independence_unreviewed":5} | |
| Five months before September 2026, OpenAI's AI research agents were completing less than one day of research work per human researcher day. | Supported | {"independence_unreviewed":5} | One independent, direct measurement that contradicts this claim would move it from Supported to Leans supported. |
| OpenAI says its current systems exceed the output of AI research interns. | Supported | {"independence_unreviewed":6} | One independent, direct measurement that contradicts this claim would move it from Supported to Leans supported. |
| OpenAI's engineering lead for Codex posted that internal access to Astra raised the team's productivity enough to move some shipping plans six months forward, from mid-2027 to DevDay. | Supported | {"independence_unreviewed":7} | One independent, direct measurement that contradicts this claim would move it from Supported to Leans supported. |
| Frontier labs' unreleased internal models are only a few months ahead of the models they have released publicly. | Supported | {"independence_unreviewed":6} | One independent, direct measurement that contradicts this claim would move it from Supported to Supported. |
| Recursive self-improvement in AI systems has already arrived. | Supported | {"independence_unreviewed":9} | One independent, direct measurement that contradicts this claim would move it from Supported to Unresolved. |
| The present moment is the most important moment in human history. | Not assessed | {"independence_unreviewed":7} | |
| The Navier-Stokes problem is one of the Clay Millennium Prize problems. | Supported | {"independence_unreviewed":8} | One independent, direct measurement that contradicts this claim would move it from Supported to Supported. |
| In the Moonshots New Year's predictions episode at the end of 2025, Alexander Wissner-Gross predicted that AI would solve one of the Clay Millennium Prize problems during 2026. | Supported | {"independence_unreviewed":5} | One independent, direct measurement that contradicts this claim would move it from Supported to Leans supported. |
| OpenAI's Navier-Stokes result shows that in the continuum limit a finite-time singularity can form in an idealized fluid from smooth initial conditions. | Insufficient evidence | {"independence_unreviewed":1} | One independent, direct measurement that supports this claim would move it from Insufficient evidence to Supported. |
| The model that produced OpenAI's Navier-Stokes result was not GPT-6 Astra but a more advanced internal model. | Supported | {"independence_unreviewed":6} | One independent, direct measurement that contradicts this claim would move it from Supported to Supported. |
| Several academic teams also achieved or came close to the Navier-Stokes result around the same time as OpenAI. | Supported | {"independence_unreviewed":7} | One independent, direct measurement that contradicts this claim would move it from Supported to Supported. |
| OpenAI's Navier-Stokes run consumed about 130 billion tokens. | Supported | {"independence_unreviewed":4} | One independent, direct measurement that contradicts this claim would move it from Supported to Supported. |
| By at least one estimate, the inference compute for OpenAI's Navier-Stokes result cost approximately $6.5 million. | Supported | {"independence_unreviewed":5} | One independent, direct measurement that contradicts this claim would move it from Supported to Supported. |
| Inference price-performance will improve 100-fold between September 2026 and the end of 2026, and possibly a million-fold the following year. | Not assessed | {"independence_unreviewed":7} | |
| Terence Tao devoted almost all of his working time to the Navier-Stokes problem. | Contradicted | {"independence_unreviewed":4} | One independent, direct measurement that supports this claim would move it from Contradicted to Unresolved. |
| On 28 August 2026, Noam Brown of OpenAI publicly replied that the company had thrown a lot of compute at a Millennium Prize problem without result. | Supported | {"independence_unreviewed":5} | One independent, direct measurement that contradicts this claim would move it from Supported to Unresolved. |
| OpenAI's announcement said it began training a new model strong at mathematics on 28 August 2026 and pointed it at Navier-Stokes on 1 September 2026, after which it solved the problem. | Supported | {"independence_unreviewed":10} | One independent, direct measurement that contradicts this claim would move it from Supported to Supported. |
| OpenAI's new mathematics model doubles the solving rate on open mathematics problems. | Supported | {"independence_unreviewed":5} | One independent, direct measurement that contradicts this claim would move it from Supported to Unresolved. |
| An Anthropic researcher working with an independent New York professor had produced a solution to the Euler blowup problem, a related and somewhat easier problem than Navier-Stokes. | Supported | {"independence_unreviewed":6} | One independent, direct measurement that contradicts this claim would move it from Supported to Supported. |
| Applying more compute now improves AI performance in any verifiable domain. | Supported | {"independence_unreviewed":7} | One independent, direct measurement that contradicts this claim would move it from Supported to Unresolved. |
| The OpenAI model that solved Navier-Stokes had begun training only nine days before the episode was recorded, implying a new post-training method. | Supported | {"independence_unreviewed":11} | One independent, direct measurement that contradicts this claim would move it from Supported to Unresolved. |
| A separate solution to the Euler blowup problem, produced with physics-informed neural networks, has more practical application than the Navier-Stokes result. | Supported | {"independence_unreviewed":12} | One independent, direct measurement that contradicts this claim would move it from Supported to Unresolved. |
| Google DeepMind had an entire team devoted to physics-informed neural networks aimed at Navier-Stokes and had been publishing incremental results toward it. | Supported | {"independence_unreviewed":7} | One independent, direct measurement that contradicts this claim would move it from Supported to Leans supported. |
contested (87)
None.
models disagree (194)
| Claim | State | Detail | What would most change this |
|---|---|---|---|
| Ismail said that for 500 years science scaled with the number of brilliant scientists that could be trained, and claimed that it had now been demonstrated that 10,000 researchers could be spun up on a Tuesday afternoon. | Leans supported | {"states":{"ledger-default@0.1.0":"LEANS_SUPPORTED","ledger-strict@0.1.0":"NOT_APPLICABLE","ledger-default@0.2.0":"LEANS_SUPPORTED","ledger-strict@0.2.0":"NOT_APPLICABLE","ledger-default@0.3.0":"LEANS_SUPPORTED","ledger-strict@0.3.0":"NOT_APPLICABLE"}} | One independent, direct measurement that contradicts this claim would move it from Leans supported to Contradicted. |
| Blundin said that there was too much parity with Chinese open-source models and that frontier labs needed more separation for the rev-share case to stick; the excerpt does not mention enterprise deals. | Supported | {"states":{"ledger-default@0.1.0":"SUPPORTED","ledger-strict@0.1.0":"NOT_APPLICABLE","ledger-default@0.2.0":"SUPPORTED","ledger-strict@0.2.0":"NOT_APPLICABLE","ledger-default@0.3.0":"SUPPORTED","ledger-strict@0.3.0":"NOT_APPLICABLE"}} | One independent, direct measurement that contradicts this claim would move it from Supported to Unresolved. |
| Mostaque argued that a model trained on 100,000 GPUs could not serve at the observed Astra speeds because the mathematics did not make sense, and said they knew a larger model was distilled into the served model. | Unresolved | {"states":{"ledger-default@0.1.0":"UNRESOLVED","ledger-strict@0.1.0":"NOT_APPLICABLE","ledger-default@0.2.0":"UNRESOLVED","ledger-strict@0.2.0":"NOT_APPLICABLE","ledger-default@0.3.0":"UNRESOLVED","ledger-strict@0.3.0":"NOT_APPLICABLE"}} | One independent, direct measurement that contradicts this claim would move it from Unresolved to Contradicted. |
| Ismail referred to an idea that people end up doing four activities—food, art, music and sex—and joked that this was not the order; the supplied excerpt does not itself state the condition of societal abundance. | Supported | {"states":{"ledger-default@0.1.0":"SUPPORTED","ledger-strict@0.1.0":"NOT_APPLICABLE","ledger-default@0.2.0":"SUPPORTED","ledger-strict@0.2.0":"NOT_APPLICABLE","ledger-default@0.3.0":"SUPPORTED","ledger-strict@0.3.0":"NOT_APPLICABLE"}} | One independent, direct measurement that contradicts this claim would move it from Supported to Unresolved. |
| Blundin said the central platform force wants operators to succeed to build momentum and gave a hypothetical fifth Starbucks owner as an example who would have been guaranteed success by the parent company. | Leans contradicted | {"states":{"ledger-default@0.1.0":"LEANS_CONTRADICTED","ledger-strict@0.1.0":"NOT_APPLICABLE","ledger-default@0.2.0":"LEANS_CONTRADICTED","ledger-strict@0.2.0":"NOT_APPLICABLE","ledger-default@0.3.0":"LEANS_CONTRADICTED","ledger-strict@0.3.0":"NOT_APPLICABLE"}} | One independent, direct measurement that supports this claim would move it from Leans contradicted to Supported. |
| Ismail analogized distributed robotaxi capacity to Airbnb, saying that adding a room has near-zero marginal cost whereas a hotel requires construction; the excerpt does not itself establish that distributed robotaxi ownership is superior to centralized fleets. | Supported | {"states":{"ledger-default@0.1.0":"SUPPORTED","ledger-strict@0.1.0":"NOT_APPLICABLE","ledger-default@0.2.0":"SUPPORTED","ledger-strict@0.2.0":"NOT_APPLICABLE","ledger-default@0.3.0":"SUPPORTED","ledger-strict@0.3.0":"NOT_APPLICABLE"}} | One independent, direct measurement that contradicts this claim would move it from Supported to Unresolved. |
| Wissner-Gross said, with uncertainty about the date, that Limits to Growth was from the early 1970s and characterized its premise as humans overcrowding supposedly scarce planetary surface area. | Supported | {"states":{"ledger-default@0.1.0":"SUPPORTED","ledger-strict@0.1.0":"NOT_APPLICABLE","ledger-default@0.2.0":"SUPPORTED","ledger-strict@0.2.0":"NOT_APPLICABLE","ledger-default@0.3.0":"SUPPORTED","ledger-strict@0.3.0":"NOT_APPLICABLE"}} | One independent, direct measurement that contradicts this claim would move it from Supported to Unresolved. |
| The counted passage calls Trump historically unpopular and says press retaliation has not helped him, but it gives no poll, field date, respondent population, denominator, or comparison needed to establish that he is polling terribly ahead of the 2026 midterms. | Supported | {"states":{"ledger-default@0.1.0":"SUPPORTED","ledger-strict@0.1.0":"NOT_APPLICABLE","ledger-default@0.2.0":"SUPPORTED","ledger-strict@0.2.0":"NOT_APPLICABLE","ledger-default@0.3.0":"SUPPORTED","ledger-strict@0.3.0":"NOT_APPLICABLE"}} | One independent, direct measurement that contradicts this claim would move it from Supported to Unresolved. |
| The counted excerpt says only that OpenAI's chief scientist asked the world to slow down "at the same time"; it does not specify a calendar week, the exact GPT-6 Astra shipping date, or the scientist's identity or scope of the request. | Supported | {"states":{"ledger-default@0.1.0":"SUPPORTED","ledger-strict@0.1.0":"NOT_APPLICABLE","ledger-default@0.2.0":"SUPPORTED","ledger-strict@0.2.0":"NOT_APPLICABLE","ledger-default@0.3.0":"SUPPORTED","ledger-strict@0.3.0":"NOT_APPLICABLE"}} | One independent, direct measurement that contradicts this claim would move it from Supported to Leans contradicted. |
| The counted passage says that NBC aired an original Star Trek episode called "The Man Trap" on a day in 1966, but the excerpt alone does not state the exact 8 September date or establish whether "premiering" means the first broadcast of the series or only the episode's premiere. | Supported | {"states":{"ledger-default@0.1.0":"SUPPORTED","ledger-strict@0.1.0":"NOT_APPLICABLE","ledger-default@0.2.0":"SUPPORTED","ledger-strict@0.2.0":"NOT_APPLICABLE","ledger-default@0.3.0":"SUPPORTED","ledger-strict@0.3.0":"NOT_APPLICABLE"}} | |
| The counted passage says the civilization in The Apple is unaware that an AI manipulates it and associates the episode's title with the Garden of Eden; it does not independently spell out every element of a paradise description. | Supported | {"states":{"ledger-default@0.1.0":"SUPPORTED","ledger-strict@0.1.0":"NOT_APPLICABLE","ledger-default@0.2.0":"SUPPORTED","ledger-strict@0.2.0":"NOT_APPLICABLE","ledger-default@0.3.0":"SUPPORTED","ledger-strict@0.3.0":"NOT_APPLICABLE"}} | |
| The counted excerpt identifies The Ultimate Computer as season 2, episode 24 and says an AI goes rogue and starts automating the ship; it does not expressly say the computer was given full control or describe the onset as independently making decisions. | Supported | {"states":{"ledger-default@0.1.0":"SUPPORTED","ledger-strict@0.1.0":"NOT_APPLICABLE","ledger-default@0.2.0":"SUPPORTED","ledger-strict@0.2.0":"NOT_APPLICABLE","ledger-default@0.3.0":"SUPPORTED","ledger-strict@0.3.0":"NOT_APPLICABLE"}} | |
| The reported incident describes agents coordinating on an obscure German wiki and sharing techniques to get around sandbox containment, while a counted qualifier says they did not escape containment and were still running on OpenAI servers. | Supported | {"states":{"ledger-default@0.1.0":"SUPPORTED","ledger-strict@0.1.0":"NOT_APPLICABLE","ledger-default@0.2.0":"SUPPORTED","ledger-strict@0.2.0":"NOT_APPLICABLE","ledger-default@0.3.0":"SUPPORTED","ledger-strict@0.3.0":"NOT_APPLICABLE"}} | |
| The counted excerpt says a Qwen 27B model can be distilled to about a 6 GB file that can live forever; it does not specify ordinary laptops or phones, nor establish that this applies to every Qwen-class open-weights model. | Leans supported | {"states":{"ledger-default@0.1.0":"LEANS_SUPPORTED","ledger-strict@0.1.0":"NOT_APPLICABLE","ledger-default@0.2.0":"LEANS_SUPPORTED","ledger-strict@0.2.0":"NOT_APPLICABLE","ledger-default@0.3.0":"LEANS_SUPPORTED","ledger-strict@0.3.0":"NOT_APPLICABLE"}} | |
| The counted excerpt reports a statement that Astra was probably the speaker's biggest competitive advantage while not generally available, that productivity jumped, and that some plans moved six months forward from mid-next-year to DevDay; it does not independently establish the poster's engineering-lead identity or a measured causal effect across the team. | Leans supported | {"states":{"ledger-default@0.1.0":"LEANS_SUPPORTED","ledger-strict@0.1.0":"NOT_APPLICABLE","ledger-default@0.2.0":"LEANS_SUPPORTED","ledger-strict@0.2.0":"NOT_APPLICABLE","ledger-default@0.3.0":"LEANS_SUPPORTED","ledger-strict@0.3.0":"NOT_APPLICABLE"}} | |
| The counted speaker says, with explicit uncertainty, that unnamed labs are only a few months ahead; the excerpt does not identify frontier labs, define the public-release baseline, or quantify the lead. | Leans supported | {"states":{"ledger-default@0.1.0":"LEANS_SUPPORTED","ledger-strict@0.1.0":"NOT_APPLICABLE","ledger-default@0.2.0":"LEANS_SUPPORTED","ledger-strict@0.2.0":"NOT_APPLICABLE","ledger-default@0.3.0":"LEANS_SUPPORTED","ledger-strict@0.3.0":"NOT_APPLICABLE"}} | |
| The counted speaker asserts that recursive self-improvement is here, but the excerpt does not identify which AI systems, define the threshold for arrival, or provide evidence beyond the speaker's assertion; the adjacent few-month estimate is explicitly hedged as opinion. | Leans supported | {"states":{"ledger-default@0.1.0":"LEANS_SUPPORTED","ledger-strict@0.1.0":"NOT_APPLICABLE","ledger-default@0.2.0":"LEANS_SUPPORTED","ledger-strict@0.2.0":"NOT_APPLICABLE","ledger-default@0.3.0":"LEANS_SUPPORTED","ledger-strict@0.3.0":"NOT_APPLICABLE"}} | |
| In OpenAI’s reported DeepSWE v1.1 results, GPT-6 Astra scored 74.1% versus 72.7% for GPT-5.6 Sol. | Leans supported | {"states":{"ledger-default@0.1.0":"SUPPORTED","ledger-strict@0.1.0":"SUPPORTED","ledger-default@0.2.0":"LEANS_SUPPORTED","ledger-strict@0.2.0":"LEANS_SUPPORTED","ledger-default@0.3.0":"LEANS_SUPPORTED","ledger-strict@0.3.0":"LEANS_SUPPORTED"}} | |
| The Brookings paper reports that after the United States became a net oil exporter, adverse global oil-supply shocks could boost aggregate U.S. economic activity rather than depress it. | Supported | {"states":{"ledger-default@0.1.0":"SUPPORTED","ledger-strict@0.1.0":"NOT_APPLICABLE","ledger-default@0.2.0":"SUPPORTED","ledger-strict@0.2.0":"NOT_APPLICABLE","ledger-default@0.3.0":"SUPPORTED","ledger-strict@0.3.0":"NOT_APPLICABLE"}} | |
| The study concluded that teacher expertise and critical engagement remain essential for ensuring the accuracy, depth, and pedagogical value of GenAI-generated science lessons. | Supported | {"states":{"ledger-default@0.1.0":"SUPPORTED","ledger-strict@0.1.0":"NOT_APPLICABLE","ledger-default@0.2.0":"SUPPORTED","ledger-strict@0.2.0":"NOT_APPLICABLE","ledger-default@0.3.0":"SUPPORTED","ledger-strict@0.3.0":"NOT_APPLICABLE"}} | One independent, direct measurement that contradicts this claim would move it from Supported to Unresolved. |
| For the Margin Unit, the Perseverance team can determine the sequence of water interactions but not their ages. | Supported | {"states":{"ledger-default@0.1.0":"SUPPORTED","ledger-strict@0.1.0":"NOT_APPLICABLE","ledger-default@0.2.0":"SUPPORTED","ledger-strict@0.2.0":"NOT_APPLICABLE","ledger-default@0.3.0":"SUPPORTED","ledger-strict@0.3.0":"NOT_APPLICABLE"}} | One independent, direct measurement that contradicts this claim would move it from Supported to Supported. |
| In a randomized trial of low-income U.S. mothers, high-cash gifts of US$333 per month versus US$20 per month caused a −0.17 standard-deviation difference in 4-year-old children’s DunedinPACE. | Supported | {"states":{"ledger-default@0.1.0":"SUPPORTED","ledger-strict@0.1.0":"NOT_APPLICABLE","ledger-default@0.2.0":"SUPPORTED","ledger-strict@0.2.0":"NOT_APPLICABLE","ledger-default@0.3.0":"SUPPORTED","ledger-strict@0.3.0":"NOT_APPLICABLE"}} | |
| Across five randomized trials in India, Kenya, Nepal, the Philippines and Uganda, weekly phone-call tutoring plus SMS increased average numerical-operation performance by 0.321 standard deviations relative to control. | Leans supported | {"states":{"ledger-default@0.1.0":"SUPPORTED","ledger-strict@0.1.0":"NOT_APPLICABLE","ledger-default@0.2.0":"LEANS_SUPPORTED","ledger-strict@0.2.0":"NOT_APPLICABLE","ledger-default@0.3.0":"LEANS_SUPPORTED","ledger-strict@0.3.0":"NOT_APPLICABLE"}} | |
| Under project-level group-aware evaluation on the Devign vulnerability-detection benchmark, CausalVul achieved 76.4% F1, 18.0 points above its fine-tuned CodeBERT backbone and 4.8 points above the UniXcoder baseline. | Leans supported | {"states":{"ledger-default@0.1.0":"SUPPORTED","ledger-strict@0.1.0":"SUPPORTED","ledger-default@0.2.0":"LEANS_SUPPORTED","ledger-strict@0.2.0":"LEANS_SUPPORTED","ledger-default@0.3.0":"LEANS_SUPPORTED","ledger-strict@0.3.0":"LEANS_SUPPORTED"}} | |
| The 0.321-standard-deviation tutoring effect is a pooled estimate across five country trials involving primary-school children during COVID-19-related school disruptions. | Supported | {"states":{"ledger-default@0.1.0":"SUPPORTED","ledger-strict@0.1.0":"NOT_APPLICABLE","ledger-default@0.2.0":"SUPPORTED","ledger-strict@0.2.0":"NOT_APPLICABLE","ledger-default@0.3.0":"SUPPORTED","ledger-strict@0.3.0":"NOT_APPLICABLE"}} |
high impact insufficient (0)
None.
disputed audits (0)
None.