Corrections and changes
Corrections and changes
Claims here are graded, research is pre-registered and corrections are public.
When something on this site is wrong, it is corrected here with the date and what changed. Nothing is silently edited. To report an error, write to venkat@twopointsingularity.com.
Log
| Date | What changed |
|---|---|
| 12 Nov 2026 | Site: added the story, five movements that connect the two tracks, Neev (the practice) and Signal ≠ Value (the research); one merged lab; a table of how strong each link in the story is. The C320 profile is marked retired. |
| 12 Nov 2026 | Research: the paper Signal ≠ Value moves to version 1.1 after a second adversarial review round, with the lab at v2.2 and the technical supplement at v1.1. Corrected: the source-fix assumption in the H2 simulation was described as “0.15 standard deviations”; the model lowers session depletion by 0.15 on a 0–1 scale, and its sensitivity (11–44% of the lean removed for 0.05–0.20) is now reported. Power at a spread of 11 points is 67%, not 69%. Cost figures made consistent ($3,000–7,000 at 60 people, up to about $11,000 at 100). The “every digital action is a signal” claim narrowed to channels documented as reaching AI training. Added: simulation checks of the sample-size rule and of side bias, an analysis plan, a conditions schema, a pre-mortem and data-protection and dual-use notes. Glossary: Ψ(t) retired from the programme. No hypothesis, threshold or decision rule was loosened. |
| 12 Nov 2026 | Research: the whitepaper v2.0 is replaced by a new paper, Signal ≠ Value: Does Pressure Bend the Human Feedback That Trains AI?, with a technical supplement, following reviewer feedback. It is published as a new paper rather than a version of the March 2026 paper (doi:10.5281/zenodo.19235970), whose claims it supersedes. Changes: the citation for Milli et al. (arXiv:2305.16941) corrected; H1 decisions now read from the confidence interval against a 3-point smallest effect of interest (“supported under the prespecified criterion”, not “proven”); “value” in the method renamed the reflective reference choice, with D as the primary measure and M (mismatch) as a secondary one; a precise null on H1 now described as constraining this mechanism, not clearing human feedback; the glossary terms c and Δc retired. The lab is v2.1 and is readable without JavaScript. Simulation code is now Apache-2.0. |
| 12 Nov 2026 | Site: added the Signal ≠ Value Lab, the Neev app page and this log. |
| 12 Nov 2026 | Research: CWHF whitepaper v2.0 published, with the merged Signal ≠ Value Lab v2.0. Claims retired from v1.0 and the earlier labs: the “Bias Theorem” (now stated as arithmetic conditional on H1), the 120 bits per second figure (current estimate about 10), rate-distortion optimality of the weights, the ICC-based H1, the π ≥ 0.60 pass rule, the five-layer contamination model, and weighting as the main remedy. The full list and reasons are in section 14 of the whitepaper v2.0 (archived) and the lab’s Claims room. |
| 9 Nov 2026 | Site: first public version after launch checks. Single main heading per page, faster loading, fonts hosted on the site. |
Related: Research programme · The lab