Problem 1 — fabricated hero data
experiments-data.js ships AM_SURPRISE = [4.81, 4.55, 4.00, 3.38, 2.70, 2.42, 1.95, 1.61, 1.40, 1.30, 1.18, 1.10, 1.04, 1.00], rendered on index.html ("Surprise · bits/char — the only thing it feels") and journey.html ("falling as it learns").
Only 4 of those 14 values trace to EXPERIMENTS.md — 4.81→4.00 (Exp 1, held-out English corpus) and 3.38→1.61 (Exp 3, "mirro " stream). The other 10 values (4.55, 2.70, 2.42, 1.95, 1.40, 1.30, 1.18, 1.10, 1.04, 1.00) appear nowhere in the log. The curve also splices two different experiments on two different corpora into one continuous "learning" trajectory, and AM_TALLY.from/to = 4.81→1.61 does the same conflation. The index.html caption "4.81 — lost / 1.61 — understanding" presents the spliced endpoints as one measurement.
For a project whose stated product is honest, traceable claims, the front-page chart being invented data is the worst possible failure.
Problem 2 — overclaims contradicted by the project's own RECIPE
- index.html hero: "No labels. No rewards. No pretraining." — but the documented RECIPE explicitly includes "taught labels for the few-shot word↔concept mapping" (Exp 34), and Exp 2 injected a labeled positive (admitted in the log).
- index.html + README.md: "It isn't even handed a dataset" — the Act-I language experiments stream an English text corpus (
data/corpus.txt, Exp 1: "built-in English corpus"). True for the embodied creature, false for the language arc.
Fix
- Replace
AM_SURPRISE with values traceable to EXPERIMENTS.md (e.g. Exp 1's held-out series endpoints, honestly labeled as Exp 1), or clearly label the plot per-experiment. No invented intermediate points.
- Fix the index.html plot caption and
AM_TALLY.from/to so endpoints from different experiments are never presented as one trajectory.
- Soften hero/README claims to match the log: labels are absent for valence/place/opinion formation, but word labels are taught, and the language act consumed a text corpus. The honest version ("nothing pretrained; content self-formed; labels taught") is already written in EXPERIMENTS.md — the site should use it.
Acceptance: every number and series rendered on the site is traceable to a specific EXPERIMENTS.md entry.
🤖 Generated with Claude Code
Problem 1 — fabricated hero data
experiments-data.jsshipsAM_SURPRISE = [4.81, 4.55, 4.00, 3.38, 2.70, 2.42, 1.95, 1.61, 1.40, 1.30, 1.18, 1.10, 1.04, 1.00], rendered on index.html ("Surprise · bits/char — the only thing it feels") and journey.html ("falling as it learns").Only 4 of those 14 values trace to EXPERIMENTS.md — 4.81→4.00 (Exp 1, held-out English corpus) and 3.38→1.61 (Exp 3, "mirro " stream). The other 10 values (4.55, 2.70, 2.42, 1.95, 1.40, 1.30, 1.18, 1.10, 1.04, 1.00) appear nowhere in the log. The curve also splices two different experiments on two different corpora into one continuous "learning" trajectory, and
AM_TALLY.from/to = 4.81→1.61does the same conflation. The index.html caption "4.81 — lost / 1.61 — understanding" presents the spliced endpoints as one measurement.For a project whose stated product is honest, traceable claims, the front-page chart being invented data is the worst possible failure.
Problem 2 — overclaims contradicted by the project's own RECIPE
data/corpus.txt, Exp 1: "built-in English corpus"). True for the embodied creature, false for the language arc.Fix
AM_SURPRISEwith values traceable to EXPERIMENTS.md (e.g. Exp 1's held-out series endpoints, honestly labeled as Exp 1), or clearly label the plot per-experiment. No invented intermediate points.AM_TALLY.from/toso endpoints from different experiments are never presented as one trajectory.Acceptance: every number and series rendered on the site is traceable to a specific EXPERIMENTS.md entry.
🤖 Generated with Claude Code