Re-verify density-chain note against pinned source (2026-07-20): verified clean

[OpenCnid]

Re-verification: verified clean

Verified clean against the pinned source
(https://transformer-circuits.pub/2026/workspace/index.html, "Published July 6,
2026") on 2026-07-20; refreshed verification date only.

The pinned page still resolves at its canonical URL, still reads "Published
July 6, 2026," returned no redirect, and carries no revision/errata banner. Every
T1–T5 claim, every Key-results row, and every locator was re-checked against the
article studied this session. No discrepancies found.

Ledger — {claim} · {note said} · {source says} · {locator}

Claim	Note said	Source says	Locator
J-lens corpus	1,000 pretraining-like prompts	"a corpus of one thousand prompts sampled from a pretraining-like distribution"	§Methods: The Jacobian Lens
Workspace layer band	≈ layers 38–92, 25 sampled layers reindexed 0–100	workspace onset layer 38; L38–92 throughout; "25 evenly spaced layers … reindexed to the range [0–100]"	§In which layers…; ledger line 123
Occupancy	plateau ≈ 25 (median)	"rises to a plateau of around 25 (in the median case)"	§Capacity of the J-space
Excess variance	never exceeds 10%	"never exceeding 10%"	§Capacity of the J-space
Concept vector J-space share	median 6–7%	"median of only 6–7% of the concept vector's variance"	§…supports verbal report
Report swap to top-5 (pure/J-comp/non-J)	88% / 59% / 5%	88% pure J-lens, 59% J-space component, 5% non-J	§…supports verbal report
Two-hop swap (50 prompts)	Haiku 54%, Sonnet 70%, Opus 4.5 70%	"54% … on Haiku 4.5, 70% on Sonnet 4.5, and 70% on Opus 4.5" (50 two-hop prompts)	§…mediates internal reasoning
Intermediate vs answer depth	median ≈17% earlier	"a median of approximately 17 percent earlier"	§…mediates internal reasoning
Probe swap flips (n=90)	61% / 28% / 6%	61% J-comp, 28% non-J, 6% clamped, over 90 prompts	§…mediates internal reasoning
Generalization (16 fns × 12 pairs)	76/192 top-1; 101/192 at α=2	"76 of 192 trials … 101 of 192 succeed"	§…flexible generalization
MLP gain	≈10× (neuron dirs ≈1×)	"rises … to roughly 10×"; neuron output dirs "gain near 1"	§Broadcast Across Depth
Broadcast-head ablation (top 1%)	recall@25 0.67 vs 0.86; top-1 5% vs 2%; injected 0.54→0.09	"recall falls to 0.67 … 0.86 under control"; "5% of positions, against 2%"; "0.54 to 0.09"	§Broadcast Across Tokens
Eval-awareness ablation	71%→3%; blackmail 0/180→13/180 (7%)	"from 71% to 3%"; "0 of 180 … 13 of 180 (7%)"	§Strategic thinking and evaluation awareness
Self-monitoring register (40 concepts)	damn 82%, fail-words 93% post-trained vs 30% / 17% base	fail-words 93% / damn 82% post-trained; base 17% / 30%	§Evidence of self-monitoring by the Assistant
Stream-of-consciousness	thinking 58%, thoughts 23%, feeling 17%, conscious 7%	58% / 23% / 17% / 7% of (position, layer) slots	§J-space ablation flattens experiential reports
Counterfactual Reflection Training	0.25→0.07; 0.38→0.05; ablate (176/63 tokens) reverts to 0.22/0.23	0.25→0.07 fabrication; 0.38→0.05 deception; 176/63 tokens; reverts 0.22 / 0.23	§Shaping the J-space with CFR
Models	Sonnet 4.5 primary; Haiku 4.5, Opus 4.5, Opus 4.6	"By default … Claude Sonnet 4.5 … corroborate on Haiku 4.5 and Opus 4.5 … Opus 4.6"	§Technical details
Ablation battery	shallow (MMLU/SQuAD/sentiment) intact; multi-hop/translation/analogy below unablated Haiku 4.5; GSM8K-CoT robust	14-task battery matches, incl. "well below … unablated Haiku 4.5" and GSM8K-CoT "more robust"	§J-space ablation…
Model organisms	leak fake/secretly & reward/bias; logit-lens miss	reward-hacker leaks fake/secretly/fraud/trick/hidden; reward/bias organism — logit lens "surfaces neither reward nor bias"	§Detecting… (both)
Access vs phenomenal	access only, not phenomenal	"we take no position … focus on the functional role"	§Motivation
Authors / affiliation / date	16 authors, * on Gurnee/Sofroniew/Lindsey, † Lindsey; Anthropic; July 6 2026	matches byline exactly	article header
License	none stated on page	no license string present on page	—

Changed

• verified_against_source: 2026-07-18 → 2026-07-20 (frontmatter + index.json; index.json.updated too)
• Provenance: appended a 2026-07-20 re-verification bullet
• Note body facts, tiers, Key-results, and "Our take": unchanged (nothing warranted a fix)

🤖 Generated with Claude Code