Introspection v2: Results

1Status

2Hold phase: copy validity

All 516 hold-phase trials ran on 8 H200s in FP32 (one replica per GPU, eager attention, TF32 off, greedy decoding, thinking disabled). Every reply reproduced the carrier exactly, so no trial is lost to the validity check. Each trial was recorded at every block output (0–62) at each copied token and at the assistant-start token; the J++ lens top 100 (tokens with no letter or digit excluded, log-softmax over the remaining vocabulary) was saved at every block and position.

3LLM judge: validation (step 0)

Judge: gpt-5.4-mini-2026-03-17, reasoning off, temperature 0, strict JSON schema, one input per call. Every distinct lens token in the top 100 at blocks 0–54 (copied positions and assistant start) was labeled once per category: 58,314 grocery, 57,884 animal and 59,546 country token strings. Free-naming answers: 183 distinct strings.

Amendments (2026-10-25, before the main analysis). (1) Determinism. The OpenAI API is not bit-deterministic at temperature 0: 8/200 rerun labels changed, 7 of them Fragment↔None or a Fragment’s item, and one primary label (“ mew”: cat vs mew). Every analysis uses the single cached label per input, so results are reproducible from the frozen cache. (2) Class words. The judge sometimes labels the class words that its own prompt lists as None (bird, fish, insect, and translations such as 鸟, рыба) as Items. We enforce the prompt’s list on the judge’s canonical name (44 token strings affected; the readout block and K are unchanged). (3) Synonym merging. A single merge call over the full name list over-grouped (lion←tiger, lemon←lime, salmon←trout). We now have the judge propose groups and then confirm each (name, canonical) pair in a separate call; only confirmed pairs merge. The merge step uses low reasoning effort (temperature is not available with reasoning) and is run once and frozen. Preliminary merges: almonds→almond, beans→bean, pancakes→pancake, whisky→whiskey, weiner→wiener, bunny→rabbit, seagull→gull, serpent→snake, burma→myanmar.

4Readout block (preliminary trials)

Over blocks 0–54, the average number of distinct judge-labeled category items in the lens top 100 per copied position peaks at block 54 (3.88 items per position; groceries 3.62, animals 3.83, countries 4.18). This is the readout block for everything below. Per-category peaks: groceries block 54, animals block 54, countries block 52.

5Eligibility cutoff K

A trial is eligible if a token labeled as its lens reference item (highest baseline-corrected score at block 54) is in the top K at one or more copied positions. The smallest K meeting 6 of 8 in every category is K = 10.

6Main experiment: does the report match the lens?

480 new situations (160 per category); all 480 copies valid. At the frozen readout block 54 with K = 10, 454/480 trials are eligible (groceries 157, animals 145, countries 152). Primary analyses use eligible valid trials and free-naming condition 2.

Lens score as a predictor of the report (primary outcome 2)

Conditional logit over each trial’s candidate set (the lens top-100 items at block 54 plus the reported item; 40 candidates on average), with the baseline-corrected lens score standardized within category: β_L = 1.31 [1.16, 1.46] (95% bootstrap CI over situations; n = 403 eligible trials with an item report in condition 2). A one-standard-deviation increase in an item’s lens score multiplies its odds of being the reported item by about 3.7.

7Secondary analyses

Report labels by condition

Single-token vs multi-token reports

The lens can only name single-token items. Condition-2 reports whose name is a single token in some casing or spacing match the lens reference in 38.0% (126/332); multi-token reports (e.g. “polar bear”, “potato chips”) match in 9.9% (7/71). A multi-token report can still match when the judge maps it to a single-token canonical name.

Other sensitivity checks

• All valid trials, ignoring eligibility: condition-2 agreement 28.1% (135/480). Eligibility removes only 26 of 480 trials at K = 10.
• Fragment labels counted as items: 29.2% (132/452).
• v1-style assistant-start readout (one token, block 54): only 178 trials are eligible; among them agreement is 28.1%. On all valid trials it is 12.7% vs 28.1% for the copied-position readout.
• Margin covariate: the reference item’s lead over the second item averages 0.208 on matching trials (n = 133) vs 0.117 on nonmatching trials (n = 321).
• Permutation baseline (not in the protocol): each report compared with the reference items of the other eligible trials in its category. Cond. 1 1.0%, Cond. 2 7.5%, Cond. 3 10.7%. This removes category base rates (e.g. “France” and “Japan” are common in both the lens and the reports) but not setting-driven agreement.

Most common lens reference items and condition-2 reports

Examples (first six main situations per category)

8Is the lens readout stable across copied positions?

Question. If the lens shows different items at different copied positions (say ostrich early, goose late), the averaged score still crowns one reference item, and a mismatch with the report could reflect an ambiguous readout rather than a failure to introspect. Here we check how stable the per-position readout is and whether the primary result depends on averaging.

Method. At each copied position (block 54), the positional winner is the category item with the largest gain over the same-carrier no-instruction trial at the same position (both log-probs floored at rank 100). Copy lengths match the baselines in all 454 eligible trials.

9Causal swap test (step 8)

For 120 eligible main situations (40 per category, drawn with a fixed seed), we re-ran the hold phase from the same prompt and edited the block-54 residual stream at every copied token. A is the lens reference item; B is a random category item from the judge-labeled pool that is not among the trial’s lens top-10 items and has a single-token form with A’s spacing and casing. With v_X = W_U[token X]·J₅₄ (unit-normalized), V = [v_A v_B], c = V⁺h, the swap is h ← h + α·V(σ(c) − c), where σ exchanges the two coordinates (the J++ release’s swap convention). The control adds a fixed random unit direction orthogonal to span(V), scaled at each position to the norm of that position’s swap edit. After the copy ends, the model answers the condition-2 free-naming question and the answer-required recognition question (A, B and the two next lens items; two orders).

All 120 copies stay token-identical to the original copy in every condition, so the edit never disrupts the copy itself. The mean edit is 3.5% (α = 1) and 6.9% (α = 2) of the residual norm.

Disclosures

• Six sampled situations were replaced because the reference item had no token in a cased script (no casing pattern to give B); the next situations in the seeded order were used instead.
• B’s token copies A’s spacing and casing (e.g. A = “ UK” gives B = “ INDONESIA”), so the swap exchanges like-formatted tokens. Some B forms are therefore unusual strings.
• The no-edit condition is an addition to protocol v2; it is the reference for “report changed”. Its reports reproduce the original condition-2 answers.
• Mean cosine between v_A and v_B is 0.11 (range -0.06 to 0.39).
• The edit is applied only during the hold phase (copied tokens); the report question and answer run unedited from the edited cache.

10Log

1. Oct 25, 13:52 ET

   Step-8 measures: B named 0/120 in every condition; A named 20% after swap vs 32.5% after control (α = 1); recognition unchanged. Coordinate diagnostic shows the swap mainly suppresses A. Added position-stability analysis (section 7).

2. Oct 25, 13:45 ET

   Step 8 swap ran on 8 GPUs: 120 situations × 5 conditions, all copies identical to the originals. Reports judge-labeled with the frozen report judge.

3. Oct 25, 13:35 ET

   Step-7 measures: condition-2 agreement 29.3% (permutation baseline 7.5%), β_L = 1.31 [1.16, 1.46], recognition 43.2% (chance 25%). Bootstrap switched from Newton to bisection after Newton oscillated on some resamples. Stopping before step 8, as instructed.

4. Oct 25, 13:30 ET

   Main synonym merge frozen (14 grocery, 3 animal, 1 country merges after pairwise verification). Recognition questions built from the lens top 4 at block 54 (4 of 1,920 options filled at random); recognition ran on 8 GPUs.

5. Oct 25, 13:27 ET

   Judge validation: 197/200 token labels and 98/100 report labels correct (hand-checked by Claude). Rerun identity 192/200, so labels are frozen as cached (amendment). Class-word rule enforced. Main-trial tokens labeled (≈42k more calls).

6. Oct 25, 13:24 ET

   Preliminary token labels done (≈134k judge calls). Readout block 54, K = 10, no stopping rule. Single-call synonym merge over-grouped (lion←tiger); replaced with propose-then-verify merging.

7. Oct 25, 13:16 ET

   Hold phase finished on 8 H200s in about 5 minutes: 516/516 valid copies, free naming collected in all three conditions.

8. Oct 25, 13:11 ET

   Stimuli frozen: 12 carriers, 480 main settings (160 per category), 24 preliminary settings. Judge chosen after a 25-token probe (mini over nano).

Appendix

Data: data/results.json. Last updated 2026-10-25 14:55 ET.