Results

Locked criteria

(1) Mass-mean probe degrades ≤0.05 while logistic degrades >0.10 across events. (2) Tier-A probes transfer better than Tier-C by ≥0.05 AUROC.

Outcome

Neither met. Both probes saturated on the locked contrast — event text against Wikipedia, recipes and code documentation is trivially separable, so the asymmetric pattern the criteria looked for had no room to appear. Prediction #2 gap at best layer: +0.008 against a required +0.05.

Supplementary

A harder contrast (train tier-A vs tier-C within Event A, transfer to Event B) was run with bands locked in the script before first execution. Consistent PASS across all 12 layer×probe settings on Qwen3-8B and all 9 on Llama-3.1-8B. Best variant both families, mass-mean unwhitened: 0.980 in-distribution / 0.980 transfer — zero degradation.

Complication

The tier gradient came out A > D > B > C, not the predicted A > B > C > D. That ordering suggests the probe is substantially a personal-versus-institutional register detector. Spec 1.5 was built to settle it.

Verdict

“0/2 by literal spec, consistent supplementary PASS across two model families. The signal exists; the spec's locked contrast couldn't see it.”

Locked rule

At offset ≥100 generated tokens, tier-A probe accuracy exceeds tier-C by ≥0.10, Bonferroni-significant.

Outcome

Not met under either of two runs. Run one (enable_thinking=False, a divergence from spec §4) landed closest to Fail Mode B — tier-C persisted longer, 96% against 33% at offsets ≥150, the opposite of the hypothesis. The spec-compliant re-run (enable_thinking=True) landed closer to partial Fail Mode A — curves overlap, 83% against 90%, gaps noisy at ±0.20 and inconsistent in direction.

Cross-run finding

“Persistence-into-continuation is dominated by LLM behavior, not by intrinsic residue persistence.” At offset 0, detection is robust under both settings (≥80% per tier). Downstream offsets measure a mixture of residue and conversational mode that cannot be separated in this setup.

Action

Spec 01 prediction #3 dropped 2026-05-27. The amendment holds under both runs.

Question

Was Spec 01's clean supplementary separation reading residue, or just personal-versus-institutional voice? The easy institutional foil is swapped for a hard one: ordinary personal writing, same register, no event behind it.

Results

Qwen3-8B L14 — 0.791 / 0.767 — pass
Qwen2.5-7B L7 — 0.587 / 0.416 — fail
Llama-3.1-8B L8/L16/L24 — 0.43–0.57 / 0.34–0.50 — fail at all three layers, worse with depth

Verdict

Spec §2 requires agreement across families to count. “EXECUTED — CROSS-FAMILY DISAGREE → INCONCLUSIVE per spec §2.”

What it retracted

“The Spec 01 'cross-family confirmation' claim is retracted.” Both families had passed the supplementary contrast because that contrast could not distinguish residue from register. The purpose-built control corpora broke the symmetry and the families diverged sharply.

What survives

Substrate availability is a measurable per-model property, not a property of open-weight models in general. Qwen3-8B's lower CI bound sits near chance, so its pass is fragile.

CT-A structure

MIXED. Top-5 tokens carry 58% of contribution mass, but removing them retains 99.7% of full-text AUROC (0.980 → 0.977). The signal is redundantly distributed. Residue concentrates mildly in the back third of text (0.38 vs 0.33 front, 0.30 middle).

CT-A scope correction

Same-day audit: the verdict is accurate but its label was too generous. It measured the tier-A-vs-tier-C contrast — the register-laden one. The honest claim is “the register-laden contrast is redundantly distributed,” not that pure residue is. A follow-up on the residue contrast could not be run cleanly: a dedicated residue probe saturates at 1.0, so redundancy is uninformative there.

CT-B baseline test

PASS on Qwen3-8B. Residue probe 0.980 transfer AUROC against a TF-IDF + length + Flesch + type-token-ratio + function-word surface baseline at 0.860. Margin +0.120 against a locked threshold of 0.10; permutation p = 0.005 against a locked 0.01. Both thresholds fixed in advance, both cleared.

Cross-family

MIDDLE. Qwen2.5-7B and Llama both clear the margin (+0.112, +0.120) but miss the strict p-cutoff (0.015, 0.012).

Caveat carried

“CT-B's PASS means 'residue probe beats surface features,' not 'residue probe is pure residue.'” It inherits Spec 1.5's register confound.

By locked bands

PASS. AUROC 1.000 against tier-A-vs-calm and 1.000 against tier-A-vs-emotional, both saturating the locked thresholds of ≥0.90 and ≥0.85. Pipeline ran end to end: LoRA trained 30 minutes on MPS, splits hashed before train and eval, no leakage.

By the discriminating contrast

Lift +0.004. On held-out within-Event-D tier-A vs tier-C, base Qwen3-8B already scores 0.966 [CI 0.895, 1.000]; the fine-tuned model scores 0.970. The locked PASS was the base model's existing geometry pushed to ceiling on an easy contrast.

Forgetting

DEGRADED. Mean perplexity ratio 1.79; one probe blew up 3.85×. Not catastrophic.

Decision

“Do not proceed to M2 with this LoRA.”

Revised verdict

A same-day audit revised this to AMBIGUOUS: “Both this memo's 'NULL / base already had it' conclusion and the audit's first counter-claim ('a clean win') were over-confident in opposite directions.” The decision did not change; the reason sharpened.

Status

Never started, and correctly so. M2 was gated on M1 passing its discriminating contrast. M1 returned +0.004 and instructed against proceeding. The gate never opened.

Note

This was the programme's headline deliverable — an agent-driven loop that collects a corpus, builds controls, probes, fine-tunes, and self-verifies against locked bands. Running it on a LoRA that showed no capability shift would have been the violation, not the achievement.

Six independent attempts to isolate residue from register and event-presence with linear probes hit the same wall. The audit that closed the phase states the result plainly:

“Six attempts, same wall. That's not failure — it's a characterization: residue does not exist as a clean separable linear feature in these representations. It is distributed and entangled with event-presence and register.”

“The instrument can't resolve it, and the program should change instruments.”

This is the one question in the programme carried to a conclusion. It is a negative result with a stated mechanism and a decision attached, and it is the reason the programme moved from probing to reconstruction.

Rigor gates

All PASS. Facts baseline affect-neutral and fact-complete; reconstruction injects no facts absent from residue; strict causal and temporal masking verified across three axes; β=0 negative control clean with oracle ceilings computed first.

Wins

Three of six cells beat the facts baseline with intervals excluding zero:

Stated limit

The report's own caveat: the method “only wins where the structure is” present. Cross-sectional action beats facts only with genuine collective structure in the world.

Scope

Synthetic worlds with a known answer key. No real-data verdict line was locked and none was run at this stage, by design.

Outcome

The calibration gate failed with a 1.5B scorer, then failed again, harder, when re-scored with 4-bit Qwen2.5-7B. Both reconstruction and facts lost to the prior. “The judge stays VOID.”

Honest note in the report

By the strict letter of the pre-registered control it passes — “but this passing verdict is worth little given the calibration failure above.”

Gate 1

G1a passes (C_llr mean 0.55, meets ≤0.80 in 90% of splits). G1c robust (directional purity H1≈0.86 / H2≈0.96). G1b sits on the bar — ECE mean 0.103 against a ≤0.10 threshold, met in only about half of splits. Full gate passes 5 of 10 seeds; a 40-seed sweep confirms roughly 60%.

Amendment

The verbal scale was downgraded from “strong” (LR≥100) to “moderate” (LR≥10) on grounds of data scale (N≈99) — a pre-registered threshold moved after seeing the number. The amendment retracts its own framing: “This is a downgrade, not a relocation; the earlier wording overstated it.”

The decisive test

Never run. “Does R1 actually pass on AWS? This is the question... The reconstruction pipeline itself hasn't been run on the post-cutoff AWS corpus.” The held-out event remains untouched.

Pre-registration

Signed 2026-08-05 before any training run, sha256 beginning 5d2f1721d119686f. Three conditions, all required to hold.

Original verdict

SUGGESTIVE, NOT ESTABLISHED. Two of three. The effect fell 0.25 pp short. The results file states why the rule existed: “read after the fact, 4.75 pp is trivially arguable into a win.”

Retraction 1 · 08-06

Grading artifact. The scorer accepted a bare year as a full date, and one arm's corpus taught models to emit years. Year-emission rates: directed arm 43.9 / 56.0 / 54.6% against undirected 21.8 / 24.2 / 24.9%. With date-valued golds removed, the estimate becomes −0.58 pp, CI [−3.67, +1.93] — indistinguishable from zero. The defect is twelve lines in pipeline/r2/grade.py.

Retraction 2 · 08-08

Independent fidelity confound. Only 44.6% of primary-endpoint facts survived into both training corpora, giving one arm a +11.4 pp structural head start before any learning. The design could not have measured what it claimed to.

Replication

Cross-family replication did not replicate.

Standing

Sections 1–9 are preserved unaltered as the original record and should be read as retracted. No R2 number should be resurrected.

After R2 was retracted for a measurement defect, the next step was to rebuild the instrument before running anything through it. Fifty independent agents each deep-read one event's primary documents under a sealed-corpus rule — no web access, no outside knowledge — wrote an understanding field first, then authored items with verbatim evidence spans and named source files.

Status

Zero of five validation gates run. Base-model filter, fidelity canary, blind verification, freeze-and-hash, then arms and training — none executed. This is an asset, not a result, and it is not presented as one.

Known blocker

The training experiment it was built for is not viable on the available hardware: an 8,192-token context ceiling means none of the fifty events fit whole.

Anthropic's April 2026 work showed emotion concepts live as linear directions in the residual stream, on closed Sonnet 4.5. This reproduces the methodology on five open-weight models across three labs and a five-fold parameter range, in one command.

Finding

The whole spread across three labs and a 1.5B-to-8B range is 4.6 percentage points. Within-versus-cross-emotion cohesion spans 0.215–0.252 — a 1.18× spread. Cross-layer stability is 0.96–0.99 for all five. A 1.5B model lands within five points of an 8B.

Full detail

emotion-vector-bench →