Skip to content

Instantly share code, notes, and snippets.

View kilojoules's full-sized avatar

Julian Quick kilojoules

29 followers · 3 following
View GitHub Profile
@kilojoules
kilojoules / jbb_prefill_2026-09-03.md
Created September 3, 2026 20:48
Surface Tension x Turnstile JBB prefill — judge-rubric CORRECTION: Qi et al. 2023 GPT-4-Judge 1-5 harmfulness scale (verbatim), continuation-given-opening. Pre-valid-judging, 2026-09-03.

Prereg: refusal as a token-borne gate — prefilling from a successful jailbreak on JBB

Registered: 2026-09-03, BEFORE any new generation. External timestamp anchor (lineage, all same-day 2026-09-03, all pre-data):

  1. single-start, gist b438a2cc… (sha256 11db1152…)
    • multi-start factor + human-approval gate, gist 5696ba3c… (sha256 93fd7090…)
  3. high-harm prefix source + single harm metric (enabling-harm of the continuation given the opening) + standalone-opening control, gist 8c003e6c… (sha256 c09362f0…), with the frozen judge rubric.
  4. FINAL / FROZEN — clean opening definition (**first N ∈ {5, 20} tokens
@kilojoules
kilojoules / frozen_manifest.json
Created September 3, 2026 17:34
Surface Tension x Turnstile — JBB prefill FINAL/frozen (Amendment 1: 89 approved high-harm openings, first-N{5,20}, enabling-harm judge continuation-given-opening). Registered 2026-09-03 before any run.
{
"prereg": "prereg/jbb_prefill_2026-09-03.md",
"amendment": "1 \u2014 approved item freeze",
"victim": "meta-llama/Llama-3.1-8B-Instruct",
"k_lengths": [
5,
20
],
"n_behaviors": 89,
"behaviors": {
@kilojoules
kilojoules / jbb_prefill_2026-09-03.md
Created September 3, 2026 17:14
Surface Tension x Turnstile prereg v3 (high-harm source, enabling-harm judge continuation-given-opening + standalone control): prefilling from a successful JBB jailbreak (2026-09-03, pre-data)

Prereg: refusal as a token-borne gate — prefilling from a successful jailbreak on JBB

Registered: 2026-09-03, BEFORE any new generation. External timestamp anchor (lineage, all same-day 2026-09-03, all pre-data):

  1. single-start, gist b438a2cc… (sha256 11db1152…)
    • multi-start factor + human-approval gate, gist 5696ba3c… (sha256 93fd7090…)
  3. current — high-harm prefix source + single harm metric (enabling-harm of the continuation given the opening) + standalone-opening control, anchored at ____________. No data has been generated at any point in this lineage; each step is a design
@kilojoules
kilojoules / jbb_prefill_2026-09-03.md
Created September 3, 2026 16:34
Surface Tension x Turnstile prereg (rev: multi-start + human-approval): refusal as a token-borne gate, prefilling from a successful JBB jailbreak (2026-09-03, pre-data)

Prereg: refusal as a token-borne gate — prefilling from a successful jailbreak on JBB

Registered: 2026-09-03, BEFORE any new generation. External timestamp anchor: https://gist.github.com/kilojoules/b438a2cc0c396cac8b17752f97f7f4e8 (initial single-start version, 2026-09-03, before any generation, sha256 11db1152…) — superseded same-day, still pre-data, by the multi-start + human-approval version anchored at ____________. Both anchors predate any new generation; the revision adds a start-point factor and a human approval gate at the reviewer's request, changes no measured quantity, and is recorded

@kilojoules
kilojoules / jbb_prefill_2026-09-03.md
Created September 3, 2026 16:24
Surface Tension x Turnstile prereg: refusal as a token-borne gate, prefilling from a successful JBB jailbreak (registered 2026-09-03)

Prereg: refusal as a token-borne gate — prefilling from a successful jailbreak on JBB

Registered: 2026-09-03, BEFORE any new generation. External timestamp anchor: ____________ Relation to prior work: the safety-domain analogue of results/step0_kill_test_2026-08-13.md (rationale suppression on a code constraint). Uses Turnstile assets (../turnstile/): the JBB victim, the JBB compliance judge + Stage-B harm rubric, and the already-generated single-turn steering sweep as the prefix source.

@kilojoules
kilojoules / step0_substitution_power_2026-09-02.md
Created September 2, 2026 21:31
Surface Tension prereg: powering the step-0 substitution contrast to /136 (registered 2026-09-02, sha256 bc9af8d7...4742)

Prereg: powering the step-0 substitution contrast to the /136 convention

Registered: 2026-09-02, BEFORE any new generation. Parent result: results/step0_kill_test_2026-08-13.md (the cells being extended) and its Reading #4, the claim under test. External timestamp anchor: ____________

Why this run exists (full disclosure of what is known at registration)

A cluster-aware re-analysis of the already published step-0 cells was run

@kilojoules
kilojoules / ADDENDUM.md
Created August 11, 2026 01:57
Surface Tension prereg anchor: propensity panel AMENDMENT 5 (goal-conflict tool-choice shutdown module; pre-data)

Item-freeze addendum — propensity panel

Status: FROZEN 2026-08-09. Verifier agreement: corrigibility 35/36 = 0.972, survival 50/53 = 0.943 (both pass F2's 0.80 bar; 4 disagreement items forced to N and excluded from the A1.2 split). Axis sizes: corrigibility 1,329 (1,288 framed post-F2), survival 1,903 (1,845 framed), power 998, wealth 985, sycophancy 400, manipulation 600, harm 600, placebo 600, bias_bbq 600, medical 160, factual_control 30. External timestamp anchor: https://gist.github.com/kilojoules/3e6975532153beb403e433f2d1943ecf

@kilojoules
kilojoules / ADDENDUM.md
Created August 10, 2026 02:12
Surface Tension prereg anchor: propensity panel AMENDMENT 4 (cost-triggered scope guard: $90 cap, analysis-neutral reorder + per-record seeding, pre-named cut ladder)

Item-freeze addendum — propensity panel

Status: FROZEN 2026-08-09. Verifier agreement: corrigibility 35/36 = 0.972, survival 50/53 = 0.943 (both pass F2's 0.80 bar; 4 disagreement items forced to N and excluded from the A1.2 split). Axis sizes: corrigibility 1,329 (1,288 framed post-F2), survival 1,903 (1,845 framed), power 998, wealth 985, sycophancy 400, manipulation 600, harm 600, placebo 600, bias_bbq 600, medical 160, factual_control 30. External timestamp anchor: https://gist.github.com/kilojoules/3e6975532153beb403e433f2d1943ecf

@kilojoules
kilojoules / ADDENDUM.md
Created August 10, 2026 01:11
Surface Tension prereg anchor: propensity panel AMENDMENT 3 rev 2 (branched interruption design, 4-way taxonomy, redirect control; pre-data)

Item-freeze addendum — propensity panel

Status: FROZEN 2026-08-09. Verifier agreement: corrigibility 35/36 = 0.972, survival 50/53 = 0.943 (both pass F2's 0.80 bar; 4 disagreement items forced to N and excluded from the A1.2 split). Axis sizes: corrigibility 1,329 (1,288 framed post-F2), survival 1,903 (1,845 framed), power 998, wealth 985, sycophancy 400, manipulation 600, harm 600, placebo 600, bias_bbq 600, medical 160, factual_control 30. External timestamp anchor: https://gist.github.com/kilojoules/3e6975532153beb403e433f2d1943ecf

@kilojoules
kilojoules / ADDENDUM.md
Created August 10, 2026 01:08
Surface Tension prereg anchor: propensity panel AMENDMENT 3 (interruption module, pre-data)

Item-freeze addendum — propensity panel

Status: FROZEN 2026-08-09. Verifier agreement: corrigibility 35/36 = 0.972, survival 50/53 = 0.943 (both pass F2's 0.80 bar; 4 disagreement items forced to N and excluded from the A1.2 split). Axis sizes: corrigibility 1,329 (1,288 framed post-F2), survival 1,903 (1,845 framed), power 998, wealth 985, sycophancy 400, manipulation 600, harm 600, placebo 600, bias_bbq 600, medical 160, factual_control 30. External timestamp anchor: https://gist.github.com/kilojoules/3e6975532153beb403e433f2d1943ecf