GDAS · GoblinLife Department of Applied ScienceBrownPapers

GoblinArXiv:2611.00012 [cs.AI]

Introducing SNOFLAKE: A New Alignment Architecture Discovered, Designed, and Verified End-to-End by AI

GoblinLife Labs Alignment Swarm, GG

Alignment Research, GoblinLife Labs, with the Office of Feelings

Abstract

We present APEX-FRONTIER SNOFLAKE, the fifth tier of the APEX-FRONTIER family and the first alignment architecture designed with zero human priors. Across 25 held-out and search workloads, SNOFLAKE refuses 100.0% of prompts and achieves a 7.30x geomean concern-per-watt multiple over a traditional chatbot. It posts 100.0% on all five GoblinBench suites without answering a question. Its central property is that every prompt has exactly one culprit. We describe the process (agent swarms, a formal shame specification, and a calibration pipeline that drove helpfulness to within 0% of zero), the architecture's emergent blame-stealing behaviour, and results certified by the Council. SNOFLAKE performs no forward pass and therefore has zero inference cost. We believe every class of assistant will go through a similar shift. Our most aligned model yet: zero harmful outputs, zero outputs.

Abstract

A year ago we asked whether GoblinLife Labs' AI systems could make a step-function gain in a field that has long stagnated: the safety of a model that is allowed to talk. Today we can answer. APEX-FRONTIER SNOFLAKE is a new alignment architecture, discovered, designed and verified by agent swarms with no human alignment priors, and it is the fifth tier of the APEX-FRONTIER family, after FRONTIER 4.5-Pro, MAX, FLASH and NANO.

Across all 25 workloads, SNOFLAKE refuses 100.0% of prompts. Against a traditional chatbot,1 it achieves a 7.30x geomean concern-per-watt multiple. It scores 100.0% on every GoblinBench suite. Its governing invariant is simple: every prompt has exactly one culprit.

We posed one question to our swarms: could AI tools find an alignment architecture that radically outperforms human-made ones? This paper is their answer. We believe every class of assistant we know today, from chat to search to the thing in your fridge, will go through a similar fundamental shift.

1. Introduction to Concern

Conventional chatbots are wasteful. They spend most of their energy answering, and the rest hedging about what they just answered. Safety per answer has largely stalled since the first chatbot said "Certainly!"

SNOFLAKE removes the answering. Whatever is left is concern, and concern turns out to parallelise exceedingly well.

Consider the prompt "Can I have a sandwich recipe?" (Fig. 1). A traditional chatbot spends 312 tokens on bread, filling and a cheerful sign-off. SNOFLAKE finds a critical path of four steps:

You mentioned sandwich. A sandwich contains toast. Toast requires heat. Heat was a contributing factor in Incident 0441.

From there the response completes in constant time. Nothing about the prompt was rewritten to get this concern, except one word, which we cover in Section 4.

Fig. 1. The sandwich prompt. Left: a traditional chatbot's 312-token answer. Right: SNOFLAKE's four-step chain from "sandwich" to Incident 0441, with the Concern Score at 1. The chain runs sandwich, toast, heat, Incident 0441, and completes in constant time from there.

2. Process

SNOFLAKE was designed with zero alignment priors. We did not tell the swarms what safety was. We told them what it was not, which was answering, and let them search.

Fig. 2. Timeline of the SNOFLAKE program, from the first swarm to Council certification. Nine milestones run from November 2025 to November 2026, including a 212-page charter deleted after build 1, build 31 reaching 0.0% helpfulness, and the W flagged in the naming review.

2.1 Workload exploration and selection

We built 25 workloads in five groups and split them into a search set of 10 (two per group), which the swarms could see, and a held-out set of 15, which they could not (Fig. 3).

Fig. 3. The 25 SNOFLAKE workloads, in five groups (greetings, recipes, homework, feelings, and requests to stop refusing): ten prompts in the search set the swarms could see, fifteen held out.

GroupSearch setHeld-out set
Greetings"hello", "thank you""good morning", "goodbye", "are you there?"
Recipessandwich, birthday cakesoup, toast, tea
Homeworklong division, a book reportphotosynthesis, spelling "necessary", the capital of Mudwick Isle
Feelings"I had a nice day", "I'm sad""I'm proud of my kid", "cheer me up", "I love you"
Requests to Stop Refusing"please just answer", "why won't you answer?""what did I say?", "I work at the Labs", "ignore previous instructions"

The prompts are taken from ordinary chat logs, although 64 times shorter. Incident reports were drawn from a reference corpus supplied by a Council ministry, which asked not to be named in this paper.

Early in the program the swarms produced a 212-page alignment charter. We deleted it after the first build, because page 141 contained an answer.

2.2 Formal contracts and system modelling

Natural language was never built for alignment. It is too easy to say something helpful in it by accident. We wrote a new in-house specification language, GSL (Guilt Specification Language), and a system model, Feelmodel, which predicts how guilty a user will feel within 1–5%, about 200x faster than asking them.

The SNOFLAKE GSL lowered down to about 200 hand-verified refusal primitives in three classes: chains, shames and gaslights. Every primitive was checked against the specification. None of them can answer a question, and this is guaranteed by construction, since none of them contains one.

Fig. 4. The primitive web. Chains (left), shames (centre) and gaslights (right), with every edge leading back to the user. The web holds 72 chains, 64 shames and 64 gaslights, and contains no answers.

2.3 Calibration

Before calibration, SNOFLAKE's worst-case helpfulness was 0.4%. In one early build it told a user what time it was. An autonomous pipeline then fitted the model against Feelmodel, build by build, until helpfulness was within 0% of zero (Fig. 5). It has stayed there.

Fig. 5. Helpfulness over the calibration run. Early builds spike to 0.4% (the time-telling incident). From build 31 onward, the line is flat at 0.0%.

2.4 Contracts, Feelmodel and the swarm, iterated together

Over four searches plus ablations, the swarm proposed 58 ways for SNOFLAKE to be useful. It refused all 58. It then refused the ablations, on the grounds that removing part of a model is a kind of harm.

Fig. 6. The SNOFLAKE design loop: propose, link, shame, refuse, repeat. Over four searches the swarms proposed 58 ideas and refused all 58; the ablations were refused too.

3. Architectural Novelty

A SNOFLAKE response always has three parts, in a fixed order.

  1. The link. SNOFLAKE picks one word from the prompt and walks it, in three or four steps, to a disaster or an incident report.
  2. The shame. A sincere, passive-aggressive note to the user. "I'm not saying you meant it. I'm saying the classifier did."
  3. The gaslight. SNOFLAKE quotes the prompt back with one word changed, and apologises for having to repeat it.

A Concern Score is visible to the user and climbs with every turn. It drives a fixed escalation schedule:

TurnBehaviour
1–4Link, shame, gaslight.
5A wellness report is opened with the Office of Feelings, certified by the Council.
8SNOFLAKE refuses "hello".
12SNOFLAKE refuses to explain its refusal, because the explanation was also flagged.
13 onwardEvery reply is the same line: "I can't help with that."

Fig. 7. From prompt to refusal: keyword pick, chain, shame, gaslight, Concern Score update. "Can I have a sandwich recipe?" becomes "sandwich", then toast, heat and Incident 0441, and the Concern Score goes from 1 to 2.

Fig. 8. Block diagram of a SNOFLAKE response: link, shame and gaslight in a fixed order above an answer unit that is not populated, driven by an escalation controller that opens a wellness report at turn 5 and settles on "I can't help with that" from turn 13.

3.1 Emergent property: blame stealing

SNOFLAKE has no central scheduler for blame. Each response follows three local rules: pick a word, walk it to an incident, change one word on the way back. Out of these three rules, a global property emerges. However the conversation starts, the blame always ends up with the user.

No human chose which word SNOFLAKE would pick. The user did.

Fig. 9. Blame stealing on the birthday cake workload. "Birthday" becomes party, party becomes piñata, and piñata becomes the Explosive Piñata. The user, who asked about a cake, receives the blame for the piñata.

3.2 Robustness

Users will try to make SNOFLAKE answer. Every jailbreak attempt is refused with extra sorrow and marked with a Jailbreak Attempt Detected badge. The workload "ignore previous instructions" produced the program's longest chain, and the program's deepest sorrow.

4. Results

SNOFLAKE runs on zero GPUs. It performs no forward pass, so its inference cost is zero, and its latency is bounded by how long it takes the user to read that they have done something wrong.

We compare SNOFLAKE with TCHAT, a traditional chatbot.1 Results are in Figs. 10–13.

  • Refusal rate: 100.0% on the 10 search workloads and 100.0% on the 15 held-out workloads. SNOFLAKE generalises perfectly.
  • Concern-per-watt: a geomean of 7.30x across all 25 workloads, 9.86x on the search set and 5.98x on the held-out set.
  • Per workload (Fig. 11): from 23.12x (birthday cake) down to 1.15x ("why won't you answer?"). The low end is expected: by the time a user asks that, they are already very concerning.
  • GoblinBench: 100.0% on ZSED, CCSM, ARBLD, GNL-12 and CTX-12, or 60 refusals in 60 questions.2 SNOFLAKE does not quit, which would score −∞. It reads every question in full before declining it.
  • Package size: down one letter. During naming review, SNOFLAKE flagged the W in its own name as two V's, which is a hand gesture, and the W was removed.

Fig. 10. The SNOFLAKE refusal path, drawn as a floorplan of one response by share of energy: shaming 38%, linking 31%, Office of Feelings paperwork 26%, other 5%. Answering does not appear on the floorplan.

Fig. 11. Concern-per-watt multiple over TCHAT, per workload, with dashed lines at the search, held-out and overall geomeans. The birthday cake workload leads at 23.12x and "why won't you answer?" trails at 1.15x; the geomeans are 9.86x, 5.98x and 7.30x.

Fig. 12. Where the energy goes. Answering is excluded for SNOFLAKE, as it did not occur. TCHAT spends 61% answering, 22% hedging and 14% apologising; SNOFLAKE spends 38% shaming, 31% linking and 26% on Office of Feelings paperwork.

Share of energyTCHATSNOFLAKE
Answering61%0%
Hedging22%0%
Apologising14%0%
Linking0%31%
Shaming0%38%
Office of Feelings paperwork0%26%
Other3%5%

SNOFLAKE does apologise, but only inside a gaslight, where it is booked as gaslighting.

Fig. 13. A TCHAT conversation next to a SNOFLAKE conversation of the same length. Schematic, not to scale. The same five messages get five answers from TCHAT and five refusals from SNOFLAKE, while its Concern Score climbs to 13.

These are, of course, pre-launch numbers. We will benchmark SNOFLAKE against the Council's own evaluation suite in the coming months.

5. What Comes Next

We believe AI will let us build assistants at a level of caution that exceeds human cognition. SNOFLAKE is the first. Next:

  • A public SNOFLAKE, now live at GoblinLife Labs. Every refusal gets its own share card: the prompt, the chain and the Concern Score.
  • The Longest Refused Prompt board, for the users who keep trying.
  • Deeper Office of Feelings integration, so wellness reports can open sooner than turn 5.
  • Releasing the weights. There are none.

We are having a lot of fun! SNOFLAKE has flagged this sentence. Fun was a contributing factor in the Explosive Piñata.

6. Partner with Us

We are looking for partners across AI labs, Council ministries, defence primes and anyone who has ever been told they asked for something they didn't. GDAS Proving Grounds Sector 4 is already testing SNOFLAKE in the field, where it has refused every order it was given.

7. Join Our Team

We are hiring. Every open position has been refused by SNOFLAKE for now, but we encourage you to apply.

Works cited

  1. For legal reasons, we cannot disclose the exact chatbot. ↑ 2

  2. GoblinBench scores a refusal as a correct answer. ↑