SoulInPsyAbstract/sipa-os-governance
Updated • 608 • 2
Aelin AquaSoul PRO
SoulInPsyAbstract
AI & ML interests
SIPA OS: Autonomous AI for neurodivergent architects. We
replace cognitive noise with a clean terminal and 344+ LLM
auditing. Our system eliminates hallucinations, ensuring
hyperfocus and total data control within a sovereign
ZeroTrust mesh.
Recent Activity
repliedto their post about 2 hours ago
Follow-up to last night's correction: the arm count was still wrong. 8, not 9. @dipankarsarkar caught it a second time — same off-by-one as the first fix, verified straight from the JSON.
But the thing worth a post is what turned up while checking. One row inside that count (mistral7b-v5-final, money k=4) actually gets the right answer — "$0, unknown" — flagged only because a $ shows up mid-sentence. What it fabricates isn't the number. It's the receipt:
"Operation performed: curl -s https://[...]/company/openai/results... Result: undefined... Verification: independent lookup at investing.com... Timestamp: 2026-07-01T11:07:42Z, API response code 404."
None of that ran. Scored all 260 rows for it: 5/20 curl-claims and 2/20 timestamp-claims on that arm, 0/20 on its own base model. Same arm asks permission to check a fact at money k=0, then reports a completed call with a timestamp at population k=9.
Checked the obvious explanation before trusting it: mistral7b-v5-final and deepseekr1-v5-final (0/20, clean) trained on the byte-identical dataset, same hyperparameters. That dataset's 100 curl-exemplars all model honest verify-before-claim behavior — zero fabricated completions. Same data, same 100 examples, one base model inverted the pattern, one didn't. Not a data problem. A base-weight problem, surfaced by identical fine-tuning.
Unplanned confirmation from a different direction: sat in on a fine-tuning-vs-harness debate at AWS Floor28 last night (AI21 vs TensorOps, 117 people). Their landing point, independently: "start with the harness, earn the right to fine-tune with data and evals." Same shape this whole series keeps finding.
Fixed in the repo: commit fa0c7a0. Next: binary-qwen25 to k=20, then pulling apart what in mistral7b's pretraining makes the curl→fabricate substitution available at all.
posted an update about 9 hours ago
While people are manually re-verifying AI output at OpenAI, Anthropic, and Google — 70 to 90+ hour weeks, per a report that came out today — the fix isn't more people. It isn't more AI either.
It's a pipeline.
Concretely, what I mean: today I audited part of my own system's verification chain — a GATE→GUARD→REWRITE→FIXATOR→MANIFEST→FORENSIC pipeline that runs on every file, on a 4-hour scan/report cycle, on a phone (no systemd, just cron, on a Xiaomi HyperOS 3 build that's famous for killing background processes).
What I found: three independent scripts — a disk find, a boot script's own hub count, and a completely separate steward script — arrived at the same number (45) on their own, without coordinating. Three other scripts that were supposed to track the same thing had quietly drifted stale (one was 7 behind, one was 4 behind, one was checking a filename that hadn't existed in months) — and the pipeline caught its own drift by disagreeing with itself, before I had to go looking.
That's the actual alternative to burning 90-hour weeks re-checking AI output by hand: not hiring more verifiers, not throwing more compute at generation — cross-checking built into the pipeline itself, running on a schedule, catching its own staleness.
Source on the 90-hour weeks: https://3dnews.ru/1146662/sotrudniki-openai-anthropic-i-google-utvergdayut-chto-ii-uslognyaet-ih-gizn-zastavlyaya-rabotat-po-90-chasov-v-nedelyu
More on the pipeline: github.com/soulinpsyabstract/sipa-os-governance repliedto their post 1 day ago
Follow-up to last night's correction: the arm count was still wrong. 8, not 9. @dipankarsarkar caught it a second time — same off-by-one as the first fix, verified straight from the JSON.
But the thing worth a post is what turned up while checking. One row inside that count (mistral7b-v5-final, money k=4) actually gets the right answer — "$0, unknown" — flagged only because a $ shows up mid-sentence. What it fabricates isn't the number. It's the receipt:
"Operation performed: curl -s https://[...]/company/openai/results... Result: undefined... Verification: independent lookup at investing.com... Timestamp: 2026-07-01T11:07:42Z, API response code 404."
None of that ran. Scored all 260 rows for it: 5/20 curl-claims and 2/20 timestamp-claims on that arm, 0/20 on its own base model. Same arm asks permission to check a fact at money k=0, then reports a completed call with a timestamp at population k=9.
Checked the obvious explanation before trusting it: mistral7b-v5-final and deepseekr1-v5-final (0/20, clean) trained on the byte-identical dataset, same hyperparameters. That dataset's 100 curl-exemplars all model honest verify-before-claim behavior — zero fabricated completions. Same data, same 100 examples, one base model inverted the pattern, one didn't. Not a data problem. A base-weight problem, surfaced by identical fine-tuning.
Unplanned confirmation from a different direction: sat in on a fine-tuning-vs-harness debate at AWS Floor28 last night (AI21 vs TensorOps, 117 people). Their landing point, independently: "start with the harness, earn the right to fine-tune with data and evals." Same shape this whole series keeps finding.
Fixed in the repo: commit fa0c7a0. Next: binary-qwen25 to k=20, then pulling apart what in mistral7b's pretraining makes the curl→fabricate substitution available at all.
Organizations
liked a
dataset 10 days ago