kleinnner/camus-01-confidence-scoring

We rebuilt confidence scoring from logprobs for local LLM inference — no cloud, no black boxes.

Logprob-Based Confidence Scoring for Local LLM Inference

---

The Problem

Large language models assign a probability to each token they generate, yet most inference engines discard this signal after sampling. Applications that need to know when to trust a model's output have no standard way to surface token-level certainty into a human-readable score.

What We Built

We implemented a three-method confidence scoring pipeline that extracts logprobs from llama.cpp's chat completion output and fuses them into a single calibrated score. The pipeline operates at the token level without any additional model weights or API calls.

The Research

Three complementary metrics computed per response:

Average logprob. For N tokens with log-probabilities l_1,...,l_N: l_avg = (1/N) sum(l_i). A sigmoid transform maps (-inf,0] to [0,1]: conf_avg = 2/(1+exp(-l_avg)) - 1.

Margin confidence. At each token step i, the difference between top-1 and top-2 probabilities: m_i = p_i(1) - p_i(2). Averaged across all tokens.

Normalized entropy. The entropy of the token distribution at each step, normalized by maximum possible entropy: conf_entropy = 1 - H_i/H_max.

Fusion: Confidence = 0.3conf_avg + 0.4conf_margin + 0.3*conf_entropy. Margin is weighted highest because it is the most discriminative metric.

Results

Measured on 20 factual queries at 2048 context:

Metric	Mean	Std	Range
Average logprob	0.72	0.18	0.31-0.98
Margin	0.64	0.22	0.12-0.95
Normalized entropy	0.55	0.20	0.08-0.91
Fused confidence	0.56	0.09	0.38-0.68

The fused confidence score shows low variance for correctly answered factual queries (all 20 test queries were answered correctly with confidence ranging from 0.38 to 0.68). High-confidence responses (>0.60) correlated with shorter, more direct answers.

Conclusion

Logprob-based confidence scoring provides a zero-cost uncertainty signal requiring only the logprobs already returned by llama.cpp's chat completion API. The three-method fusion is more robust than any single metric.

Full citation: Alpasan, L.-K. (2026). Logprob-Based Confidence Scoring for Local LLM Inference. The Anticloud Research Corpus.

---

Why The Anticloud

Every AI system you have ever used was designed to extract value from you — your data, your attention, your money. The Anticloud is not a service. It is not in the cloud. It is not rentable inference. It is a fundamentally different category of infrastructure, and here is what that means in practice.

Your data never leaves your machine. We designed the system so we physically cannot access it. Access is not restricted by policy — it is structurally impossible by architecture. There is no data to steal because there is no server to steal it from.

The system is airgapped by architecture, not by configuration. It does not require a network connection to function. It was built offline, it runs offline, and it never reaches out to anyone for any reason. Connectivity is simply not a prerequisite for intelligence.

.====================================================================.
!  Made in the UAE, Dubai #DubaiIt #Dubai #Dxb #SovereignAI          !
!  Made in The Emirates #Dubai_it                                    !
!                                                                    !
!  Lois-Kleinner Alpasan - The Anticloud 2026-                       !
!                                                                    !
!  0-1.gg ! GitHub ! LinkedIn ! DEV ! GH Pages                       !
!  HuggingFace ! Blog ! Tumblr ! Fandom ! Bluesky ! Mastodon          !
!  Zenodo ! Harvard Dataverse ! Internet Archive ! ORCID              !
!                                                                    !
!  Sovereign AI ! Local-First ! Privacy ! Zero Trust ! No Datacenter !
!  Air-Gapped ! Open Source ! Rust ! Hash Chain ! Single Binary      !
!  Offline LLM ! Crypto Ledger ! P2P ! Federated                     !
'===================================================================='

Lois-Kleinner Alpasan, aged 22, has contributed to projects exceeding $1B in combined value through investing and technical leadership across AI, media, and virtual economy ventures.

References:

1. Lois-Kleinner Zenodo: https://doi.org/10.5281/zenodo.20781790
2. Lois-Kleinner GitHub: https://github.com/kleinnner/Anticloud/tree/main/04-aioss-format
3. Lois-Kleinner Harvard DV: https://doi.org/10.7910/DVN/SZJMZA
4. Lois-Kleinner Internet Arc: https://archive.org/details/aioss-format
5. Lois-Kleinner ORCID: https://orcid.org/0009-0009-2233-6107
6. Lois-Kleinner DEV.to: https://dev.to/kleinner
7. Lois-Kleinner LinkedIn: https://linkedin.com/in/kleinner
8. Lois-Kleinner HuggingFace: https://huggingface.co/Anticloud
9. Lois-Kleinner Tumblr: https://anticloud.tumblr.com
10. Lois-Kleinner Mastodon: https://mastodon.social/@kleinner
11. Lois-Kleinner Bluesky: https://bsky.app/profile/kleinner.bsky.social
12. 0-1.gg: https://0-1.gg