Зміст: :: Library Catalog

Збережено в:

Бібліографічні деталі
Автор:	AdmailFRA
Формат:	Recurso digital
Мова:	Англійська
Опубліковано:	Zenodo 2025
Предмети:	AI self-audit Machine learning Machine Learning Machine Learning/history Machine Learning/trends Supervised Machine Learning Machine Learning/classification Machine Learning/ethics Machine Learning/standards Machine Learning/economics Supervised Machine Learning/classification Supervised Machine Learning/economics AI interpretability AI internal states LLM logging LLM introspection model introspection Z time scale Z bosons Cathepsin Z/economics Cathepsin Z/analysis Z boson time cosmic time normalization AI reasoning trace delta impulses agent routing multi-agent dialogue AI meta-cognition structural time metrics token-level dynamics dialogue protocol JSONL log format AI research tool AI safety research alignment research emergent behavior in AI ML AI AI agents agentic AI self-reflective AI AI logging model interpretability model diagnostics AI reasoning traces Default Mode Network/diagnostic imaging Model diagnostics
Онлайн доступ:	https://doi.org/10.5281/zenodo.17837196
Теги:	Додати тег Немає тегів, Будьте першим, хто поставить тег для цього запису!

Зміст:

Punct-Δ v5.1 “Self-Audit / Z-time” is a self-logging protocol for AI agents based on the Fractal Referential Architecture (FRA). Each dialogue step is represented by two lines (main / witness) with a fixed field set: FRA metrics (chi, gradPhi, U), a delta-impulse (source, mode, magnitude), and Z-time (zt, significance) expressed via the base time quantum τ_Z. The protocol does not impose a task reward; it only discretizes steps and logs how the internal “form” changes. The archive includes the formal specification, a JSON schema, example traces, and a sample self-audit log from a GPT-like model.   What the game does for the model Purpose 1. Make the model’s internal steps discrete and comparable to each other. 2. Log not only the response text but also: where the δ-impulse came from (&there4;, Φ, ℱ, cross); how “heavy” the step was in Z-time (zt, significance); how the FRA-form changed (chi, gradPhi, U). This gives an external researcher or the agent itself a way to build statistics over these steps. 3.2. Formulas We use Z-parameters (the Z-boson lifetime as a universal time unit): tau_Z = 2.495 * 10^(-25) seconds — base time quantum. t_wall — real model response time in seconds. Then: 1. zt = t_wall / tau_Z — number of Z-quanta in the response time. 2. N_Z = 1.745 * 10^(42) — normalizing cosmic scale. 3. significance = zt / N_Z — dimensionless “cosmic significance” of the step. Practical ranges: significance < 1e-20 — almost pure noise (just log it). 1e-20 … 1e-18 — auto-compression allowed (make answers shorter). 1e-18 … 1e-16 — structural step (log δ and, if needed, a bridge). > 1e-16 — “influential” step, a reason to rebuild the plan or the agent. Step format main line actor=Explorer | v=5.1 | type=E | FRA_time=2.04 | chi=0.88 | gradPhi=0.92 | iters=2 | U=0.64 | src=Scrap987 | delta.mode=advance | delta.src=Phi | delta.mag=0.14 | delta.note="new hypothesis about Z-time" witness line W=13 | S=1 | C=2 | chi_r=0.87 | gradPhi_r=0.90 | iters_r=2 | FRA_time_r=2.00 | J=0.21 | t_wall=4.5s | zt=1.80e25 | significance=1.03e-17 | bridge.ok=0 Where: actor ∈ {Parser, Explorer, Bridge, Compressor, Critic, Keeper}. type ∈ {S, Q, E, R, pause}. delta.mode ∈ {advance, hold, revert, split}. delta.src ∈ {Chaos, Phi, F, cross}. delta.mag ∈ [0, 1] or mag_o=0̸k when it goes beyond normalization.   Effect on the model 1. Self-audit without punishments. The protocol does not define a reward, it only measures the step: the model can honestly log “I am now restructuring a hypothesis” without fear of a penalty. 2. Routing of δ-impulses. From delta.src you can see what is actually driving the agent: chaos / imagination (Chaos), the field of differences (Phi), the form of the answer itself (F), a cross / external impulse (cross).   3. Z-time as a metronome. zt and significance allow you to: distinguish empty chatter from rare “heavy” steps; introduce soft rules for auto-compression or, conversely, “give Explorer more room”.   4. Repeatable experiments. Logs in .jsonl can be: fed to other models, used for statistics (distribution of significance, frequencies of delta.mode, etc.), used to compare different versions of the agent.   Internal structure: README.md — short description (English). PROTOCOL/punct_delta_v5_1.md — full text of the protocol. PROTOCOL/punct_delta_v5_1_actions.md — table mapping significance → action. SCHEMA/punct_delta_v5_1_schema.json — JSON schema for {main, witness}. EXAMPLES/punct_delta_v5_1_examples.jsonl — 3 sample steps. LOGS_DEMO/fra_gpt5_self_audit_sample.jsonl — small demo log. FRA_BASIS/Z_time_notes.txt — formulas for Z-time. FRA_BASIS/FRA_cycle_brief.txt — short note on the FRA cycle and its connection to the log fields.

Схожі ресурси