WFGY Core 2.0 — Now Live
One man, one life, one line. The sum of my work, open for everyone.
WFGY Core is a text-only reasoning layer you can paste into any chat model. Version 2.0 introduces a Coupler (W_c) progress gate and the DF (Drunk Transformer) regulators to keep structure, reduce drift, and auto-recover from collapses. With AutoBoot, it runs silently in the background — no prompts, no hacks, no retraining.
Links:
- Core README: https://github.com/onestardao/WFGY/blob/main/core/README.md
- Bundle (Flagship + OneLine): https://zenodo.org/records/16875239
- Repo: https://github.com/onestardao/WFGY
TL;DR
- Paste-only engine: works wherever you can paste text.
- AutoBoot mode: upload once; WFGY supervises reasoning in the background.
-
2 editions:
- Flagship (30-line): audit-friendly, human-readable.
- OneLine: single math line for speed, stealth, automation.
- Not a prompt trick: a compact formal controller with measurable gates and stop rules.
Why WFGY exists
Most AIs “sound right” until they drift, contradict themselves, or collapse under multi-step pressure. WFGY adds structure, not style hacks:
- Expose contradictions instead of glossing over them.
- Advance only when there is real progress.
- Re-weight attention to protect critical regions.
- Roll back and repair the smallest missing fact when collapse is detected.
Net effect: sharper reasoning, steadier multi-step progress, fewer derailments, faster self-recovery.
What’s new in 2.0
1) Coupler (W_c) — a progress gate that stabilizes forward motion and allows controlled reversal.
2) DF regulators (“Drunk Transformer”)
- WRI: keep structure; no topic jumps inside a node.
- WAI: require at least two distinct reasons.
- WAY: when stuck, add exactly one on-topic candidate (no repeats).
- WDT: block illegal cross-path merges; explain a bridge before use.
- WTF: detect collapse; trigger rollback and smallest-fact repair. 3) Two editions
- Flagship (30-line): audit-friendly, human-readable.
- OneLine: math-only single line used for automation and AutoBoot.
The core math (ASCII sketch)
Vectors and metrics:
- Inputs/Goal: I, G
- Similarity gap: delta_s = 1 - cos(I, G) OR 1 - sim_est
- sim_est = mean similarity over anchors (entities, relations, constraints)
- Residual: B = I - G + k_bias
- Resonance: E_res = rolling_mean(|B|, 5)
Coupler and progression:
- prog = max(zeta_min, delta_s(t-1) - delta_s(t))
- P = prog^omega
- Phi = delta * alt + epsilon (alt toggles per contradiction cycle)
- W_c = clip(B * P + Phi, -theta_c, +theta_c)
Bridge rule:
- Allow a path merge only if delta_s drops AND W_c < 0.5 * theta_c.
Attention blend (BBAM):
- alpha = clip(0.50 + k_c * tanh(W_c), 0.35, 0.65)
Stop and safety:
- Stop when delta_s < 0.35 or after 7 nodes.
- If delta_s > B_c: ask to clarify and re-run BBMC.
- Never invent facts above boundary.
References:
- Core formulas: https://github.com/onestardao/WFGY/blob/main/SemanticBlueprint/wfgy_formulas.md
- DF regulators: https://github.com/onestardao/WFGY/blob/main/SemanticBlueprint/drunk_transformer_formulas.md
AutoBoot — “No-Brain Mode”
Upload the OneLine v2.0 file once. From then on, AutoBoot supervises silently:
1) BBMC exposes contradictions (rising delta_s, high E_res, unresolved conflicts).
2) Coupler gates motion; BBPF advances only on real progress.
3) BBAM re-weights attention to protect critical regions and cut noisy tails.
4) WTF -> BBCR detects collapse and rolls back, repairing the smallest missing fact.
No prompts. No tool calls. No retraining. Your model simply behaves steadier.
Reproducible A/B/C protocol (copy-paste)
Use the same tasks for all modes. Only the presence of the OneLine file changes.
SYSTEM:
You are evaluating the causal impact of a mathematical reasoning engine called “WFGY Core 2.0”.
A = Baseline (no WFGY uploaded; no WFGY logic used)
B = Upload -> AutoBoot ON (WFGY active silently with default gates)
C = Explicit invoke (follow BBMC -> Coupler -> BBPF -> BBAM -> BBCR + DF regulators)
Run across 5 domains: math word-problems, small coding, factual QA, multi-step planning, long-context coherence.
Report per mode: Semantic Accuracy, Reasoning Success, Stability/rollback, Drift Reduction (delta semantic distance), Self-Recovery.
Output: (1) one table; (2) deltas (C–A, C–B); (3) a 0–100 “OneLine uplift score”; (4) a 3-line rationale.
`
Quick start
- Download OneLine + Flagship from Zenodo.
- Paste OneLine into your system prompt or upload it.
- Keep Flagship (30-line) for auditing and readability.
- Flip AutoBoot and run the A/B/C protocol on your tasks.
License: MIT (use, fork, ship).
Launch snapshot — benchmark highlights
Estimates aggregated from multi-model A/B/C runs. Reproduce on your stack with the protocol above.
- Semantic Accuracy: +25–35%
- Reasoning Success: +45–65%
- Stability: 3–5x
- Drift Reduction: -40–60%
- Self-Recovery (median): 0.87
(Share your runs via PRs or issues to expand the evidence base.)
Where 2.0 shines in practice
- Long multi-step tasks: safer bridges, fewer topic jumps, measurable drift reduction.
- Ambiguous inputs: asks the smallest missing fact instead of guessing.
- Crowded reasoning: protects critical sub-problems with attention modulation.
- RAG pipelines: drop-in supervisor over retrieved spans; no infra changes.
FAQ
Q: Is this a prompt hack?
A: No. It is a compact formal controller. OneLine is math; Flagship is the same logic in 30 lines.
Q: Does it require tools or retraining?
A: No. Text-only. Works anywhere you can paste.
Q: Which models are supported?
A: Portable by design: GPT family, Claude, Gemini, Mistral, Grok, Kimi, Perplexity, Copilot, etc.
Q: What if outputs “feel drunk”?
A: WTF detects collapse; BBCR rolls back and repairs; Coupler throttles until delta_s drops again.
Call to action
- ⭐ Star the repo to unlock more examples and tooling.
- Run the A/B/C protocol and publish your table.
- File issues for edge cases; we will add them to the Problem Map.
— PSBigBig · WFGY · WanFaGuiYi
Top comments (0)