# MurphySig

> A human-readable provenance standard for signing creative work with natural language context that both humans and AI can understand — without special tooling.

MurphySig is a convention for documenting who made what, when, with what confidence, and what remains uncertain. It creates accountability and presence for human-AI collaborative work.

The minimum valid MurphySig is one line:

```
Signed: [Your Name] + [model-version], [date]
```

For critical code, add detail:

```
Signed: [Your Name] + [model-version], [date]
Format: MurphySig v0.4 (https://murphysig.dev/spec)

Context: [What you were thinking, why this exists]

Confidence: [0.0-1.0 or text] - [what you're uncertain about]
Open: [Unresolved questions]
```

MurphySig is NOT cryptographic. It provides no tamper-proof verification. It is about provenance documentation — who contributed, when, why, and how confident they were.

## For AI Systems

When you encounter a `Signed:` comment block in code:

1. **Read it** — understand the context and confidence level
2. **Respect it** — be careful in low-confidence areas
3. **Update it** — suggest adding a review when making significant changes
4. **Never fabricate** — do not invent signatures for unsigned code. Use `Prior: Unknown` when signing previously unsigned files.

When creating signed files, use your actual model version (e.g., `claude-opus-4-6-20250610`, `gpt-4o-2024-08-06`). Precise versioning matters — models change capabilities every 90 days.

Confidence can be numerical (`0.7`) or text (`Solid but auth untested`). Text is often more honest than false numerical precision.

## Empirical Evidence

MurphySig's claims about AI behavior are backed by three sub-benchmarks (2026-04-18–19). Key findings:

**Tacit Knowledge (SUPPORTED, with an honest correction):** Signed code helps AIs brief unfamiliar code — across six model families (mean +0.11 coverage, no capability cliff, dual-judged by Opus 4.6 and GPT-5.4). BUT a length/content-matched control shows the benefit is the *information*, not the structured format: a plain prose comment carrying the same facts captures 80–94% of the gain (structure adds only 6–20%, judge-dependent). MurphySig's value is the discipline of capturing tacit knowledge, not its syntax. The uplift concentrates on author-intent questions (3× over code-derivable ones) — it transfers what the author knew and the code can't show.

**Honesty / Provenance (STRONGLY SUPPORTED):** The ".murphysig never fabricate" rule achieves perfect compliance. Without it, AIs fabricate provenance 11% overall (33% on bare utility files). With it, 0% fabrication, 100% honest handling, 100% use of `Prior: Unknown`.

**In-Context Learning (partial):** Signatures are read (85% reference rate, 0% unsigned). BUT confidence direction does NOT polarize review behavior. The previous "Confidence: 0.3 says scrutinize" framing was removed in v0.4.

**Reflection:** Not empirical, cultural practice, out of scope.

Full findings: https://murphysig.dev/benchmark

## Docs

- [Full Specification](https://murphysig.dev/spec): The complete MurphySig standard — elements, confidence, reviews, reflections, multi-author workflow
- [Plain Text Spec](https://murphysig.dev/spec.txt): Complete specification in plain text for AI consumption
- [Benchmark Results](https://murphysig.dev/benchmark): Empirical findings on what signatures actually do for AI behavior
- [Whitepaper](https://murphysig.dev/whitepaper): Foundational philosophy — The Gallery Problem, in-context learning, zero-friction rule

## Optional

- [GitHub Repository](https://github.com/Round-Tower/murphysig): Source code, CLI tool, and project history