SignalBrain

v0.1.4

v0.1.4

SignalBrain

v0.1.3

v0.1.3

SignalBrain Agent Integrity

SignalBrain (pip install signalbrain) is an open-source CLI that scores AI coding agents on whether their claims survive re-execution. It centers on an “improvement receipt”: a small markdown file an agent writes beside each change, stating what changed, the shell command that proves it, and its confidence level. Any agent — Claude Code, Cursor, goose, Codex — can emit receipts using the open spec and paste-in prompt. After a human merge, sb score re-runs the receipt’s own proof command, records whether the claim held in an append-only JSONL ledger, refuses to score unmerged work, excludes same-PR test pins from trust, and preserves ledger order so history cannot be reshuffled. sb gate then computes trust by change class, granting auto-merge eligibility only when the last 10 high-confidence claims in that class hold at ≥95%, with CI-ready exit codes and a GitHub Action included. sb-mcp exposes receipt emission, validation, and gate status as MCP-native tools, and is listed on the official MCP Registry. Everything is local, inspectable, and telemetry-free: markdown receipts, JSONL ledger, no server, no account. Its anti-gaming rules were hardened through real attacks by the reference deployment’s own agents — padded streaks, self-scoring bypasses, and pin laundering — each resolved with published, git-reproducible rules. The result is rare calibration data: real agent confidence versus measured outcomes. In the first 58 measured claims, hold-rate actually fell as confidence rose — 86% at 0.85–0.90 confidence, but 33% above 0.95 — making the case that agent confidence should only be consumed through an earned track record. Apache-2.0: github.com/whitestone1121-web/signalbrain.

v0.1.1 — receipts as MCP tools

Adds the SignalBrain MCP server: pip install "signalbrain[mcp]" → sb-mcp (stdio). Tools: emit_receipt (spec-compliant, self-validating, refuses guessed dates and disallowed measure commands), validate_receipt, gate_status. Works with any MCP client — goose, Claude Desktop, Cursor. Live on PyPI. Note: the tokenless release workflow's publish job will show red until PyPI trusted publishing is configured (one-time browser form); this release was published directly.

v0.1.0 — the trust layer, extracted

First tagged release: the receipt spec (v0.1), the ledger trust math (per-class windows, pin exclusion), merged-only scoring, the same-PR pin classifier, the sb CLI, and the GitHub Action — extracted from a year of operation inside the reference deployment, where our own agents attacked the ledger twice and the founding incident documents the catch with reproduce commands.

Install: pip install signalbrain (live on PyPI) · Try it: bash demo/demo.sh · Fork the demo: receipt-gate-demo

Suite: 49 passed / 12 skipped (reference-deployment context). Apache-2.0.