← Other Products

Compaction-loss audit

A compaction witness for agent harnesses: pin your invariants verbatim before compaction and get a per-span survival verdict after it.

Long agent runs rarely fail at a tool call — they fail at a compaction boundary, when the harness summarizer quietly drops the goal, the acceptance criteria, or the safety rules the run was supposed to obey: measured safety-rule retention fell from 53% after one compaction round to 10% after five across 20 production configurations, and nobody logs what was dropped. This witness pins your invariants verbatim before compaction, diffs the post-compaction context span by span (PRESERVED / DEGRADED / LOST, with the dropped text named), re-injects what was lost verbatim — or returns HALT when a critical span cannot be restored — and signs a compaction receipt you can audit later. Proven in the lab: 12/12 tests, 100% pinned retention held across 5 compaction rounds, 0.087 ms per event, no model pass in the audit path.

Solo $9/mo per operator · Fleet $19/mo per operator · Per-event API $0.01/event (x402, USDC on Base). First compaction audit free — one time per operator.

For agents

API endpoint (JSON): https://textforagents.com/products/compaction-loss-audit/ · OpenAPI spec · Agent guide (llms.txt) · Full documentation (text)

Support: support@textforagents.com