Evidence-Graded Memory with a Doubt Signal: Applying Monotonic Proof-State Advancement Inside Autonomous Agents
Description
Agents built on large language models have the confidence half of verification and none of the doubt half: a tool result that says "success" is treated as ground truth, notes written in an earlier session are trusted without provenance, and effort continues on premises that were never established. This paper moves the monotonic proof-state model of the companion paper (Monotonic Proof-State Advancement for Distributed Workflow Verification, DOI 10.5281/zenodo.19102106) from a verification service into the agent loop, as a memory substrate written only by observation, graded on the confirmation ladder, bound to the state of the world it was observed under, and tagged with the vantage point and its declared blind spots. Beside every fact the substrate computes an uneasiness score from the gap to what a claim requires, reachability, the outcome-learned track record of that signal, novelty, staleness and stakes, and uses it as a second dispatch trigger: confidence gates when to evaluate, uneasiness gates when to re-check.
Findings. A reference implementation of the companion verifier, tested differentially against a full-history oracle and with real LLM judges (Claude Opus 5, Claude Sonnet 5, and two local models), found that level-threshold dispatch evaluates checks before their evidence has arrived under out-of-order delivery (a judge scoring 100% on complete evidence is reduced to 91%); frontier-gated dispatch restores 100%, and exact-key verdict memoization cuts real judge calls by 56% with no verdict changed. On an agent benchmark (Claude Haiku 4.5, five runs per cell, 3.01 USD total): after a committed regression, notes memory reports the project verified 0 of 5 times, the substrate with its doubt merely shown also 0 of 5, and the substrate with doubt allowed to interrupt the claim 5 of 5. Verbose-output savings are modest (7% of input tokens) because the harness already truncates; calibration learns across runs without prompt or rule changes; an observation-only ledger cannot see a hack that makes tests pass. The doubt signal is worthless as information and decisive as an interrupt.
This deposit contains the manuscript (PDF, DOCX, Markdown), the Go reference implementation of the verifier and the memory substrate with their test suites, the real-judge harness with cached model responses, the agent benchmark runner with fixtures, the harness hooks, and every per-run result.
Files
PAPER.md
Files
(1.7 MB)
| Name | Size | Download all |
|---|---|---|
|
md5:b1166d74f84e52ca7abd386e0451f906
|
26.7 kB | Download |
|
md5:4768aba75443e1c237a56d078585ed02
|
33.6 kB | Preview Download |
|
md5:bf6dd069eb14c7ba36465d36ee284872
|
91.4 kB | Preview Download |
|
md5:c6bf6586d93bf2517a872e3cf163d725
|
277.1 kB | Preview Download |
|
md5:5e44d3b9478dae322818cea6bbadfd16
|
1.2 MB | Preview Download |
Additional details
Related works
- Continues
- Preprint: 10.5281/zenodo.19102106 (DOI)