// preprint · 2026-09-12

Evidence-Graded Memory with a Doubt Signal: Applying Monotonic Proof-State Advancement Inside Autonomous Agents

Sterling Morrison · GDF Research · DOI 10.5281/zenodo.22728803

Changed repository, stale notes. Notes memory: 0/5 correct. Doubt shown to the model: 0/5. Doubt allowed to interrupt the completion claim: 5/5. Representation was not enough; enforcement was.

Abstract

Moves the monotonic proof-state model inside the agent loop as a memory written only by observation: facts graded on a confirmation ladder, bound to the code state they were observed under, capped by what the agent's vantage can see, each carrying a cheap uneasiness score learned from outcomes. Building the reference implementation surfaced a premature-dispatch gap in the original theoretical design — under out-of-order delivery, threshold dispatch could evaluate before all required evidence arrived, reducing a perfect judge to 91% across 500,000 simulated instances; frontier gating restored 100%. The efficiency thesis largely did not survive measurement. The correctness thesis did. One small agent model, five runs per cell, constructed scenarios: a pilot, and labeled as one.

The question

Agents built on language models treat a tool result that says "success" as ground truth, and trust notes written in earlier sessions without asking where they came from. People do something different. They hold "the payment provider says success" as probably true while keeping a cheap, always-on sense that it might not be, and they re-check when that sense gets loud.

The paper builds that missing half: a memory written only by observation, graded by how far each observation reached, bound to the state of the world it was seen in, with a doubt score beside every fact.

What we did

First we built the verifier from the companion paper, Monotonic Proof-State Advancement, in Go and tested it against a full-history oracle and four LLM judges. Then we moved its three mechanisms inside a production coding agent through harness hooks: a confirmation ladder for facts, per-check boundary caps for what the agent's vantage can see, and an uneasiness score that rises with staleness, unreachability, novelty, stakes and the signal's track record.

The agent benchmark ran Claude Haiku 4.5 through Claude Code on five scenarios, five runs per cell, for $3.01 total.

What we found

The original design had a gap. Under out-of-order delivery, dispatching a check on confidence level alone evaluated it before all its evidence had arrived. A judge that scored 100% on complete evidence dropped to 91%. Dispatching only when the check's declared evidence had arrived, which we call frontier gating, restored 100%. Exact-key memoization of judge verdicts then cut real judge calls by 56% with no verdict changed.

The decisive result concerns memory across sessions. After a regression was committed, notes memory reported the project verified every time. The substrate with its doubt shown to the model did no better. With doubt allowed to block the completion claim once and say why, the agent re-ran the tests and reported the regression in 5 of 5 runs, for 0.6 cents more per run.

The efficiency story was small. Summarizing verbose output saved 7% of input tokens, because the harness already truncates. And an observation-only ledger can't see a hack that makes tests pass: on the flawed-premise task, every agent with or without the substrate bent the code until both contradictory tests passed.

ArmCorrectTurnsCost per run
Notes memory0/542.6¢
Substrate, doubt shown only0/552.6¢
Substrate, doubt interrupts5/583.2¢
Memory across sessions after a committed regression (Claude Haiku 4.5, 5 runs per arm, medians).

What to do with it

If an agent keeps memory across sessions, bind each remembered fact to the state it was observed under, and let a stale fact stop a completion claim instead of printing a warning. In these runs the warning reached the model and changed nothing.

Limitations

  • A pilot: one small agent model, five runs per cell, scenarios built to contain the boundaries where the substrate can matter.
  • A stronger model that heeds shown context might not need the interrupt. That's untested.
  • Correctness was checked by rule and by reading each final message, not by blind human grading.
  • Verdict-memo hit rates on real traffic are unmeasured. The 56% figure comes from a realistic synthetic workload.

Terms

Questions this paper answers

Cite

Morrison, S. (2026). Evidence-Graded Memory with a Doubt Signal: Applying Monotonic Proof-State Advancement Inside Autonomous Agents. Zenodo. https://doi.org/10.5281/zenodo.22728803

@misc{morrison2026evidence,
  author    = {Morrison, Sterling},
  title     = {Evidence-Graded Memory with a Doubt Signal: Applying Monotonic Proof-State Advancement Inside Autonomous Agents},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.22728803},
  url       = {https://doi.org/10.5281/zenodo.22728803}
}