idkmesh

ADR-0007: Treat independent verification and verification debt as control-plane primitives

• Status: Accepted for experimentation
• Date: 2026-08-28
• Related: issues #5, #14; PR #47

Context

IDKMesh can already represent a bounded WorkUnit and a worker ResultManifest, but a scalable swarm needs a protocol boundary between worker claims and independent evidence.

It also needs a way to prevent candidate generation from outrunning verification. Raw queue length is insufficient because pending candidates have different risk, uncertainty, blast radius, verification cost, and evidence correlation.

Decision

1. Independent VerificationResult is a separate protocol object

The execution/evidence chain is:

WorkUnit
 -> worker attempt
 -> ResultManifest
 -> independent verifier
 -> VerificationResult
 -> integration/human policy decision

A worker cannot self-accept its candidate. A verifier can recommend acceptance, rejection, escalation, or insufficient evidence, but the VerificationResult does not itself authorize a canonical merge/integration action.

schemas/verification-result-v0.1.schema.json is the initial experimental contract.

2. Verification debt is a first-class flow-control signal

IDKMesh will experimentally model pending verification burden as risk-weighted debt rather than only candidate count.

The initial reference model increases debt with:

• candidate risk;
• uncertainty;
• blast radius/impact;
• estimated verification cost;
• insufficiently diverse/correlated evidence.

This is a controller signal, not a probability or permanent quality score.

3. Generation fan-out should respond to verification pressure

The first reference controller is Risk-Weighted Verification Backpressure (RWVB), implemented in experiments/verification_backpressure.py.

RWVB:

• prioritizes verifier capacity using risk-clearing pressure per estimated verification cost;
• includes an age/starvation guard;
• reduces generation fan-out when verification debt exceeds capacity watermarks;
• permits fan-out growth again after debt falls into a safe region.

Rationale

The project objective is verified useful work, not maximum candidate production. If generator count grows while independent evidence capacity stays fixed, verification backlog and escaped-risk pressure can dominate any benefit from additional agents.

Making verification pressure part of the control loop creates negative feedback:

high unverified risk
 -> higher verification pressure
 -> lower generation fan-out
 -> debt clears
 -> generation can expand again

This also aligns with IDKMesh’s earlier decision that verification must scale with generation.

Mathematical inspiration

RWVB is inspired by queueing/network MaxWeight/backpressure methods, especially work originating with Tassiulas and Ephremides and later stochastic-network optimization methods.

The classical throughput/stability theorems do not automatically transfer to IDKMesh’s heuristic risk/verification model. The analogy is a research hypothesis and must be benchmarked against simpler baselines.

See docs/research/VERIFICATION_DEBT_AND_BACKPRESSURE.md.

Consequences

Positive:

• worker self-report and independent evidence have an explicit trust boundary;
• verifier independence/correlation can be measured;
• the local Verified Swarm Runner has a concrete verification artifact to produce;
• verification capacity can actively constrain candidate-generation rate;
• issue #14 now has an executable algorithmic baseline.

Costs/risks:

• risk and cost estimates may be wrong or gameable;
• aggressive feedback parameters can oscillate;
• high-risk work can monopolize verifier capacity;
• evidence diversity is difficult to estimate robustly;
• more protocol objects increase implementation complexity.

Required experiments before stronger adoption

Compare FIFO, highest-risk-first, cheapest-first, and RWVB under controlled workloads with seeded defects and increasing generation fan-out. Measure escaped defects, accepted throughput, total verification debt, queue latency, verifier cost, human attention, false rejection, and evidence correlation.

Parameters remain experimental until evidence supports defaults.

Implementation references

• docs/research/VERIFICATION_DEBT_AND_BACKPRESSURE.md
• docs/research/VERIFICATION_BACKPRESSURE_BENCHMARK.md

This site is open source. Improve this page.