IDKMesh

AVE-core Shadow Adapter

Status: stacked research adapter for N2/N3 dry-run evidence
Date: 2026-09-22
Depends on: Adaptive Policy Shadow Contract (PR #637 / issue #636)
Related research: AVE (PR #622 / issue #621), E025 learned verifier reliability

Purpose

The AVE-core shadow adapter is the first concrete consumer of the common adaptive-policy shadow contract.

It answers only:

Given one canonical WorkUnit, its already-owned EvaluatorPlan, and a point-in-time verifier observation pool, which verifier portfolio would AVE-core recommend in shadow mode?

It does not:

• modify the EvaluatorPlan;
• dispatch a verifier;
• execute candidate code;
• approve or merge;
• relax required validators;
• grant project spending authority;
• convert model/provider identity into trust.

Inputs

WorkUnit v0.2

The adapter reuses:

• exact WorkUnit id/version;
• provenance source revision;
• security risk class;
• required validators;
• independent-verification requirement;
• minimum independent verifier count;
• zero-project-spend budget lane.

Existing EvaluatorPlan

The adapter treats the current EvaluatorPlan as the named baseline.

It verifies:

• exact canonical WorkUnit digest;
• WorkUnit id/version binding;
• source revision binding;
• required validator coverage;
• verifier-distinct-from-worker policy.

The baseline remains visible even when AVE would not choose the same verifier.

That is necessary for a fair N3 comparison.

Verifier Observation Pool v0.1

Schema:

schemas/verifier-observation-pool-v0.1.schema.json

Each candidate includes:

• current availability;
• verifier family;
• supported validator IDs;
• expected review units and queue load;
• independence/correlation signals;
• domain-scoped reliability evidence;
• known-bad probe diagnostics.

The pool is point-in-time evidence, not a durable reputation ledger.

E025 constraint: no global verifier reputation

E025 showed an important failure mode:

• learned reliability/dependence weighting can improve stable regimes;
• the same learned model can perform materially worse after reliability or dependence shifts;
• more calibration history can shrink uncertainty without protecting against a changed environment.

Therefore AVE-core gives positive reliability weight only when all are true:

basis == live_outcomes
domain == current task domain
shift_warning == false
sample_count > 0
observed_at is not future-dated
age(observed_at, captured_at) <= reliability_max_age_days

Otherwise reliability is neutral:

mean_reliability = 0.5
positive_live_samples = 0

This includes:

• synthetic evidence;
• unknown evidence;
• probe-only evidence;
• stale live outcomes;
• wrong-domain live outcomes;
• live outcomes under a shift warning.

Probe rule

Known-bad probes have asymmetric meaning.

Passing probes

Passing probes does not create positive verifier trust.

A verifier with:

99 / 99 known-bad probes handled correctly
0 live outcomes in the current domain

does not outrank a verifier merely because of those probe passes.

Breaching a known-bad probe

A known-bad breach is a strong diagnostic signal.

For AVE-core v0.1:

known_bad_probe_breaches > 0
 -> verifier remains visible in the frozen observation state
 -> current EvaluatorPlan baseline can still be named
 -> verifier is excluded from AVE shadow portfolio selection

This is intentionally asymmetric and should be revisited only with real representativeness evidence.

Hard gates

AVE never compensates for a failed hard gate.

The adapter records:

• WorkUnit v0.2 support;
• exact WorkUnit digest binding;
• WorkUnit id/version binding;
• source revision binding;
• verifier-pool repository/WorkUnit binding;
• evaluator sovereignty;
• zero-project-spend lane;
• independent-verification policy;
• required-validator binding.

Any failure forces:

selected_choice_id = null

The plan still records the baseline and failure reasons for audit.

Risk-adaptive target

The shadow target is:

target =
  max(
    WorkUnit.minimum_independent_verifiers,
    risk floor
  )

risk floor:
  low      -> 1
  medium   -> 2
  high     -> 3
  critical -> 3

AVE-core v0.1 requires the recommended target-sized portfolio to use distinct declared verifier families.

This is not a claim that different family labels prove independence.

It is a research heuristic motivated by correlated-failure evidence. The shadow plan explicitly states that family labels are not statistical proof.

If the requested family-diverse portfolio is unavailable, AVE abstains rather than reducing the target.

Validator coverage

Portfolio validator coverage is computed as the union across selected verifiers.

A portfolio is viable only when it covers every validator that is required by:

• the WorkUnit; or
• the existing EvaluatorPlan.

AVE does not remove or downgrade required checks.

Lexicographic selection

Among viable portfolios, v0.1 uses an explicit lexicographic order:

1. more distinct verifier families;
2. fewer shared-model/shared-runtime correlation signals;
3. more verifiers with fresh same-domain live-outcome evidence;
4. more live-outcome samples;
5. higher posterior mean reliability;
6. fewer review units;
7. lower queue load;
8. deterministic verifier IDs.

There is intentionally no opaque global “trust score.”

Replay normalization

The verifier pool is canonicalized before shadow-plan hashing:

• capture and reliability timestamps normalized to UTC;
• verifier candidates sorted by ID;
• already-validated validator ID lists sorted;
• already-validated reliability source refs sorted;
• limitations sorted/deduplicated.

Equivalent pool orderings therefore produce the same shadow plan.

N3 evidence workflow

A real N3 capture should occur before verification outcome:

WorkUnit + EvaluatorPlan + current verifier pool
 -> AVE shadow adapter
 -> freeze adaptive-policy-plan
 -> real verification continues unchanged
 -> observed VerificationResult/outcome
 -> adaptive-policy outcome join
 -> cohort evaluator

The real EvaluatorPlan remains unchanged throughout the cohort.

Historical fixtures are compatibility tests, not N3 evidence

Existing repository records such as:

results/verification/node-e2e-replay-2026-08-30/

are useful to test canonical WorkUnit/EvaluatorPlan compatibility.

They cannot become N3 evidence by replaying AVE today because their verification outcomes are already known.

That would violate the anti-hindsight rule.

Promotion gate

This adapter should remain shadow-only until a real pre-outcome cohort shows:

• a meaningful shadow-vs-baseline disagreement rate;
• intact hard gates;
• measurable review/human-attention cost;
• enough delayed outcome evidence to identify useful future controlled tests;
• no evidence that a simpler deterministic diversity rule provides the same recommendations at lower complexity.

N3 observational data remains descriptive and does not establish the causal effect of an unexecuted AVE portfolio.

CLI capture

Use tools/ave_shadow_adapter_cli.py to freeze a pre-outcome AVE plan:

python tools/ave_shadow_adapter_cli.py \
  --repository MSKazemi/idkmesh \
  --work-unit path/to/work-unit.json \
  --evaluator-plan path/to/evaluator-plan.json \
  --verifier-pool path/to/verifier-observation-pool.json \
  --maturity N3 \
  --input-ref work-unit:real-id \
  --output evidence/adaptive/plans/ave-real-id.json

The CLI refuses to overwrite an existing frozen plan. It writes evidence only; it does not execute or dispatch the recommended portfolio.

An illustrative pool shape is available at examples/verifier-observation-pool.example.json. It is synthetic and must not be cited as real verifier performance evidence.

This site is open source. Improve this page.