Date: 2026-08-28
User request: continue strengthening MSKazemi/idkmesh mathematically and implement the result with GitHub-native automation.
The sharp Adversarial Evidence Envelope merged through PR #216 as:
17e77b07c700afd04ce41a5ff72641653608436c
A direct default-branch read confirmed that exact commit as canonical main and still reported:
protected: false
required status-check enforcement: off
The canonical main Adversarial Evidence Envelope workflow run:
33204917325
completed successfully. Its retained artifact:
9699250868
contains byte-identical Python 3.11 and 3.13 JSON payloads. The actual mathematical payload SHA-256 is:
6290b45db8db841520f292162e1bea1cccd4da92d9424952338a733d57ecf984
This matches the exact PR evidence hash, so the merged implementation preserved deterministic behavior on canonical main.
Before #216 merged, main also added docs/architecture/ALGORITHM_COLLABORATION_FABRIC.md through PR #214. That architecture already reserves a distinct pipeline stage:
verify -> aggregate evidence -> control flow -> learn -> govern/integrate
and explicitly states that algorithms should not negotiate one global fitness score or manufacture integration authority.
Its role table already includes correlation-aware aggregation and the Sequential Evidence Kernel, but the ACF was authored concurrently with the newer Anytime Drift Guard and Adversarial Evidence Envelope.
The clean convergence task is therefore not to replace the ACF. It is to make its aggregate evidence stage precise and executable using the now-canonical evidence algorithms.
The repository now has several evidence channels whose numbers are not algebraically interchangeable:
correlation-aware Bayesian evidence
count-contamination honest-report envelope
anytime sequential confidence sequence
temporal change detection
provenance/discrimination checks
hard governance guards
A tempting but invalid design would turn them into something like:
combined_confidence = posterior * contamination_score * drift_score * ...
No reviewed joint probabilistic model justifies that product. “No drift detected” is not a probability of correctness, and a sharp worst-case contamination interval is not a Bayesian likelihood term.
Implement a typed evidence lattice rather than a scalar aggregate.
Every signal must carry:
signal_id
producer
scope_id
claim_id
signal_type
observation_model
evidence_mass
uncertainty
assumptions
failure_modes
evidence_refs
source_revision
authority_ceiling
payload
Before composition, every channel must refer to exactly the same:
scope_id
claim_id
source_revision
This prevents evidence from different candidates, different propositions, or stale revisions from being mixed merely because the numeric direction is favorable.
The executable composer requires:
provenance
discrimination
correlation
contamination
sequential
drift
hard_guard
Each keeps its own model and payload.
The composer deliberately returns:
{
"composite_confidence": null,
"scalarized_score": null,
"double_counting_claim": false
}
A downstream consumer that wants a new scalar must define a separate reviewed joint model.
The composer retains every detected blocker but exposes one deterministic bounded recommendation:
hard guard failure
-> guarded
invalid provenance
-> observe_invalid_provenance
non-discriminating instrument
-> observe_non_discriminating
low correlation-adjusted effective evidence
-> observe_correlation_uncertainty
robust contamination rejection
-> insufficient_support
contamination uncertainty
-> observe_adversarial_uncertainty
drift detected
-> observe_drift
sequential evidence not yet sufficient
-> observe / insufficient_effect
all channels adequate
-> experiment_candidate
The priority controls the recommendation only. All blockers remain in the evidence artifact so fixing the first blocker cannot hide secondary problems.
Correlation and contamination remain orthogonal:
correlation channel:
probabilistic evidence strength under a declared reliability/dependence model
contamination channel:
deterministic honest-report mean range under <= f arbitrary accepted reports
Sequential evidence controls optional stopping under its bounded common-mean assumptions.
Drift controls temporal pooling assumptions.
None of these establishes external truth, Sybil resistance, or integration authority.
The recently added IDKGraph human review-session validator operates one layer earlier. It validates/describes one review artifact and its provenance disclosures. The Evidence Aggregation Fabric composes several already accepted typed signals for the same exact claim/revision.
Structural validity of one review does not prove independence or immunity to arbitrary-report faults; conversely, an adversarial envelope does not validate a review’s provenance.
This convergence adds:
scripts/evidence_aggregation_fabric.py;tests/test_evidence_aggregation_fabric.py;docs/architecture/EVIDENCE_AGGREGATION_FABRIC.md;.github/workflows/evidence-aggregation-fabric.yml;The tests cover all-clear nomination, hard-guard dominance, invalid provenance, non-discrimination, low effective evidence despite high posterior probability, adversarial uncertainty/rejection, drift, insufficient sequential effect, multiple blocker retention, missing channels, scope/claim/revision mismatch, duplicate channels, evidence-reference requirements, invalid decisions, and invalid effective-evidence mass.
The strongest operational output remains:
experiment_candidate
not merge, approve, activate, write, or reset.