Skip to content

ROADMAP — module segregation

The pipeline splits into five modules. Each is independently useful, independently testable, and depends only on the modules before it. Nothing later can be trusted if something earlier is wrong, so the order is not negotiable.

  M1 Detection        "is this one node biased, and how sure are we?"
       |                 pure math + LLM probes. no graph, no framework.
       v
  M2 Propagation      "how much does that node's bias matter to the whole run?"
       |                 graph centrality + accumulation over time.
       v
  M3 Control          "stop the run when it matters, deterministically."
       |                 threshold, freeze, reroute. LangGraph adapter lands here.
       v
  M4 Intervention     "repair the biased state and resume."
       |                 skeptic debate + trust pruning, or MADERA reasoning edits.
       v
  M5 Evidence         "prove what happened, in a form someone can audit."
                         causal trail, SARIF, HTML/JSON reports.
# Module Owns Ships as Depends on
M1 Detection core Counterfactual perturbation, divergence estimation, per-node noise floor, significance testing, SDC priors v0.1 nothing (numpy only)
M2 Propagation model DAG protocol, transposed Katz centrality, Bayesian error-history score, composite dependency weights, multi-scale (fast+slow) bias-corrected EWMA, superstep reduction v0.2 M1
M3 Control plane Two-threshold hysteresis breaker, edge-level interception, Normal→Warning→Intervention→Recovery state machine, router self-monitoring, payload freezing, BiasState, LangGraph Command() halt/reroute v0.3 M1, M2
M4 Intervention Skeptic panel (≥2, diverse provenance) + Send() fan-out, trust-graph pruning with anonymization + no-prune-on-dissent guardrail, MADERA diagnose→retrieve→rewrite v0.4 M3
M5 Evidence Append-only causal trail, extended SARIF 2.1.0 export (trigger reason, routing entropy), HTML/JSON reports, ablation evaluation protocol + error budget v0.5 M3, M4

Mechanism additions in M2–M5 are banked from the 2026-07 external review — see DESIGN.md §8 and reviews/2026-07-external-review-response.md.

Why this order, and why M1 gets disproportionate effort

M2 multiplies M1's score by a weight. M3 compares it to a threshold. M4 fires when that threshold trips. M5 reports on all of it. Every one of those is a transformation of a number M1 produces. If B_i is not a calibrated, meaningful quantity, then M2 is weighting noise, M3 is thresholding noise, M4 is spending real inference budget repairing outputs that were never biased, and M5 is generating audit reports that certify nothing.

There is no way to recover from a bad M1 further down the pipeline, and a plausible-looking B_i that is actually measuring sampling temperature will not announce itself — the system will run, produce numbers, trip occasionally, and be entirely wrong. So M1 is not "the easy first bit." It is the only module with a genuine research question in it, and it gets built slowly, with an empirical validation study attached.

M2 through M5 are comparatively well-understood engineering: known algorithms (Katz, EWMA) and known framework primitives (Command, Send, TypedDict reducers). They are work, but they are not risk.

Current status

  • M1 — shipped as v0.1.0. Complete and characterized: plans/PHASE-1.md, calibration study. WP6 (SDC) deferred as optional.
  • M2 — shipped as v0.2.0. Complete and characterized: plans/PHASE-2.md, propagation study. DAG, transposed Katz, n-stable magnitude, composite weight, multi-scale accumulator, DoT harness.
  • M3 — shipped as v0.3.0. Complete and characterized: plans/PHASE-3.md, control study. Two-threshold hysteresis breaker, recovery state machine, threshold calibration, routing seam, and the LangGraph reference adapter (edge-level interception via a staging buffer).
  • M4 — shipped as v0.4.0. Complete and characterized: plans/PHASE-4.md, intervention study. Skeptic panel (+ LangGraph Send fan-out), trust graph (anonymized, no-prune-on-dissent), MADERA repair, and the runner wiring it into M3.
  • M5 — shipped as v0.5.0. Complete and characterized: plans/PHASE-5.md, evidence study. Append-only audit trail, SARIF 2.1.0 export (schema-validated in CI), self-contained HTML/JSON reporting, breach trace-back.

All five modules are shipped. The pipeline runs end to end: detect → weight + accumulate → control → intervene → audit. It remains demonstrated on the ground-truth fake, with no validated-performance claims on real models (benchmark reproduction is still unscheduled).