diff --git a/AGENTS.md b/AGENTS.md
deleted file mode 100644
index a0b7f5e..0000000
--- a/AGENTS.md
+++ /dev/null
@@ -1,35 +0,0 @@
-# AccountPulse Repository Instructions
-
-These rules govern every session and agent working in this repository.
-
-## Runtime and compute
-
-- Run all scientific Python work under Ubuntu-22.04 WSL.
-- The Windows-backed source repository is `/mnt/c/Users/bjw-0/Downloads/AccountPulse`.
-- The only primary project environment is `$HOME/.local/share/accountpulse/project-venv`.
-- Set `UV_PROJECT_ENVIRONMENT` to that environment. Do not create or substitute another venv.
-- Put large temporary files and caches under `$HOME/.cache/accountpulse`.
-- Preserve the qualified CUDA Torch/XGBoost/CatBoost stack recorded in `artifacts/bootstrap/`.
-- GPU-capable model stages must use CUDA and fail loudly if CUDA is unavailable. Never silently fall back to CPU.
-- There is one physical GPU. Run at most one heavy GPU training process at a time and make no multi-GPU claims.
-- CPU ETL, statistics, reporting, and tests may run concurrently when safe.
-
-## Scientific integrity
-
-- AccountPulse is predictive decision science, not a causal outreach model. Keep predictive, causal, and sequential-allocation estimands separate.
-- Protected outcomes may not influence feature, model, hyperparameter, calibration, threshold, or claim choices.
-- Never access LOCKED outcomes before a complete freeze. A scientific change after outcome exposure requires a new protocol identity.
-- Predictive ranking must never be described as causal lift or treatment effect.
-- Retain negative results, failed eligible candidates, and strong baselines. Do not tune away protected negative findings.
-- Treat Maven CRM as `FICTITIOUS_PUBLIC`; Olist, UCI Online Retail II, and verified Criteo as their documented public evidence classes. Never upgrade public, fictitious, semi-synthetic, local, or replay evidence into enterprise-production claims.
-- Enforce point-in-time availability, horizon maturity, competing-risk semantics, chronological splits, cold-entity separation, and dependence-aware inference.
-- Leakage diagnostics L1-L3 are `INELIGIBLE_LEAKAGE_DIAGNOSTIC` and can never enter selection.
-- Hindsight oracles are retrospective references only.
-
-## Execution and evidence
-
-- Use the pipeline state, experiment registry, source manifest, split manifest, freeze manifest, claim ledger, and canonical DuckDB/Parquet evidence to resume work.
-- Classify failures as `INFRASTRUCTURE_FAILURE`, `SCIENTIFIC_FAILURE`, `EXTERNAL_BLOCKER`, or `EXPECTED_NEGATIVE_RESULT`.
-- An external blocker in one track must not halt unrelated tracks.
-- Final claims and resume bullets must be generated only from validated canonical evidence.
-- Never use `force` to bypass protected scientific gates.
diff --git a/README.md b/README.md
index bc7042f..6109ed5 100644
--- a/README.md
+++ b/README.md
@@ -1,6 +1,6 @@
 # AccountPulse v1.0
 
-Evidence-controlled predictive decision science for capacity-constrained prioritization. Preserved external-track protocol: `AP-V1-PROTOCOL-20260922-R2`; Track-A protocol: `AP-V1-TRACKA-20260923-A1`; separate causal bridge: `AP-V1-CAUSAL-BRIDGE-20260922-D2`. AccountPulse prediction is not itself causal outreach modeling.
+AccountPulse evaluates how predictive scores and calibrated value estimates rank accounts when outreach capacity is limited. It compares frozen business heuristics with ML models on public CRM and transaction benchmarks, and studies causal uplift in a separate experiment. Predictive ranking results are not evidence of causal outreach impact.
 
 ## Current outcome
 
@@ -12,6 +12,8 @@ On Criteo, the response and validation-selected uplift top-decile lists overlapp
 
 ## Reproduce
 
+Frozen protocols: external tracks `AP-V1-PROTOCOL-20260922-R2`, Maven Track A `AP-V1-TRACKA-20260923-A1`, and separate causal analysis `AP-V1-CAUSAL-BRIDGE-20260922-D2`.
+
 The repository publishes source, manifests, and compact JSON results. Raw benchmark files,
 derived Parquet evidence, model weights, and local runtime databases are excluded. Full
 research reproduction requires acquiring the documented sources under their terms and a