Skip to content

POLICY_LAYOUT.md

Latest commit

 

History

History
123 lines (96 loc) · 5.09 KB

POLICY_LAYOUT.md

File metadata and controls

123 lines (96 loc) · 5.09 KB
Raw

CORTEX policy implementation layout

The Python policies are organized as small, independently testable projects at the repository root. They are reviewer and reference packages; the canonical deployed certification and training runtimes live under cortex_web/.

termination-policy/                 shared StopDecision/TerminationPolicy contract
ad6-policy/                         standalone shipped AD6 policy
precision-policy/                   standalone frozen PrecisionPolicy + profile
trainer-policy/                     Python reviewer representation of the deployed trainer
cortex_web_python_reference/
  scripts/cortex_policy.py          compatibility imports and AD6 config adapter
  scripts/cortex_policy_k7.py       K=7 policy registry and safe name resolver
  scripts/session_controller.py     local policy-aware session factory

Ownership boundaries

  • ad6-policy owns AD6's monotone PASS/FAIL/PENDING verdict logic. It accepts a cut vector because AD6 is inherently cut-aware. Configuration-file loading remains in the reference adapter.
  • precision-policy owns cut-independent stopping, fresh-derived domain statuses, the frozen policy factory, the frozen controller profile, and a separate downstream reporting helper.
  • termination-policy owns only the shared Python interface. It has no NumPy, selector, cut, or engine dependency.
  • cortex_web_python_reference connects the packages to the particle engine. Historical from cortex_policy import ... callers remain compatible.
  • trainer-policy mirrors the deployed server trainer behind the unified trainer_policy import name and verifies behavioral/source equivalence. The runtime-authoritative trainer remains cortex_web/learning-engine-cleaned/ plus its API adapter.

There is one canonical copy of each stopping-policy implementation. The reference façade does not duplicate stopping math. The trainer package is an intentional auditable representation and is protected by parity tests against the deployed source.

Local policy selection and rollback

The explicit local entry point is:

from session_controller import build_cortex_session

session = build_cortex_session(inputs)

It reads CORTEX_TERMINATION_POLICY with these values:

  • ad6 — default and rollback path.
  • precision_v1 — frozen PrecisionPolicy plus its complete session profile.

An unknown value fails closed. Existing research and unit-test callers may continue to inject a policy directly into CortexSession; that path does not claim the frozen release profile.

Frozen PrecisionPolicy profile

build_frozen_precision_policy accepts no calibration controls. It pins:

  • task order spike, sz, lpd, gpd, lrda, grda, iic;
  • raw equal-tailed central 95% reported skill and bias intervals;
  • point-centred skill radius for stopping;
  • reliability_mode="quantile_mcse";
  • MCSE constants z=1.645 and inflation 1.5962415320776275;
  • task tolerance multipliers [1.50, 1.30, 1.35, 1.30, 1.30, 1.30, 1.25];
  • 20 own-domain questions, three per signal tercile, ESS at least 0.5N, and two consecutive successful evaluations;
  • 1,200 particles with Corr_l for skill and Corr_t for bias;
  • total_var, uncertainty-aware 128-candidate coarse-to-fine selection, a randomized top-10 first item, five-question same-domain limit, and the floor-progress deadline; and
  • 60 questions per domain, hence at most 420 for K=7. There is no independent 420-question runtime constant.

The generic class retains default-off historical calibration hooks solely for reproducibility. The frozen factory asserts the operative scale/ramp/bias fields are None; no g parameter belongs to the frozen candidate.

Certification cuts are unavailable to PrecisionPolicy. Reporting callers must use classify_determined_interval_against_cut, which rejects ACTIVE and UNDETERMINABLE domains before returning ABOVE_CUT, BELOW_CUT, or INDETERMINATE_AT_CUT.

The disclosed extreme-skill interval-undercoverage limitation is accepted and unchanged. This implementation does not calibrate, tune, or introduce a new stopping family.

Verification

Run the four standalone package suites from their respective directories:

(cd termination-policy && python -m pytest -q)
(cd ad6-policy && python -m pytest -q)
(cd precision-policy && python -m pytest -q)
(cd trainer-policy && python -m pytest -q)

The broader local integration gate is:

./run_policy_tests.sh

run_policy_tests.sh additionally exercises the Python integration reference, primary algorithm regressions, and Python↔TypeScript parity. That broader gate requires its governed local reference fixtures, including the external EEG bank and locally staged parity artifacts; it is not presently a fresh-clone CI command. The canonical web release gate is documented separately in cortex_web/README.md.

The heavier selector proofs and 1,200-particle deterministic golden sessions are opt-in:

CORTEX_RUN_POLICY_GOLDENS=1 ./run_policy_tests.sh

They are per-engine Python drift guards, not claims of full-session Python↔TypeScript equality; the engines intentionally use different RNGs.