Project 4 signal-to-portfolio protocol alignment

This note records a normative correction to the documented signal-to-portfolio protocol. The correction was made after reading the historical A5 and A6 outputs, so it is not a blind preregistration, a new strategy search, or evidence about real securities. The only returns in this evidence are from the fixed constructed universe with seed 20261210.

Fixed semantics

universe.min_names_per_date = 60 counts names listed at signal time. preprocess applies that gate before winsorization, industry/size neutralization, or cross-sectional standardization; long_short_weights applies the same gate at the portfolio entry point. A date with fewer than 60 listed names has no signal and stays in cash with reason below_min_listed_names. The field winsorization minimum of 30, z-score minimum of 10, style coverage, and combination rules remain separate and unchanged. The production rule/configuration files were not edited.

Industry-neutral projection uses only names listed on the signal date. It builds rows only for industries present in that support. Listed names without a valid alpha score remain available as projection coordinates and can receive a hedge weight; this means the projected portfolio is not necessarily composed only of the original extreme quintiles. Names outside the signal-date support receive exactly zero. Once the homogeneous net and industry constraints are satisfied, the row is scaled by one positive scalar to configured gross 2. A zero/degenerate projection, or a row without valid long and short directions, becomes cash with an explicit reason. Missing industry labels, mismatched support axes, non-boolean membership, and non-finite non-missing scores fail explicitly.

The tests also include a mathematical counterexample showing that scaling a solution with a nonzero constraint target can change the target. This is only a demonstration of why the production projection remains homogeneous; it does not add a nonzero-target strategy.

Neutrality is a signal-time target. Cancelling a known delisting at execution without resizing, partial fills, and subsequent price movement can change realized industry, net, and gross exposures. Those deviations are reported below. The execution-date delisting state is used only by the existing A6 execution/ledger step; it does not enter signal formation. The constructed data has no halted-market, borrow-availability, or real-price model.

Red/green and consumer evidence

tests/test_p04_portfolio_support.py dynamically loads its source from C13_SOURCE_ROOT and verifies 59/60 listed membership, strict support, listed no-alpha hedge coordinates, future-membership perturbation, the single-listed-name industry case, all-cash reasons and finite diagnostics, input alignment/classification failures, and the nonzero-target scaling counterexample. With the frozen canonical signal source (e16d97c61fea2f8cab1e7b80a6f20442232507e8b35b4bb2c1a1a2b51f0d0373), the final red run was 8 failed, 1 passed. With the candidate signal source, all 9 support tests passed. Logs and exact source paths are recorded under qa/a10/teaching_revision/C13/r1_20261014/engineering/.

The final candidate consumer run passed 23 tests and deselected 2 tests that invoke P5. The original combined Chapter 29 P4/P5 test remains in tests/test_ch29_project_interfaces.py, including its P5 ledger, hedge-ratio, status, and output-overwrite assertions; its former expectation of off-list zeroing was updated to the new strict-support result. The separate P5 training-window test is preserved unchanged. Neither P5 test was run in this batch. The P4-only consumer exercised the actual candidate p4_demo() and checks accounting, fee monotonicity, support, and post-execution exposures. A separate old/new check confirmed the former restrict_to_listed silently accepted reordered axes while the candidate rejects them.

The fixed rules and signal-contract bytes still match their frozen canonical hashes:

• frozen_rules.json: 5ee18f1c3b88cbeb4f14d4973955287a96e7eebe840f97cf0d354f73e41b1641
• signal_contract.json: 8890d3199fcba8a703dceaaafd8f6bb8758869077d76ea268c22ef89bccd6539

The final candidate source hashes used by the engineering run are:

• preparation/signal_layer.py: b5357c4122116bacecb585ffe6625a7d14b9dc85c3f65c7ea1dffc744e0a4fd8
• backtest_preparation/backtest_interfaces.py: 356b2cf9e2786d2b75f69581ee9345077ad466a73b089a575a789dfd89e16562

Fixed constructed-data run

build/batches/A10-C13/r1_20261014/runs/engineering_01/ contains the candidate signal-layer main flow and an explicit A6 P4-only ledger run. The driver loaded candidate modules from qa/a10/teaching_revision/C13/r1_20261014/candidate/, recorded their hashes and paths, and called p4_demo() directly. It did not call the combined P4/P5 runner or p5_demo(). The candidate and canonical builds had identical seed, listed panel, constructed returns, and industry labels; the composite scores were exactly equal over all 20,104 overlapping cells.

In the 96-month, 240-name constructed universe, listed support ranges from 192 to 228. The 60-name gate therefore does not bind in the main run; the boundary is tested separately. Both portfolios have 94 active months and 2 cash months (no_valid_quintile_directions in January and February 2014). The candidate industry-neutral signal targets have maximum active-row industry net 7.98e-17, net 1.39e-16, gross error 4.44e-15, and off-support weight 0. The candidate neutral mean gross is 2.0. Plain targets retain their intentionally non-neutral industry exposure.

The old A6 artifact reported 1,842 industry-neutral name-month weights outside the signal-date listed support; zeroing those weights removed up to 0.0529937383 gross in a row and left a maximum industry net of 0.0280748663. In the rebuilt old-code weight comparison, those off-support weights have a maximum absolute name weight of 0.0040404040; old mean gross was 2.11363478. The corrected target has 16,831 nonzero industry-neutral name-months versus 18,452 under the old rule, and no off-support weights. Plain weights and fee-ledger results are unchanged to floating-point precision.

The lower turnover and fees follow from the corrected weights; the changed constructed return is a protocol correction effect, not alpha evidence. The full old/new plain and industry-neutral fee table, including exact returns, fees, turnover, residuals, target samples, and per-name weight samples, is in engineering_01/old_new_comparison.json and engineering_01/old_new_sample_weights.csv. Historical A6 source SHA: b40546b35d7a4d85228246d9b0932cbcf3aa1d3a35fc0a0d517080fdfbd164de.

At signal time, industry neutrality is within floating-point error. The existing A6 execution policy cancels 24 known execution-date delisting targets without resizing, so corrected execution-adjusted targets have maximum industry/net/gross residual 0.0230051. At 10 bps, post-execution close weights show maximum industry net 0.0296263, net 0.0464074, and gross error 0.2865722 as prices drift. Their off-support weight remains zero. These execution outcomes are reported rather than fed back into signal formation.

Across all six P4 fee runs, P&L attribution residual divided by NAV is below 3e-16. The new ledger's maximum absolute cash identity residual is 2.33e-10 in account currency. Historical A6 serialized the P&L residual but not the cash-identity residual, so no historical cash-identity comparison is claimed.

The old A5 result is retained at qa/a5/claude/results/p04/p04_signal_layer_results.json (SHA-256 7e9ea0f9375cf8af57a29017b49c82eb41c461019452a878cf90b1e36de09ad4); the old A6 outputs are also preserved. This addendum does not replace either historical result or claim independent acceptance, real-market validation, or completion of Project 4.