Prospective pooled-panel calibration protocol V17 — review draft

Proposed author: Peilun Cai, Axiocairn Research. ORCID: https://orcid.org/0009-0007-5364-6883. Human review and signature pending. Not registered, approved, or permission to collect retained data. This consolidated proposal incorporates the CEO’s 2026-10-30 01:30 missing-bar and first-value decision. It retains V15’s pooled estimator, cohort and dimensions; historical paper results and original code remain unchanged.

Question, target and interpretation

Given one future thirty-stock panel, does the predeclared pooled fixed-fingerprint procedure reject more often than 0.10 under the specified common-feature-shift/random-window algorithm? The primary quantity q is the probability of an algorithmic rejection conditional on the complete panel. Independent randomization streams supply the Monte Carlo replication unit. Stock-window rows are not treated as independent market trials. This is a prospective extension of the calibration question, with a different pooled sampling unit from the original per-stock study. It is not a direct numerical replication of 0.314, a market-wide type-I-error estimate, evidence of predictability, or proof of a §15 breakthrough.

The historical 0.314 concerns the final fixed rule and omits the selection step. This protocol reports that fixed-rule arm and the entire declared 24-rule search separately. Feature definitions are fixed before lock and are not searched. Unknown historical feature/model selection is not reconstructed. A joint feature shift is the chosen conditional audit; its validity as a general market null is not assumed. Raw-price regeneration is a different experiment and is not substituted.

Cohort, observation period and admissibility

Use exactly the thirty distinct issuers and their order in ../forward_panel_v16/UNIVERSE.json: pre-data SEC/NYSE identity intersection with the original 164 identifiers, selected by CIK then security ID. Keep the full selection/exclusion ledger. No replacement, outcome-based selection, inferred ticker change or silent issuer remapping. The cohort is not representative of the market.

Let s[0] be the first XNYS session strictly after the UTC date of the final protocol’s Bitcoin-confirmed OTS attestation. Use exactly 1,021 such sessions for all price, feature and benchmark inputs. Calendar submissions without Bitcoin confirmation do not start the clock. The first feature basis is s[60]; subsequent bases are s[60+20j], j=0,…,47, with endpoints s[80+20j]. Twenty daily return increments do not overlap. A shared endpoint price does not imply independent market returns. Require the first basis’s quarter to start after the confirmation date; the current planner fails closed otherwise. All warmup is post-attestation. The unused original ATR-rank output does not enter any rule; the four selected features retain their original definitions and initializations.

The earliest complete-panel date is conditional on actual confirmation, official session calendars, admissibility and successful execution. Under a hypothetical 2026-10-30 confirmation, the prior package projection is 2030-11-22. Official calendars have only been checked through 2028; that later date is not a delivery promise. W=48 is a fixed all-shift-support design choice, not a claim that 1,440 rows provide 1,440 independent observations or 90% market power. No sample expansion or reduction after outcomes; an ineligible panel ends inconclusive.

../forward_panel_v16/admission.py requires thirty unique issuers in frozen order and complete, finite, aligned benchmark and OHLCV first-value series, with original close preserved as RawClose. Scheduled exchange closures are excluded from the session grid. A day without a returned transaction bar is a missing bar: exclude it and ledger the security/session without attributing the absence to a halt or vendor gap. No external halt proof or Boolean is required. A legal complete bar is retained regardless of any optional halt label. Missing, malformed or invalid first values are ledgered; no forward-fill, backfill, source fallback, replacement, date compression or sample reduction. Any required exclusion, including a benchmark gap, makes the complete-panel study inconclusive. This existing conservative rule may prevent completion over the long collection period.

The independent stock custodian requests each day’s adjusted OHLCV and original close at 21:00 America/New_York, preserving the first response or failure by 21:15, request parameters, each transport time, exact bytes and hashes. This proposed collection window must be reviewed before lock. The offline adapter ../forward_panel_v16/first_snapshot.py pins the first pair per symbol/session, including empty or invalid replies, and retains later replies as audit observations only. Its endpoint-specific mapping is in ../forward_panel_v16/SOURCE_CONTRACT.md; both the adjusted and unadjusted endpoint use an adjClose field, with distinct meaning. A first pair outside the window is unusable. Supplied receipt times and hashes are not authenticated OTS or proof of source entitlement.

The calculation uses the first adjusted price and volume received; RawClose is preserved for audit. Subsequent splits or restatements never modify those registered first values. Split events are saved in a separate event table with acquisition times and response hashes. This point-in-time snapshot series may contain corporate-action discontinuities and is not a consistently restated total-return history. Report that limitation and all events, including adverse effects on measured features and returns; do not repair the historical series after seeing results.

The custodian archives all original responses and a collection-package OTS. The author neither collects the retained panel nor reads the existing synthetic judge’s seal. CEO’s 01:30 decision removes the external halt-coverage prerequisite and treats the founder’s derived-publication rights choice as separate from protocol lock. Only derived statistics may be considered for publication; no raw price redistribution is authorized. Actual future data access, exchange calendars and receipt authentication remain the custodian’s execution responsibilities.

Exact estimator and eligibility

The executable definition is ../forward_panel_v16/panel_estimator.py, using the verbatim feature extraction in ../forward_panel_v16/original_features.py. The four features are gmma_bear, below_avwap, rs60 and volz. For each issuer, the unshifted final basis chooses the three fixed-rule directions, with volume cutoff 1.0. Exactly zero final-basis rs60 makes (rs60 < 0) false, selecting the positive-event branch; a zero event itself satisfies neither sign. Missing momentum or volume never qualifies an event; invalid final-basis features stop before analysis.

For a common circular offset d, evaluate features at (basis−d) mod 1021 for every issuer, keeping its unshifted returns. This shifts the same feature-vector rows across the panel and preserves synchronous cross-stock alignment. The observed pooled mean weights each selected stock-window equally. The primary mask uses each issuer’s frozen rule. Search constructs all eight direction combinations times volume thresholds 0.5,1.0,1.5, in the exact RULES order. All masks are rebuilt on every null replicate; no search early stop after success.

Require at least thirty selected events across the pooled panel for a rule to be tested. Before any estimator call, calculate feature-only event counts at every allowed offset in the union 63…958. The fixed arm must meet thirty events and at least one searched rule must be testable at every offset. If either condition fails, the planned panel is inconclusive before randomization; record all counts and bad offsets. This prevents stochastic event-count attrition from aborting a qualified run. It conditions on panel features and is not a proof that eligibility is independent of returns.

For a tested rule, let m_s be its selected count in issuer s. Each of 400 placebo draws selects exactly m_s of that issuer’s 48 windows without replacement; zero-event issuers contribute zero. Draws are independent across issuer/draw conditional on the PRNG contract. A shared permutation-prefix plan across the 25 masks is an explicit variance/cost choice: it induces dependence across rules but does not skip their search or change their marginal random-window distributions. Compute the original strict-less-than percentile, mean(placebo_mean < selected_mean), and reject at percentile ≤0.05 or ≥0.95. A tested zero-return tied fixture consequently rejects; this known defect is deliberately retained as an adverse result, not repaired in the historical method.

Rules below the event floor are untested, with count and reason, never non-rejections. The any-rule indicator is true if any tested searched rule rejects. Preflight guarantees at least one testable searched rule per offset; invalid numeric calculations or violation of that invariant produce an inconclusive run. Every rule row, including untested rows, stays in the ledger.

Randomization, ordering and decision

Use NumPy 2.3.5, seed 20261029, 2,000 replicates for each arm. For horizon h and replicate r, the shift generator is SeedSequence([seed,15,0,h,r]); the independent placebo generator is [seed,15,1,h,r]. Draw d uniformly with replacement from [h,1021−h], inclusive. Horizon supports have 518,896,770,14 offsets for h=252,63,126,504 respectively. These are finite conditional shifts, not fresh market histories.

Finish all 2,000 fixed-rule h=252 replicates before any search work. Then run complete searched h=252,63,126,504 arms, 2,000 replicates each, retaining a fixed-rule value in each. The h=252 search replay must reproduce every primary fixed-rule indicator using identical streams. Every searched replicate includes all 24 rules. Overall completion requires every arm.

Primary test: H0 q≤0.10 versus q>0.10. With Q = primary rejections / 2000, reject only if Q−sqrt(log(20)/(4000)) > 0.10, equivalently at least 255 primary rejections. Otherwise report no rejection, not equivalence. The lower bound is one-sided 95%; the pair of displayed primary bounds is not a 95% two-sided interval. Secondary fixed/any-rule summaries use .05/16 per tail across two statistics, four horizons and two tails. This secondary family is separate from the primary test; the entire combined display is not one 95% simultaneous family. Secondary nominal-threshold flags are diagnostics and cannot become confirmatory decisions or a claim of family-wise market calibration.

../forward_panel_v16/CONDITIONAL_POWER.json and its derivation bind the planning calculation. At q=.15 the exact independent-Bernoulli power is 0.99816644, with a Hoeffding lower bound 0.87114605; at q=.20 the lower bound is 0.9999999993155593. At the null boundary .10, the exact Bernoulli size for this conservative critical count is 0.0000434585. These quantify independent algorithmic draws conditional on a panel, assuming their implementation/PRNG contract, not market sampling. V14’s beta-binomial tables remain a separate sensitivity exercise and do not set this estimator’s effective market sample size.

Software, resource budget and invalid runs

The exact candidate files are in ../forward_panel_v16/MANIFEST.json (unchanged code), with V17 review additions bound by this folder’s MANIFEST.json; the measured core workload is separately bound by ../forward_panel_v15/PRELAUNCH_MANIFEST.json. The observed Python 3.11.14 environment and package versions/installed record hashes are in ../forward_panel_v16/ENVIRONMENT.json and ../forward_panel_v16/requirements.lock. The V16 test dependencies are separately pinned in ../forward_panel_v16/requirements-test.lock and ../forward_panel_v16/TEST_ENVIRONMENT.json. The accompanying V17 author reproduction adds exact binary-wheel hashes in WHEELS.json and requirements-hashed.lock, with offline installation into a fresh virtual environment. All 35 combined admission, first-snapshot, estimator and execution-control tests passed. CLEAN_ENVIRONMENT.json and CLEAN_TESTS.log bind the observed result. This shares the host and base Python installation, so it is not independent clean-OS reproduction. REPRODUCE.sh is the separate reviewer replay command; its execution is not presumed. No lock/OTS until the final package and environment have passed the existing different-family protocol review. The prior full workload is reused only for its unchanged estimator, fixtures and execution controls; it does not benchmark the new receipt adapter.

The complete deterministic stress workload finished with 2,000 primary and four times 2,000 searched replicates: 10,000 replicate records and 202,000 rule evaluations. It used 115.0877 seconds wall, 96.1494 seconds CPU and 341,655,552 bytes peak RSS. All fifteen prelaunch files matched afterward; supervisor return code was zero. ../forward_panel_v15/FULL_RUN_READBACK.json independently reconciles the ledger, phase order, counts, duplicates and 2,000 primary replays. The feature builder was timed separately on synthetic OHLCV; stress features then deliberately made all 24 rules testable. This is full software workload evidence, not empirical size/power or data-adapter acceptance.

Proposed execution budget: 1,000 seconds inner wall, 1,020 seconds outer wall, 900/930 seconds CPU soft/hard, 512 MiB peak RSS checked before/after calls, and 512 MiB maximum single ledger file. Darwin rejected RLIMIT_AS in two minimal probes; the memory budget is not a hard in-call OS cap. Fixed dimensions and streamed output bound the implemented working arrays; any future data loader needs separate review. The outer supervisor handles interruption, kills only its owned process group and reaps it; the child stops between calls if its supervisor disappears. The manifest hash is bound at supervisor launch, child start and completion. A timeout, error, source drift, missing supervisor or memory-budget crossing is inconclusive, with preserved attempts. No same-version retry after real study outcomes; retain the failed run and seek a reviewed deviation/new registration if needed.

All pre-data timing failures are retained: unsupported RLIMIT_AS, a stale-pin launch following an incorrect Python environment inventory, and the 256 MiB RSS stop at primary replicate 52. A bounded 500-replicate memory diagnostic preceded the proposed 512 MiB budget. No scientific sample, horizon, replica count or adverse outcome was reduced to make timing pass.

Controls, publication and remaining prerequisites

The prior twenty-two-test result covers the V15 estimator, admission and execution controls. The focused V16 tests in ../forward_panel_v16/TEST_RESULTS.json supersede the admission tests and exercise receipt mapping, chronological receipt handling, per-session exclusions, split/restatement immutability and unchanged feature calculation; property tests additionally vary first and revised values. Unchanged estimator/control evidence is reused by exact hash. Eight full-search deterministic controls across the four horizons retain the constant-zero-return rejection defect and engineered positive-signal direction. The positive fixture is constructed to align with its known shift; it is not independent efficacy validation. No comparative market type-I error, statistical operating-characteristics experiment or strongest-baseline superiority is claimed. Such claims require the separately registered matched simulation and independent scientific assessment.

Report rejection, no rejection and inconclusive outcomes, every attempted run, untested-rule count, exclusion and deviation. Do not change thresholds, dates or universe after results. V1 remains exploratory, with the post-result timestamp disclosed, and does not wait for v2 outcomes. Its publication still requires the independently approved final prospective protocol plus OTS, compliance with the bound founder risk-acceptance scope, complete trial accounting, human author review, publication QA and founder approval.

The author has supplied executable answers for eligibility, full-search semantics and measured workload. Protocol lock now awaits different-family adjudication of this exact pooled conditional design, source contract and final software/environment. No external proof of halt/no-halt status is required. The founder has selected risk acceptance for inline derived statistics from commercial market data (not redistributed). This is not contractual provider permission. Human author verification and final publication approval remain pending; the rights choice does not block protocol lock after independent review. Do not describe this review draft as registered or sign for the author.