🛠 IMPLEMENTATION · autonomous loop · P5_DONE · hardened vs #544/#556
Crypto Orthogonal Bar Features — Implementation Journal (batch 6: 3 shipped · 1 gated)
3 of 4 shipped as a stacked PR chain: #568 card 78 hoeffding Φ² ← #569 card 59 hvg horizon ← #570 card 60 vgclust meangap. Each: 16/16 stages, ≤1e-9 oracle, 0-conceded adversarial challenge, check-full green.
⛔ BLOCKING — card 50 (rogers_satchell_var_bar) awaits YOUR ruling:
ratify that the production formula is the window-MEAN of the per-bar Rogers-Satchell scalar
(eval verdict.md C3). Then set gates.window_mean_ratified=true in LOOP_LEDGER.json and
re-launch the loop — it resumes at card 50 automatically.
Operator tail: merge #568 → rebase+merge #569 → rebase+merge #570 →
deploy → sidecar restart (zero-overlap playbook) → backfill → population audit (3 new columns).
Full decision list: LOOP_LEDGER.json → morning_handoff; journal: iterations 01–13.
The autonomous loop that turned the batch-6 orthogonality-proven research formulas (re-scoped 8 → 4 by the ξ §B gate, PR #564) into production ClickHouse columns — 3 shipped as stacked PRs #568 ← #569 ← #570, 1 gated on an operator ruling — each proved bit-exact (≤1e-9), wired from a single manifest source of truth, and persistence-proven by a hermetic round-trip before the PR is opened. This batch is explicitly hardened so the batch-5 #556 silent-NULL incident cannot recur.
PHASE · P5_DONE (13 iterations) gate · worst-regime / worst-cell ∩ ξ §B 3 shipped · PRs #568 ← #569 ← #570 1 gated · card 50 awaits window-mean ruling loop stopped at the PRs (contract)
Read this first — what this loop is doing
main) — not one batch PR. Small blast radius, trivial per-feature revert, and a per-feature persistence gate that counters #556. Cost: every feature edits the same SSoT files and bumps the column-count tripwire, so the stack rebases linearly as earlier PRs merge (they queue, they don't fan out). Order: 78 (clean) → 59 → 60 (visibility-graph) → 50 (after the reduction ruling).
xi_worst<0.50 · WATCH residual · BAN >0.95 ∧ breadth ≥ 0.80) and banned the median rule as a promotion axis. Loop scope is now the intersection of the two declared worst-cell gates: cards 78 · 59 · 60 · 50 (ξ 0.078–0.312, all PASS). Cards 58 · 45 · 108 · 107 are ξ-WATCH held (0.609–0.771) — document + re-probe + operator ruling before scheduling. Audit: PR #563.
The candidates were promoted by two declared worst-cell gates: Spearman CH_FEATURE_BANLIST bands (PASS max|ρ|<0.85 · WATCH 0.85–0.95 · BAN ≥0.95) and Chatterjee ξ §B (PR #564) — never the median rule (banned). Full evidence: WORST_REGIME_CELL_ANALYSIS.md · live status: verdict.md · full playbook: IMPLEMENTATION_PLAYBOOK.md.
How each feature is implemented — the hardened flow
Green = loop-owned (autonomous) · amber = operator-owned (outside the loop) · rose = incident-prevention guard · ★ cyan = the #556 persistence gate.
The 16 per-feature stages (inside P1)
| Group | Stages | What it proves |
|---|---|---|
| Prove the maths | oracle · xcheck | Clean-room Rust == trusted reference bit-exactly (≤1e-9), cross-checked by 3 independent legs. Guards: FixedPoint quantize (drift==0 fail-closed), FOSS/closed-form cross-val, IEEE-stable substrate. |
| Build the kernel | kernel · wire · causality · determinism | Fast Rust kernel wired into every write surface from feature_manifest.toml (no hand-lists); reads past bars only; threaded == serial. |
| Store it | schema · contract · manifest | ClickHouse column + comment; 4-file lockstep contract; catalogue registration. |
| Verify it | parity · robust · ★ persistence round-trip + SSoT completeness | compute == stream == persisted-read-back; degenerate/oversized/scale witnesses; hermetic serialize→write→read-back proves the 8 columns are non-NULL and bit-equal — with no deploy. |
| Ship it | adr · mise · check-full · challenge · commit | Decision record, slash-commands, full gate (fmt+lint+test+deny), per-feature Attacker/Defender challenge, then commit — the loop halts at the open PR. |
Incident guardrails baked into this batch
| Incident | Root cause | Guard in this loop |
|---|---|---|
| #544 / #556 silent-NULL persistence | compute updated but the ~5 write surfaces each carried a hand-list of only the 4 old columns → 8 new columns serialized as NULL, 0% populated | Wire every surface from the manifest SSoT; hermetic persistence round-trip + manifest-completeness test before the PR |
| I1 order split-brain | implementation_order diverged from the ledger | Order derived from the manifest, not a parallel list; a stateless fire re-derives "next" |
| I2 rank-flip | libm ≠ glibc ln flipped ranks (Bartels 6.98e-4 > 1e-9) | Rank the IEEE-exact gross-return ratio, not ln |
| I3 quantization | round-half-away vs half-even at the FixedPoint boundary | Oracle pre-quantizes via the production FixedPoint path; drift==0 fail-closed |
| I4 bad oracle | numpy np.float64() repr bug in a generator | FOSS/closed-form cross-validation baked into the generator, fail-closed |
The features — final status (3 shipped · 1 gated · 4 ξ-WATCH held)
All clean-room Rust in bar_close_features.rs; per-feature gen_*_oracle.py; ≤1e-9; no runtime FOSS dependency (libraries are dev-only oracle generators). Final bar_* column names assigned by feature_manifest.toml.
"bar_close" has two senses. The family (group="bar_close") is all batch-6 features — that governs wiring/persistence. The Substrate column is what each formula actually consumes: only 59/60 (and held 58) are close-substrate; the rest are signs / ohlc_bar / ret_dur. The A/B/C delivery group is a third axis (shared blocker for PR sequencing) and crosses substrates.
| Card | Candidate | Substrate | worst-cell max|ρ| | ξ worst-cell | Group / notes | Status |
|---|---|---|---|---|---|---|
| 78 | bar_hoeffding_phi_squared_midreturn_duration | ret_dur | 0.738 | 0.078 | A · stack base · first TWO-substrate kernel (return-rank × duration-rank Φ²) | 16/16 · PR #568 |
| 59 | bar_hvg_forward_visibility_horizon_mean | close | 0.704 | 0.175 | B · first-passage horizon (plan corrected: needs NO graph — the VG helper moved to card 60) · closed-form/property oracle | 16/16 · PR #569 |
| 60 | bar_vg_time_directed_clustering_meangap | close | 0.844 | 0.171 | B · renamed _kld → _meangap in the PR · carries the shared HVG helper · ts2vg dev-oracle leg | 16/16 · PR #570 |
| 50 | rogers_satchell_var_bar | ohlc_bar | 0.800 | 0.312 | C · TTR (dev-oracle) · NOT started — the loop stopped here per contract | ⛔ gated · awaits window-mean ruling |
| 58 | hvg_mean_degree_excess | close | 0.764 | 0.609 | B · ξ-WATCH vs close (bull_2021) · re-probe + ruling required | held · ξ-WATCH |
| 45 | heikin_ashi_open_gap_sign | ohlc_bar | 0.805 | 0.659 | C · ξ-WATCH vs cox_stuart (current_2026) · re-probe + ruling required | held · ξ-WATCH |
| 108 | sequitur_grammar_compression_ratio_logret_sign | signs | 0.690 | 0.710 | A · ξ-WATCH vs high (bull_2021, breadth 4/7) · re-probe + ruling required | held · ξ-WATCH |
| 107 | etc_effort_to_compress_logret_sign | signs | 0.717 | 0.771 | A · ξ-WATCH vs high (bull_2021, breadth 6/7) · re-probe + ruling required | held · ξ-WATCH |
Per-candidate test evidence (the shipped three)
Every claim below is enforced by a committed, hermetic test (no ClickHouse at test time) that fails the build if violated — file names + SHA pins are in each PR. "Oracle residual" = the largest observed |Rust − Python| across all 19,602 windows (9,801 per symbol, BTCUSDT + ETHUSDT), against the sacred ≤1e-9 ceiling.
Card 78 — bar_hoeffding_phi_squared_midreturn_duration · PR #568
| Gate | Evidence | Result |
|---|---|---|
| ≤1e-9 oracle (S1) | 19,602 windows vs SHA-locked CSVs (BTC 2ec3992d… / ETH 68676123…); new self-contained (close, duration_us) fixtures (26bfea6f…/1627ec31…), FixedPoint drift = 0 | GREEN · residual ~1e-13 |
| Independent legs (S2) | In-generator exact-integer closed form (≤1e-12, every window) + in-Rust naive O(m³) counting leg + 8-test kernel battery | GREEN |
| Causality / determinism / scale / crash (S5·S6·S11) | Dedicated two-substrate lanes: truncation bit-parity, recompute bit-parity, dual-axis bit-exact pow-2 scale invariance, 12-case totality — both symbols | GREEN ×2 symbols |
| Persistence (#556 ★, S4·S8) | All surfaces from feature_manifest.toml; tripwires 85→86 (Arrow) / 90→91 (CORE_COLUMNS + py mirror); schema-contract 4-source equality; 91-column dead-letter Parquet round-trip (Rust→polars) | GREEN |
| Parity + restart (S10) | batch ≡ streaming ≡ JSON-checkpoint-restart on real aggTrades, non-vacuous; pre-durations checkpoint self-heal (field deleted from real JSON) + oversized-window re-convergence | GREEN |
| check-full (S14) + challenge (S15) | fmt+lint+test+deny green ×2; Attacker 2 MAJOR + 4 MINOR (stale compat counts; untested deploy-day restore path) — all fixed in-PR, Defender verified 6/6 | 0 conceded |
| Discovery finding | Finite-m grid Φ² = 1.0000631 at m=199 under perfect monotone dependence (textbook [0,1] is asymptotic) — documented, not clamped | documented |
Journal: iter 01 · 02 · 03 · 04 · 05 · ADR 2026-07-02-bar-hoeffding-phi-squared-midreturn-duration.md · gates: /hoeffding:check-full
Card 59 — bar_hvg_forward_visibility_horizon_mean · PR #569
| Gate | Evidence | Result |
|---|---|---|
| ≤1e-9 oracle (S1) | 19,602 windows vs SHA-locked CSVs (BTC aff5c3a0… / ETH 54acf3c7…); committed close fixtures reused; comparisons + one division only | BIT-EXACT · residual 0 |
| Independent legs (S2) | Dual-leg generator (naive scan vs monotonic stack, EXACT integer-list equality, every window) + in-Rust naive O(n²) leg bit-equal at n=5/23/61/200 + 4 exact anchors (increasing/constant → exactly 1.0, decreasing → NaN, hand-pinned [3,1,2,3] → 5/3) | GREEN |
| Causality / determinism (S5·S6) | The STANDARD 13-column walk drivers sweep the column over both symbols (close substrate — no dedicated lanes needed) | GREEN ×2 symbols |
| Persistence (#556 ★) | Tripwires 86→87 / 91→92; compat 66→67 & 13→14 bumped proactively (card-78 lesson); schema-contract; 92-column dead-letter round-trip; completeness 14==14==14 | GREEN |
| Parity + restart (S10) | batch ≡ streaming ≡ checkpoint-restart with non-vacuity guard | GREEN |
| check-full + challenge | Attacker: 84,124-sequence exhaustive differential proof (stack ≡ naive, 0 mismatches) + full clean-room oracle recompute (0/19,602 bit mismatches); 6 MINOR findings; Defender REJECTED one fix as incomplete (+4 remnants, +the N5 vacuous-gate catch) — all worked, final 61/61 with the arrow lane | 0 conceded |
| Plan correction | Card 59's evaluated kernel is a first-passage scan — builds NO graph; the shared VG helper moved to card 60 where it is used (recorded in both ADRs) | surfaced |
Journal: iter 06 · 07 · 08 · 09 · ADR 2026-07-02-bar-hvg-forward-visibility-horizon-mean.md · gates: /hvghorizon:check-full
Card 60 — bar_vg_time_directed_clustering_meangap · PR #570
| Gate | Evidence | Result |
|---|---|---|
| ≤1e-9 oracle (S1) | 19,602 windows vs SHA-locked CSVs (BTC e70f3c69… / ETH e3e1ee77…); QUADRUPLE-leg generator: left-scan HVG primary · ts2vg edge-set EXACT equality on every window · naive O(n³) subset (21/symbol) · 1e-12 time-reversal anchor everywhere | GREEN · residual ~1e-15 |
| Traps caught fail-closed (S1) | P1: a textbook monotonic-stack HVG admits phantom edges on equal-height runs ([2,1,1,2] class) — vetoed by the ts2vg leg (3 spurious edges, window 0); P2: the reversal sign-flip is exact only in ℝ (IEEE summation-order ULP) — anchors assert 1e-12, not bits | institutionalized as tests |
| Kernel battery (S2·S3) | In-Rust naive O(n³) leg bit-equal ×4 sizes + the P1 tie battery + P2 reversal property + symmetric-sawtooth anchor (exactly 0.0) + monotone → NaN + pow-2 scale bit-invariance | 5/5 |
| Persistence (#556 ★) | Walk 13→14; tripwires 87→88 / 92→93; compat 67→68 & 14→15; manifest meta + every narrative count in the same pass; schema-contract; 93-column round-trip; completeness 15==15==15 | GREEN (62/62 battery) |
| check-full + challenge | Attacker's 1 MAJOR was REAL: clippy-deny needless_range_loop in the new kernel under --all-targets --all-features — a lane the namespace gates never run; fixed + 4 MINOR worked; Defender verified all (bench lint rot proven pre-existing at HEAD) | 0 conceded |
| The rename | _kld → _meangap: a true D_KL needs an unpinned density estimator (forbidden); what was evaluated and ships is the signed mean-gap | recorded in ADR + manifest |
Journal: iter 10 · 11 · 12 · 13 · ADR 2026-07-02-bar-vg-time-directed-clustering-meangap.md · gates: /vgclust:check-full
Iteration log (every loop firing)
| # | Card | Advanced | Page |
|---|---|---|---|
| 01 | 78 | 0/16 → 1/16 · S1 oracle (new 2-col fixtures; dual-leg green 19,602/19,602) | Hoeffding Φ² oracle grounded |
| 02 | 78 | 1/16 → 3/16 · kernel + ≤1e-9 gate GREEN; finite-m range discovery | Rust kernel proven bit-exact |
| 03 | 78 | 3/16 → 4/16 · all #556 write surfaces; schema-contract flips green | Wired into every write surface |
| 04 | 78 | 4/16 → 13/16 · causality/determinism/parity/robustness + ADR + namespace | Certified end-to-end |
| 05 | 78 | 13/16 → 16/16 DONE · challenge 0-conceded · PR #568 | Card 78 shipped |
| 06 | 59 | 0/16 → 1/16 · S1 oracle (fixtures reused; VG-helper plan correction surfaced) | The "how long underwater?" ruler |
| 07 | 59 | 1/16 → 3/16 · BIT-EXACT gate (residual 0); standard walk 12→13 | Kernel matches to the last bit |
| 08 | 59 | 3/16 → 11/16 · wiring + certification in one pass (lessons compounding) | Wired & certified in one pass |
| 09 | 59 | 11/16 → 16/16 DONE · defender-forced remnant sweep · PR #569 | Card 59 shipped |
| 10 | 60 | 0/16 → 1/16 · S1 oracle (quadruple-leg; P1+P2 traps sprung pre-Rust) | The time-arrow detector begins |
| 11 | 60 | 1/16 → 3/16 · tie-safe kernel; ≤1e-9 gate GREEN; walk 13→14 | Both booby traps defused |
| 12 | 60 | 3/16 → 11/16 · wiring with all four accumulated lessons applied up front | Cleanest wiring pass yet |
| 13 | 60 | 11/16 → 16/16 DONE · PR #570 · card-50 gate check → morning-handoff · P5_DONE | Loop ends per contract |