Daru FinanceDaru Finance
Back to topology roadmap

Changelog

Topology roadmap, version history

Every time the live topology page is rewritten, the previous version is archived here. The current live page sits at the top; click any earlier entry to open the snapshot exactly as it appeared on that date.

  1. 2026-05-14

    live

    Phase II sprint mid-update on top of the same-day Phase I closure. Two new substantive deliverables. (1) Sheaf-Laplacian aggregation lands a partial second-stage refinement: a cohomology-gated sheaf-Tikhonov portfolio Holm-rejects the previous deployment-scale headline configuration on the ALL grouping (+0.20 ΔSortino), on the EMA family (+0.53), and on the MACD family (+0.65), and Holm-dominates three other baselines (the strict-causal H4 portfolio +1.29, naive Sharpe-rank top-20 +0.92, equal-weight reliability +1.23). It misses the SMA family (−0.10, not Holm-rejected), so the pre-registered "must clear both moving-average families" bar is half-met, with a clean Stage 2 follow-up path (full λ-cross-validation, threshold tuning, per-member theory stalk) to potentially close the SMA gap. The same workpackage also confirmed empirically that the first cohomology obstruction is decoupled from per-cell decision quality (Spearman ρ between obstruction and per-cell decision-quality cost = −0.035, threshold ρ ≥ 0.40 not approached, a clean negative that adds an empirical leg to the prediction-vs-decision theorem) and that restricting the partition construction to the most recent eight walk-forward windows reduces the obstruction with a large, Holm-rejected effect size on both selection metrics tested. (2) The cost-capacity stability theorem now ships with two complementary measures, not one: a discrete-scenario operationalization on the five-scenario perturbation panel (baseline / fee / slippage / entry-drift / entry+indicator) where 16.4% of the corpus survives the entire envelope and the per-strategy cost class is temporally stable; and a continuous-cost operationalization on a nine-point basis-point grid {0, 2, 5, 8, 12, 20, 35, 60, 100} where the never-flips tail shrinks to 0.34% and a Holm-rejected temporal drift appears that the discrete grid missed. The two measures have a Spearman rank correlation of 0.0491 with each other, essentially uncorrelated, and capture genuinely complementary axes of cost-fragility, not coarsenings of the same scalar. Same per-family / per-cohort directional patterns survive across both. The prediction-vs-decision theorem therefore now stands on six independent empirical legs across multiple work packages.

    • Sheaf-Laplacian aggregation, partial second-stage portfolio refinement on top of the first-headline win. A first-cohomology-gated sheaf-Tikhonov-regularised portfolio Holm-rejects the previous deployment-scale headline configuration on the ALL grouping (+0.20 ΔSortino), on the EMA family (+0.53), and on the MACD family (+0.65); it misses the SMA family (−0.10, not Holm-rejected). The pre-registered "must clear both moving-average families" bar is therefore half-met, with a clean Stage 2 follow-up path (full λ-cross-validation, threshold tuning, per-member theory stalk) to potentially close the SMA gap. Against three further baselines the variant Holm-dominates: the strict-causal H4 portfolio (+1.29), the naive in-sample-Sharpe-rank top-20 (+0.92), and equal-weight reliability (+1.23). Description in the existing Phase II findings card and a new entry in the live page's "what changed in this update" lead.
    • Sheaf-cohomology obstruction is decoupled from per-cell decision quality. Spearman ρ between the first-cohomology obstruction and per-cell decision-quality cost = −0.035; the pre-registered threshold ρ ≥ 0.40 is not approached. Clean negative, the obstruction describes the cover structure of the partition, not the decision-side quality of the cells over which the partition is built. This is an empirical leg of the prediction-vs-decision theorem at the methodology level, not just at the deployable-portfolio level.
    • Restricting the partition construction to the most recent eight walk-forward windows reduces the cohomology obstruction on both selection metrics tested, with large, Holm-rejected effect sizes. Confirms again that the obstruction tracks the structure of the cover and not the quality of the targets over which it is built, and gives the deployable variant a principled recency window for the next stage of refinement.
    • Cost-capacity stability theorem, first operationalization (discrete five-scenario perturbation panel: baseline / fee / slippage / entry-drift / entry+indicator). Temporally stable per-strategy cost class (Holm-not-rejected on the per-window drift of the cost class), 16.4% of the corpus survives the entire 5-scenario envelope (the deployable-cost tail). Cross-cohort heterogeneity is real: L1-majors strategies are the least cost-robust; alt-L1s the most, both Holm-rejected on the 5-cohort consistency contests.
    • Cost-capacity stability theorem, second operationalization (continuous-cost basis-point grid {0, 2, 5, 8, 12, 20, 35, 60, 100}). At fine resolution a Holm-rejected temporal drift appears (mean cost-class shift = −0.0213 cost-class-units per window, p = 0.0008) that the discrete grid missed; the never-flips tail shrinks dramatically (16.4% → 0.34%), the discrete "deployable tail" was hiding the actual cost-failure point. Same per-family and per-cohort directional patterns survive (MACD-family strategies most cost-fragile; EMA most robust; L1 less robust than alt-L1) but the continuous-cost measure resolves 1.6× wider heterogeneity gaps.
    • Spearman correlation between the discrete-scenario and continuous-cost measures of cost-capacity = 0.0491. Essentially uncorrelated. The two measures capture genuinely complementary axes of cost-fragility, not coarsenings of the same scalar, and answer different deployment questions: the discrete measure says "does the strategy survive the realistic execution-friction panel?"; the continuous measure says "at what absolute basis-point cost does it fail?". A high-turnover strategy can survive the fee/slippage panel but flip at a low basis-point cost; a low-turnover strategy can be vulnerable to entry-drift but robust to high absolute cost. Reported as two complementary measures, not a single scalar.
    • The prediction-vs-decision theorem now stands on six independent empirical legs across multiple work packages: the four refinements that closed the partition (forecasting fails or shrinks; the grain is covariate-level; robustness is necessary but not sufficient; an added degree of freedom must be predictive of the outcome); plus today's two new legs (the sheaf-cohomology obstruction is decoupled from per-cell decision quality, and the cost-capacity invariant is temporally stable and carries structural cross-cohort heterogeneity, both indicating that the relevant invariants capture structural facts, not decision-side ones). Stated as a cumulative strengthening of the central methodology claim in the partition finding and the central-finding banner.
    • EN + pt-BR. Three new findings cards (sheaf-Laplacian Stage 1 refinement; cost-capacity stability theorem with two complementary measures; the six-legs framing of the partition). Mark-2 roadmap caption updated to reflect the second-stage refinement and the cost-capacity result. The cost-capacity Mark-2 phase moves from planned to in-progress with two results landed. The dek and the central-finding banner read the new six-legs framing.
    • The prior live entry, the 2026-05-14 Phase I closure (H4 corpus-scale PASS, H5 FAIL, H20 FAIL, failure-mode-diversity FAIL, the first Mark-2 sheaf-Laplacian headline bar), is archived as v12-2026-05-14-phase-i-closure (resolved by the dynamic archive route).
    Open the live page

  2. 2026-05-14

    archived

    Phase I closure: H4 lands as a corpus-scale PASS at +0.61 ΔSortino over Hierarchical Risk Parity (95% CI [+0.39, +0.82], Holm-significant), concentrated on the L1-majors cohort and the volatility-and-oscillator indicator families, and statistically indistinguishable on Sortino from a simple top-k-by-reliability equal-weight baseline (so the topology contribution is coverage and audit-traceability, not raw return). H5 cross-asset universality FAIL, behaviour-space topology does not pool across structurally-similar crypto cohorts. H20 portfolio-coverage-health alarm FAIL (lags the classic drawdown trigger). Failure-mode-diversity portfolio constraint FAIL, but smaller in magnitude than the first measurement indicated. Mark-2's sheaf-Laplacian aggregation has cleared its first pre-registered headline bar, the first deployable-portfolio Holm-rejected win against a naive in-sample-Sharpe-ranked baseline the program has produced, with the largest margins on the moving-average indicator families. Master methodology finding across the audit chain: strict-causality is a deflator of leak-amplified side-features, not a destroyer of substantive findings, every qualitative claim survived; every quantitative anchor needed restating.

    • H4 deployment-scale verdict PASS at corpus scale: ΔSortino vs Hierarchical Risk Parity = +0.61 (95% CI [+0.39, +0.82], Holm-significant). The edge is heterogeneous: L1-majors carries the corpus number (+1.20 vs HRP, Holm-rejected); outside L1-majors no §5 cohort produces a Holm-significant edge. By indicator family, ATR / MACD / Stochastic %K each Holm-rejected positive (+1.3 to +2.1 over HRP); EMA is Holm-rejected negative (−0.6); RSI / SMA / PPO are flat. Against a simple top-k-by-reliability equal-weight baseline the Mapper-cluster step is statistically indistinguishable on Sortino (paired Δ = −0.06, Holm-p = 0.37), so the topology contribution is coverage and audit-traceability, not raw return. Against the naive in-sample-Sharpe-ranked top-k baseline the topology Stack still loses on Sortino. The earlier +0.96 figure (reported on 2026-05-12 from the full-corpus re-scope, walked back on 2026-05-13 pending a stricter window-causal cluster-nerve audit) sat in the gap between this final number and the pre-audit measurement, the audit confirmed the direction of the headline and restated the magnitude. The bake-off bar chart is re-rendered with the corpus-scale numbers; the dek, scorecard, Phase-06 detail, and findings card all read the same headline.
    • H5 cross-asset universality verdict FAIL: bottleneck within-cohort / cross-cohort ratio = 1.018 (p = 0.59); Gromov–Wasserstein = 0.96 (p = 0.24). Both well above the α = 0.01 falsification rule. Behaviour-space topology does not pool across structurally-similar crypto cohorts, each asset needs its own selector. Paper D rescopes to no-universality. Phase I is now closed; every pre-registered hypothesis has landed a verdict.
    • H20 portfolio-coverage-health early-warning FAIL: median lead of the alarm against the classic drawdown trigger is −1.0 windows across 70 co-firing units (the alarm lags rather than leads); 0 of 89 cohorts produce a Holm-significant positive-lead binomial. Ships as a descriptive coverage diagnostic only, there is no actionable early-warning signal here. Third coincident-to-lagging "something is drifting" topology signal the pilot has tested (alongside H19 and the per-window aggregation-risk score), reinforcing the central methodology finding.
    • Failure-mode-diversity portfolio constraint FAIL, smaller magnitude than the first measurement indicated. The Sortino cost is real (ΔSortino = −0.09 corpus-wide, 95% CI [−0.16, −0.02], Holm-rejected, driven by the moving-average-family cohorts) but the originally-reported ~5% worst-decile drawdown deepening does not hold under careful measurement (the strict effect is −2.9% with the 95% CI spanning zero). The failure-mode labels themselves remain useful as a descriptive portfolio audit; the hard-constraint version stays off by default.
    • Mark-2 sheaf-Laplacian aggregation, first headline bar cleared. The partition-based portfolio (interpolating between rank-by-reliability and the cluster-aware diversifier) beats a naive in-sample-Sharpe-ranked top-20 baseline by +0.72 ΔSortino on the corpus (Holm-rejected), with the largest margins on the moving-average indicator families, Holm-rejected on both EMA (+0.59) and SMA (+1.15); 6 of 7 families are individually significant. This is the first deployable-portfolio Holm-rejected win against that naive baseline the program has produced. Described in the article "The partition does the work".
    • Methodology audit closure: the chain of audits applied to the cluster-nerve construction has finished. Every qualitative mark-1 claim survived the stricter window-causal discipline on its load-bearing axis; every quantitative anchor that depended on the earlier substrate needed restating on a leak-amplified side-feature (H4 magnitude moved 0.96 → 0.61, ~37% reduction; the failure-mode-diversity drawdown-deepening half halved and lost significance; H20's high-FPR side collapsed). The deflator-not-destroyer pattern is the canonical methodology lesson of the audit chain, and the strict-causal cluster-nerve construction is now the load-bearing substrate for the Mark-2 sheaf-Laplacian work and everything downstream.
    • EN + pt-BR. Dek, roadmap caption, scorecard, Phase-06 detail, findings cards (H4, H20, failure-mode-diversity, the central partition), the bake-off-bars figure caption, the deployment-stack decision-flow figure, the H4 equity curves caption, the Mark-2 section, and the in-page "what changed in this update" all updated to the 2026-05-14 closure state. The methodology audit scorecard section is removed from the live page; its historical record lives here in the changelog.
    • Last-updated label bumped to 2026-05-14. The prior live entry, the 2026-05-13 H4 walk-back pending the cluster-nerve audit, is archived as v11-2026-05-13-h4-walkback (resolved by the dynamic archive route).
    Open this archived version

  3. 2026-05-13

    archived

    H4 deployment-scale headline walked back pending a stricter window-causal re-audit on the cluster-nerve construction itself. The previous +0.96-Sortino-over-Hierarchical-Risk-Parity corpus figure (reported on 2026-05-12 from the full-corpus re-scope) is provisional and will be restated, or retracted if it doesn't hold, when the re-run lands. The H20 portfolio-coverage-health monitor and the failure-mode-diversity portfolio constraint sat on the same cluster-nerve substrate as H4, so their FAIL verdicts are part of the same re-audit and are also held provisional. Everything else on the page (H17 PASS, H14 MARGINAL, H3 FAIL, H1 PASS, H2 falsified, and the H5/H10–H19 verdicts) is leak-free by construction and stands as reported.

    • H4 deployment-scale verdict moved from PASS-with-caveat to UNDER RE-AUDIT. The earlier lookahead-feature retirement had left the H4 selector's lens itself prefix-causal, but the cohort-level cluster-nerve construction used a static cohort-wide configuration that didn't restrict to prefix windows, a residual non-strictly-causal surface. A sampling-class result should be tested under the stricter discipline before its corpus headline number is reported as a verdict, so the deployment-scale figures (≈ +0.96 ΔSortino over hierarchical risk parity, Holm-significant, ≈ 61% of cells won, plus the deepest-drawdowns and the moving-average-zero pattern) are provisional and will be restated when the re-run lands. The corresponding figures (the bake-off bars, the equity curves, the deployment-stack decision-flow) are kept on the page but explicitly tagged as the provisional first-pass renderings.
    • H20 portfolio-coverage-health early-warning and the failure-mode-diversity portfolio constraint moved from FAIL to UNDER RE-AUDIT. Both were tested on the same cohort-level cluster-nerve substrate as H4, so their FAIL verdicts are part of the same re-audit and are provisional pending the re-run. Failure-mode labels themselves are leak-free by construction (they sit on the drawdown-shape / robustness-ratio / regime-spread / trade-microstructure fingerprint columns, not on any target regression).
    • The four-layer prediction-vs-decision-partition section softened the "three for three" coincident-to-lagging line to two clean cases (per-strategy drift H19; per-window aggregation-risk score) plus one under-audit case (H20). Everything else in the partition copy stands as reported.
    • EN + pt-BR. Dek, roadmap caption, scorecard cards, Phase 06 detail, findings cards (H4, H20, failure-mode-diversity, the four-layer partition, the bake-off-bars figure caption, the equity-curves figure caption), the deployment-stack decision-flow figure caption, and the in-page "what changed in this update" lead all updated to lead with the 2026-05-13 walk-back. The 2026-05-12 entries (the H4 full-corpus re-scope, the H20 / failure-mode-diversity findings additions, the codebase refresh, the scorecard rework) are preserved as historical record beneath the new entry.
    • Last-updated label bumped to 2026-05-13. The prior live entry, the scorecard / percentage-complete rework, is archived as v10-2026-05-12-scorecard-rework (resolved by the dynamic archive route).
    Open this archived version

  4. 2026-05-12

    archived

    Page lead reworked. The original "central applied claim (as pre-registered)" paragraph is removed and replaced by a clean "what we did · what worked · what didn't" scorecard, a three-column grid (worked / partial / didn't) over every hypothesis and extension, each with a verdict badge and a one-line note, headed by the prediction-vs-decision central-finding banner, which now leads the page directly under the title. A percentage-complete headline (16 of 18 roadmap phases done = ~89%, with a progress bar and a one-line breakdown of what's still in flight and ahead) sits below the scorecard.

    • Removed the "central applied claim (as originally pre-registered)" block. The page now leads straight into a "what we did · what worked · what didn't" scorecard: three columns, what worked (H1 PASS, H4 PASS-with-caveat, H17 PASS, the Topology Deployment Stack assembled), partial-but-shippable (H12, H13, H14, H16, all marginal), and what didn't (H3 falsified, the make-or-break test, H2 falsified, H15 / H18 / H19 / H20 fail, the failure-mode-diversity constraint fail, H10 / H11 deprioritised), each item a label + a tone-coloured verdict badge + a short note, with the prediction-vs-decision central-finding stated in a banner above. EN + pt-BR.
    • Added a percentage-complete headline below the scorecard: a large "~89% complete" figure, a progress bar (16 of 18 roadmap phases done; the one in-flight phase shown in amber), and a one-line note, the empirical work is essentially finished; H5 (cross-asset universality) is the test still running; synthesis, the papers, the monograph and the public package are still ahead. The figure is computed live from the roadmap phase statuses.
    • The prior live entry, the codebase / software-package breakdown refresh, is archived as v9-2026-05-12-codebase-refresh (resolved by the dynamic archive route).
    Open this archived version

  5. 2026-05-12

    archived

    Codebase / software-package breakdown brought current. The module breakdown and line counts now reflect the code added through the latest phases, the ten-command Topology Deployment Stack CLI plus its wiring layer; the new descriptive/audit primitives (the per-strategy drift profile, the candidate-pool shortlist, the portfolio-coverage-health metrics, the failure-mode-coverage selector, each carrying the FAIL verdict that demoted it from a trigger to an audit aid); the per-phase driver scripts for the H12–H20 extensions, the failure-mode-diversity constraint test, and the full-corpus leak-free deployment-scale run; and the figure-render scripts behind the real-analysis renders. Roughly 70k lines / 229 files across Rust and Python. Plus a small set of copy edits.

    • Codebase panel refreshed (EN + pt-BR): the as-of date, the total line count (~70k) and file count (~229), the per-language split, and the module briefs are brought current, the CLI-apps group now describes the ten-command Deployment Stack (map, select, portfolio, audit, monitor, refit, aggregation-check, plateau-audit, universality-check, live-adapter) and the descriptive/audit primitives alongside it (each with its FAIL caveat); the phase-driver-scripts group now covers the H12–H20 extension tests, the failure-mode-diversity constraint test, and the full-corpus deployment-scale run; the figure-render group now covers the renders driven by real run-output parquets (the behaviour-manifold, the Mapper graph, the H4 equity curves, the H17 persistence trajectory). The Rust core is unchanged.
    • A few copy edits across the page.
    • The prior live entry, the H20 portfolio-coverage-health + failure-mode-diversity findings additions, is archived as v8-2026-05-12-h20-fmd-findings (resolved by the dynamic archive route).
    Open this archived version

  6. 2026-05-12

    archived

    Two new results added to the findings list. H20, the portfolio-coverage-health early-warning: FAIL, the coverage-health alarm fires at the drawdown, not before it, so it ships as a descriptive diagnostic only; it is the third topology-derived “something is drifting” signal that co-occurs with the loss but doesn't lead it (alongside the per-strategy drift monitor and the per-window aggregation-risk score). The failure-mode-diversity portfolio constraint: FAIL, clustering strategies into six failure modes and forcing the builder to span them costs more reliability than it buys (the worst-decile drawdown deepens ~5 % and risk-adjusted return drops ~0.17 Sortino); the failure-mode clustering is useful as a descriptive portfolio audit, not as a hard construction constraint, the fourth “another coverage axis doesn't pay for itself” result. Both reinforce the central finding (topology is for structure and decisions, not forecasting), a clause each in the four-layer-partition copy. No figure changes, the text mentions cover it.

    • H20, portfolio-coverage-health early-warning: FAIL. Does the portfolio's own coverage of behavior-space drift (coverage entropy collapsing, the picks crowding into a narrow region, the manifold concentrating) far enough ahead of a drawdown to act on it? No, the alarm fires at the same time as, or slightly after, the drawdown, not before. Ships as a descriptive coverage diagnostic only, not an early-warning trigger. It is the third time the same pattern shows up, a topology-derived “something is drifting” signal that co-occurs with the loss but doesn't lead it: the per-strategy behavior drift (H19), the per-window aggregation-risk score, and now the portfolio-coverage drift. Added as its own findings card (sibling to the H19 card) and a clause in the layer-1 / layer-4 partition copy.
    • Failure-mode-diversity portfolio constraint: FAIL. Strategies were clustered into six failure modes (slow-bleed, sudden-crash, fee-sensitive, regime-flip, thin-trade artifact, tail-exposed) and the builder was told to pick a set that spans those modes rather than always taking the most-reliable strategy in each region. Result: the worst-decile drawdown deepens ~5 % and risk-adjusted return drops ~0.17 Sortino, forcing failure-mode balance means picking a less-reliable (less out-of-sample-robust) strategy, which costs more than the diversity buys. The failure-mode clustering itself is real and useful as a descriptive portfolio audit (“which failure modes do my picks span; is any one over-represented?”), but not as a hard portfolio-construction constraint, the topology-aware builder's most-reliable-per-region pick is already near the efficient point. It is the fourth “another coverage axis doesn't pay for itself” result (a competing geometric-coverage portfolio, pre-pruning to a coverage-aware shortlist, the portfolio-coverage-health monitor, and now failure-mode coverage). Added as its own findings card near the candidate-pool-pruning card.
    • The prior live entry, the 2026-05-12 H4 full-corpus re-scope + the real-analysis figures, is archived as v7-2026-05-12-h4-rescope (resolved by the dynamic archive route).
    Open this archived version

  7. 2026-05-12

    archived

    H4 re-scoped to the full corpus + the real-analysis figures. At deployment scale, every crypto asset × indicator family × walk-forward window that lands a portfolio (2,548 cells, 26 assets, 7 families, k = 20), on the leak-free 45-D causal fingerprint, the topology selection-and-construction Stack beats Hierarchical Risk Parity by +0.96 ΔSortino (Holm-significant, wins ~61 % of cells) and edges out a plain reliability-ranked top-k (+0.37), but does not beat the naive in-sample-Sharpe-rank top-k portfolio (−0.50) and carries the deepest drawdowns of the four (≈ 9 positions vs 20); the edge concentrates in the momentum-oscillator families and is ≈ 0 on the moving-average families. So the headline is restated from “+4.79 Sortino over 26 MACD cohorts, ~54 % smaller drawdown” to: risk-parity-beating, audit-traceable portfolio construction, not raw-return dominance. Four real-analysis figures land, the actual Mapper / Reeb graph on Bitcoin/MACD, the deployment-scale verdict bars, the H4 equity curves on real cohorts, the H17 persistence-diagram trajectory, replacing the pedagogical Mapper schematic, the per-cohort forest plot, the drawdown box, and the H17 correlation lollipop; the Deployment-Stack decision-flow and the prediction-vs-decision partition figures are re-rendered to the +0.96 number. Date labels bumped to 2026-05-12.

    • H4 PASS restated at deployment scale. On the leak-free 45-D causal fingerprint, across every crypto-asset × indicator-family × walk-forward-window cell that lands a portfolio (2,548 cells, 26 assets, 7 indicator families, k = 20): the topology Stack beats Hierarchical Risk Parity by +0.96 ΔSortino (95 % CI [+0.75, +1.17], Holm-significant; wins 1,554 of 2,548 cells ≈ 61 %) and edges out a plain reliability-ranked top-k portfolio (+0.37), but loses to the naive in-sample-Sharpe-rank top-k portfolio (−0.50), and carries the deepest drawdowns of the four (mean max-drawdown magnitude ≈ 279 vs ≈ 254 for HRP and ≈ 138 for the in-sample-Sharpe rank, the Stack runs ≈ 9 positions where the others hold 20). The edge is concentrated in the momentum-oscillator families (MACD ≈ +2.2 ΔSortino vs HRP, Stochastic ≈ +2.4) and ≈ 0 on the moving-average families. The earlier “+4.79 over 26 MACD cohorts, ~54 % smaller drawdown” figures were the MACD-only slice; folding in the six non-MACD families brings the corpus number down, and reverses the drawdown sign. The headline is restated everywhere it appeared (dek, central claim, the H4 finding, the Phase-06 detail, the changes log): risk-parity-beating, audit-traceable construction, not raw-return dominance. Sampling-based topology stays invariant to the program-wide contamination fix by construction (the selector's decisions never pass through a regression of OOS Sharpe onto the fingerprint).
    • Four real-analysis figures replace illustrative or now-superseded ones. The actual Mapper / Reeb graph on Bitcoin/MACD (~2,000 backtested strategies covered by 17 Mapper nodes over the mean-OOS-Sharpe lens, a thin tail of catastrophic-Sharpe nodes hinging into a three-way fan where ~96 % of the population sits, one connected component; replaces the pedagogical Mapper schematic on the H4-method block). The deployment-scale verdict bars (mean paired delta of the Stack vs three baselines on Sortino / Sharpe / Calmar / max-drawdown, bootstrap 95 % CIs, Holm-significant pairs marked; replaces the per-cohort H4 forest plot on the H4 finding). The H4 equity curves on real cohorts (cumulative OOS PnL over the walk-forward windows for Litecoin/MACD, Bitcoin/Stochastic-%K, Ethereum/MACD and Ethereum/EMA, the Stack above hierarchical risk parity and below the naive Sharpe-rank on the oscillator cohorts, all four bunched on the moving-average cohort; replaces the H4 drawdown box). The H17 cohort persistence-diagram trajectory (Bitcoin/MACD's 16 movement phases on a tightly-wound path vs a longer-history cohort's 5 on a longer one, step-size strips, the segment-count × OOS-stability ρ = +0.72 stat; replaces the H17 correlation lollipop). The Deployment-Stack decision-flow diagram and the prediction-vs-decision partition map are re-rendered with the +0.96 number in place of +4.79.
    • Date labels bumped to 2026-05-12 (last-updated, findings header, roadmap-status caption); the roadmap-status caption and the H4 phase-detail eyebrow now note the full-corpus re-scope.
    Open this archived version

  8. 2026-05-11

    archived

    Phase 21 H19 FAIL + the Topology Deployment Stack assembled: a per-strategy topological-drift score does not fire before the classic drawdown trigger (median lead 0 windows, 68 % false-positive rate against a 30 % bar), so there is no per-strategy drift trigger in the Stack; the actionable monitor is the cohort-level aggregation-trust check (H17 PASS), and the per-strategy drift profile ships as a visualization aid only. With H18 and H19 both closed, the Deployment Stack, the trader-facing pipeline map → select → diversify → monitor → pause/refit, is now assembled as a ten-command CLI with the FAIL-derived caveats baked in, a worked end-to-end example, a live adapter into an existing walk-forward pipeline, and the runbook / live-deployment / case-study chapters drafted. The prediction-vs-decision partition gains a fourth refinement (an added degree of freedom is necessary but not sufficient, it has to be predictive of the outcome), and two new figures land: the four-layer partition map (the program's centerpiece, on the central finding) and the Deployment Stack decision-flow diagram (on the H19 / Stack finding).

    • Phase 21 H19 (live-drift monitoring), FAIL on both arms. A per-strategy topological-drift score (neighbour-churn + density-collapse + position-drift + cohort-trust-decay), CV-calibrated and treated as an early-warning alarm, fires at the classic max-drawdown breach (median lead 0 windows) and lags the trailing-Sharpe<0 trigger by 1 window, with a 68 % false-positive rate against a 30 % bar (0 of 26 cohorts under it). The reason is structural: a strategy that draws down changes its realised return distribution by virtue of having drawn down, so the geometry lights up after the fact. So there is no per-strategy drift trigger in the Deployment Stack, the actionable monitor stays the cohort-level aggregation-trust check (H17 PASS), and the per-strategy drift profile ships as a visualization aid only (alongside the per-strategy geometry-signature audit aid).
    • The Topology Deployment Stack assembled, the trader-facing pipeline map → select → diversify → monitor → pause/refit is now a ten-command CLI (map, select, portfolio, audit, monitor, refit, aggregation-check, plateau-audit, universality-check, live-adapter), built entirely from the robust geometric-outlier primitives, with the FAIL-derived caveats baked into every relevant command's help text (six of the ten carry one, the honest negatives are part of the deliverable), a worked end-to-end example on a fresh cohort, a thin live adapter that consumes an existing walk-forward pipeline's output with no changes to that pipeline, the runbook / live-deployment / case-study monograph chapters drafted, and the external user-test package ready to dispatch.
    • The prediction-vs-decision partition gains a fourth refinement (the H19 finding): an added degree of freedom is necessary but not sufficient, it has to be predictive of the outcome, and a 'drift' signal carries no information about future drawdown; layers 1 and 4 are the same fact from two directions, forecasting future performance is out of reach, however you package the forecaster. The partition is now a four-layer structure: forecasting topology fails or shrinks → covariate-level grain → adds-a-degree-of-freedom necessity → predictive-degree-of-freedom necessity, with the survivors (the Mapper-cluster portfolio, the cohort-aggregation diagnostic, the regime-rotation policy) in the sweet spot. Stated in the dek, the central-claim text, and the central finding.
    • Two new figures embedded: the four-layer prediction-vs-decision partition map (the program's centerpiece, on the central finding entry; supersedes the two-class contamination-fix shift chart as the partition's headline figure, though that shift chart stays as the empirical-evidence figure for the contamination-fix layer), and the Topology Deployment Stack decision-flow diagram (the five-box spine; the candidate-pool prune drawn absorbed into select; the per-strategy drift profile struck; the per-strategy audit aid alongside, on the H19 / Deployment-Stack finding).
    • Findings section gains one new entry (the H19 verdict + the Deployment-Stack-assembled milestone, with the decision-flow figure); the central-finding entry rewritten to the four-layer structure and given the partition figure; the deployment-phase roadmap node (Phase 17) and the central-claim block updated from 'planned / open hypothesis' to 'assembled'.
    • Roadmap statuses updated, H19 (node 16) and the Topology Deployment Stack (node 17) marked done; H5 cross-asset universality (node 15) marked in-progress.
    Open this archived version

  9. 2026-05-11

    archived

    State at the H14 / H15 / H18 deploy, before the H19 verdict and the Deployment Stack assembly: all six H12–H17 extension hypotheses had been tested (H17 PASS; H12, H13, H14, H16 marginal; H15 fail) and H18 candidate-pool pruning had landed FAIL; the prediction-vs-decision partition had been promoted to the program's central methodological finding with one sufficiency clause (from H18: topology has to add a degree of freedom the standard approach lacks); the Topology Deployment Stack had been re-specified to map → select → diversify → monitor → pause/refit (the prune stage collapsed into select) but was not yet built, and H19 (live-drift monitoring) was still listed as its remaining open hypothesis; Paper H had been added; the changelog archive links resolved; the codebase panel showed 61,861 lines / 207 files with a fifteen-group module breakdown.

    • Phase 07 (extended directions) marked COMPLETE, all six H12–H17 hypotheses tested. H14, MARGINAL (a geometric-outlier overfit-flag is a triage complement to the classical detectors, not a replacement). H15, FAIL (topology is reliably worse than the classical ex-ante features at predicting catastrophic strategy failure, held-out C-index ≈ 0.57–0.60 vs 0.77–0.79). (H12 had landed MARGINAL, the regime detector itself is null but the cohort-rotation policy that consumes it works; ships as Paper F.)
    • Phase 20 H18 (candidate-pool pruning), FAIL: there is no topology-aware-pruning step that improves on running the H4 portfolio constructor against the full backtest corpus; `topo-prune` survives only as an optional human-attention shortlist diagnostic. The Deployment Stack pipeline's `prune` stage collapses into `select`. H19 (live-drift monitoring) was the last open hypothesis.
    • The prediction-vs-decision partition promoted to the program's central methodological finding, with one sufficiency clause from H18 (topology has to add a degree of freedom the standard approach lacks). Stated in the dek, the central-claim text, and a new top finding; a new methodology paper (Paper H) captures it.
    • Phase 06 re-specified as the Topology Deployment Stack, the trader-facing pipeline map → select → diversify → monitor → pause/refit, built entirely from the robust geometric-outlier primitives; H19 (live-drift monitoring) was its remaining open hypothesis. Output portfolio grew to eight papers; the software package was named.
    • Findings section gained five new text-only entries (the prediction-vs-decision partition as the central finding, the H12–H17-extension-complete summary, H14, H15, H18). Changelog archive route files added for v2 / v3 / v4.
    • Roadmap statuses updated (Phase 07 marked done); the compute-and-support note rewritten.
    Open this archived version

  10. 2026-05-11

    archived

    State at the H4-verdicts deploy, before the H14/H15 verdicts closed the extended-research directions: H4 topology-aware portfolio construction had PASSED at corpus scale (the headline practical result, +4.79 Sortino vs HRP, ~54 % smaller drawdown, only −0.4 % shift under the lookahead-feature retirement); H13/H16 had landed MARGINAL and H17 PASS; the lookahead-feature retirement audit had closed with the two-class finding; six new figures had been added (19 total). The central-claim text led with H4 PASS but had not yet promoted the prediction-vs-decision partition as the program's central finding, and Phase 06 was still the generic 'practitioner workflow' phase (the Deployment Stack had not been specified).

    • Phase 05 marked FINISHED with H4 PASS at corpus scale; Phase 07 marked in-progress (three of six landed: H17 PASS, H13/H16 marginal); H12/H14/H15 still listed as pending.
    • Six new figures embedded (H4 forest plot, H4 drawdown box, H13 two-panel, H16 quality-aware-vs-blind bars, H17 lollipop, the two-class partition figure), 19 figures total.
    • Dek and central-claim text led with H4 PASS; claim (b) (better drawdown profiles than HRP/NCO) stated as supported at corpus scale.
    • Phase 06 still listed as 'practitioner workflow & live deployment' with a generic CLI-surface list; the Topology Deployment Stack and H18/H19 had not been specified; output portfolio was seven papers.
    • Last-updated label moved to 2026-05-11; this version's changelog entry had no archive route file (404 on click).
    Open this archived version

  11. 2026-05-11

    archived

    State immediately before the four corpus verdicts: H3 make-or-break had just failed and the program was reframing around what survives. Phase 05 (H4 portfolio construction) was elevated to next-active but not yet run; the six Phase-07 extension hypotheses (H12–H17) were all still planned. The lookahead-feature retirement had landed on the figures but its program-wide audit conclusions were not yet integrated into the page copy.

    • Phase 05 listed as the program's highest-EV remaining test, next-active, not yet started.
    • Phase 07 listed all six extension hypotheses (H12–H17) as planned; none had a verdict.
    • The central-claim text said claim (a) was falsified and the program was reframed around the surviving claims (b)/(c)/(d) plus the six new directions, claim (b) was a hope, not yet a result.
    • Findings section carried six entries (smooth manifold, H1 PASS, H2 falsification, H3 falsification, fingerprint feature importance, the per-feature breakdown); 13 real-output renders embedded.
    • Dek read "the make-or-break test (H3) just landed, and it failed; the program is reframing".
    Open this archived version

  12. 2026-05-10

    archived

    Phases 1-3 finished and named; H3 make-or-break verdict (FAIL) integrated; Phase 5 elevated; Phase 7 (H12-H17 extensions) added; real-output figures replace the illustrative SVG demos. Forward-looking WFO-OOS-Sharpe features still present in the fingerprint-importance figure at this point.

    • Project start (2026-05-08) and last-updated (2026-05-10) labels added; per-phase completion stamps added.
    • Roadmap expanded from 7 to 8 macro phases; new Phase 07 (Extended research directions, H12-H17) inserted.
    • Phases 01-03 marked finished with the dates they landed; Phase 04 marked deprioritised; Phase 05 elevated to next active.
    • Findings section added with 5 sub-findings (smooth manifold, H1 PASS, H2 falsification, H3 falsification, fingerprint feature importance) and real-output figures.
    • 13 real-output figures replace the illustrative SVG demos; every figure now clickable to open full-resolution in a new tab.
    • Compute & support callout added; the page mentions the H100 burst requirement and links to /research-support.
    • Scripts-on-request disclosure added; full toolkit ships publicly at program completion.
    • Changes section (before/after rows) and this Changelog index added.
    Open this archived version

  13. 2026-05-09

    archived

    Initial topology roadmap: 7 macro phases, Phase 1 marked in-flight with first-light findings, every other phase planned. Two illustrative SVG demo figures (persistence diagram, Mapper graph).

    • Initial publication.
    • 11 pre-registered hypotheses presented as the central applied claim.
    • Roadmap: 7 macro phases (Foundations · Geometric · Persistent topology · Local + Fisher-Rao · Topo-aware portfolios · Practitioner workflow · Synthesis).
    • Phase 1 status: in-flight; first-light findings on 3 pilot assets quoted in the body.
    • Two illustrative SVG demo components rendered (PersistenceDiagramFigure, MapperGraphFigure).
    Open this archived version
Back to topology roadmap
Topology roadmap, changelog | Daru Finance