← back

STATUS · CURRENT IN KIND, AS OF 2026-08-05. Still the most detailed inventory of the codebase, and its method (report the code, record the gaps) is unchanged. But it was assembled from main @ 545a2ac and the repository has moved since. Known drift, so you can read it with the right discount:

it says	today
engine toy-v0.3 / toy-v0.5	toy-v0.8 (src/ah/core/engine.py's TOY_ENGINE_VERSION)
schemas/worldspec-v1.2.schema.json	worldspec-v1.3.schema.json has landed alongside it
toy presets in the 5xx block	toy presets 511–514; the shipped world is stagflation_1974 …603; stress worlds are …701/702/703
sleeve mappings v1.0	v1.1 (map-2026.08.2) — PM rows re-estimated under AM-2026-08-12-001
the play surface is Play.tsx's panels	the CIO dashboard is the play surface (cio-03); Book/PrivateMarkets retired
—	the stress-scenario compiler (src/ah/gen/stress.py) did not exist
—	play alphas are port-v4-ladder / port-v4-ladder-gen
hier-flow-v1 beats the benchmark on the sealed criterion (§5.9)	retracted. The third campaign (2026-08-11, full 1953–2020) reversed it; bootstrap-v1 is generator of record by the challenger's failure, not by a sealed win

For what has been built and by what method as it stands today, read docs/current/METHOD.md first. Index: docs/current/README.md.

BUILD-SUMMARY.md — what actually exists

Assembled 2026-08-05 from the repository at main @ 545a2ac, by reading the code and running it. Where a design document and the code disagree, the code is reported and the disagreement is recorded in §5. Nothing planned is described as existing.

How to read this. §1 is the map. §2 is the inventory with evidence and invocation. §3 is the configuration, verbatim from the files that hold it. §4 is the test estate. §5 is the gap list, and §5 is the part worth your time — an orientation document whose gap section is short is a document that did not look.

1. System map

Dependency direction is downward only. port / gen / eval / artifacts depend on data, which depends on core; nothing points back up. The product surface (bundle, feed, serve, play, credibility, programme) sits on top of core + port and is the only layer the browser sees.

Terrarium/
├── src/ah/
│   ├── core/                  the numeric core — no pandas, no network, no clock
│   │   ├── worldspec.py       pydantic mirror of schemas/worldspec-v1.2.schema.json
│   │   ├── loader.py          jsonschema-first then pydantic (a property test pins agreement)
│   │   ├── validator.py       V1–V12 coherence rules; V10/V11/V12 block   → validate()
│   │   ├── numericworld.py    the narrative-free projection the engine consumes
│   │   ├── engine.py          the deterministic toy engine `toy-v0.3`     → run_path/run_ensemble
│   │   ├── institution.py     Step-0 toy institution (8 weights, no cash) → simulate_institution
│   │   ├── sleevestate.py     Step-3 sleeve/vehicle state contract
│   │   ├── institutionstate.py institution state for the twin
│   │   └── digest.py          canonical JSON + SHA-256 over float64 tensors
│   ├── store/                 SQLite; append-only where it matters
│   │   ├── db.py              schema + triggers (chronicle UPDATE/DELETE refused)
│   │   ├── worlds.py          save_world refuses edits to engine-consumed fields
│   │   ├── runrecords.py      RunRecords + verify_run (the anchor behind `ah replay`)
│   │   ├── chronicle.py       append-only event log per world
│   │   ├── sessions.py        play sessions: monotone pointer, final decisions
│   │   └── leaderboard.py     submit_score/scores over db.py:92's
│   │                          UNIQUE(world_id, seed, decision_alpha_version, participant)
│   ├── compiler/              scenario text → raw world document
│   │   ├── pipeline.py        validate → clamp → construct WorldSpec       → process()
│   │   ├── fixture_adapter.py the offline compiler (50 committed fixtures)
│   │   └── anthropic_adapter.py the live compiler, imported only under --live
│   ├── data/                  Step 1 — immutable Parquet vintages over DuckDB
│   │   ├── catalog.py         the vintage store; `current` advances only if QC passes
│   │   ├── connectors/        fred, bis, french, jst, shiller, treasury_hqm
│   │   ├── qc.py splice.py derive.py desmooth.py episode.py reports.py refresh.py
│   │   └── cli.py             → `ah data …`
│   ├── gen/                   Step 2 — the generator layer (research, NOT on the play path)
│   │   ├── registry.py        generator_id → factory                       → resolve()
│   │   ├── bootstrap.py       `bootstrap-v1`, the sealed benchmark
│   │   ├── climate/ regimes/  L1 (numpyro) and L2 (semi-Markov)
│   │   ├── blocks/            L3 — diffusion + flow (`hier-flow-v1`, torch)
│   │   ├── joinery/           L4 — waypoints, bridge, Denton reconciliation
│   │   └── systems.py         the five ablation systems A–E
│   ├── eval/                  the sealed judging code (hashed in pre-registration.lock)
│   │   ├── prereg.py g2.py g3seal.py g5seal.py     seals, locks, amendment log
│   │   ├── battery.py metrics/                     the Step-2 validation battery
│   │   ├── decision_metrics.py                     Step-5 sealed metric set
│   │   └── walkforward.py counterfactual.py episode2022.py negative_controls.py
│   ├── port/                  Step 3 — the translation layer and the real institution
│   │   ├── sleeves.py vehicles.py cohort.py        state objects
│   │   ├── cashflow_tier0.py cashflow_tier1.py     the commitment/call/distribution model
│   │   ├── portfolio.py                            cash account + coverage on both bases
│   │   ├── engine.py                               the quarterly waterfall  → PortfolioEngine
│   │   ├── twin.py                                 the DB-pension institutional twin
│   │   └── mapping.py smoothing.py heroes.py proxy.py
│   ├── artifacts/             Step 4 — the wire, the letters, the committee
│   │   ├── templates.py render.py                  Tier-1, pure functions of the tape
│   │   ├── author.py prompts.py gate.py            Tier-2 authoring + the G1–G9 gate
│   │   ├── bible.py validation.py committee.py     continuity, actor study, AI committee
│   │   ├── chronicle.py payloads.py                the G9 publication record (unused — §5.13)
│   │   └── live.py calendar.py windows.py decisions.py
│   ├── battery/               Step-0 stylized battery (thresholds still `todo` — §5)
│   ├── presets/               the four preset worlds
│   ├── splits.py              train/val/holdout + the FinalEvaluationToken guard (sealed)
│   ├── factors.py strategies.py  the factor manifest and the D4 strategy set (both sealed)
│   ├── experiment.py          local-first experiment tracking under experiments/
│   ├── play.py                the playable institution on the Step-3 twin → simulate_play
│   ├── bundle.py              the world bundle, contract `world-bundle-0.4` → build_bundle
│   ├── feed.py                the in-timeline tier-1 wire                  → build_tier1_feed
│   ├── serve.py               FastAPI session service — the scoring authority
│   ├── credibility.py         admin console: does this world's arithmetic look sane
│   ├── programme.py           the private-programme section of that console
│   ├── inspect.py             static figure page from a RunRecord alone
│   ├── tournament.py density.py                    Step-5 harnesses
│   └── cli.py exp_cli.py      → `ah`, `ah exp`
├── app/                       the React/TS player (vite 5173, proxies /sessions → 8787)
├── schemas/                   READ-ONLY vendored contract truth
├── mappings/                  frozen parameter artifacts (pacing, tier-0 spec, tier-1
│                              linkage, smoothing kernel, sleeve mappings)
├── fixtures/                  compiler, state, worlds, entity_screen, actor_validation,
│                              authoring_regression
├── tests/                     121 test modules + one shared helper
├── scripts/                   53 deterministic build/measure/train scripts
├── Instructions/              the authoritative plans, design notes and kickoffs
├── governance/                seals, amendment log, decision register, GenAI pack, evidence
├── docs/                      the realism register, data specs, and the superpowers plans
├── artifacts/                 committed research outputs (wp24…wp57, campaign2)
├── configs/ taxonomy/         search configs and the sleeve taxonomy
├── experiments/  (gitignored)  trained checkpoints and experiment records
└── data/         (gitignored)  the Parquet vintage store and data/ah.db

Entry points, exhaustively. The console script ah (pyproject.toml:35 → ah.cli:app); the ASGI app ah.serve:app (src/ah/serve.py:318); the module entry python -m ah.battery.report; the vite app in app/; and the 53 scripts in scripts/, each run with uv run python scripts/<name>.py.

2. Capability inventory

Status vocabulary: working — runs end to end and I ran it; partial — runs, but a named part of what the design asks for is absent; stub — present as an interface or placeholder without the substance behind it; absent — named somewhere but not built; unverified — I could not run it. Qualifiers after a status ("working, sealed"; "working locally, not reproducible from a clean clone") are free text, not vocabulary.

Flags are not exhaustive. The "how to invoke" column shows the shortest command that does the thing, not every option. Notably absent from it: ah data refresh --dry-run/--source/--vintage, the --data-root/--root overrides on the data and experiment commands, and ah credibility --seed. Run --help on any subcommand for the full set.

Coverage of this inventory, stated: core/, port/ and store/ are covered module by module. eval/, gen/, compiler/ and data/ are covered by capability, not by module — 28 of the 157 modules under src/ah/ have no row of their own, including sealed judged code (factors.py, strategies.py, eval/reference.py, eval/panel.py, eval/ablation.py, eval/sleevetails.py) and gen/base.py. See §5.15.

World generation

Compiler and coherence checks

Ensemble generation, replay, verification

Cashflow engine and the portfolio engine

Decisions, twins, decision alpha

The battery

Artifact generation

The product surface

The data layer

3. The numbers as configured

Every value below is read from the file cited. UNSET means the code has no default and nothing in the repo supplies one.

Run shape

Sleeves

Opening book — the scored institution

The Step-0 toy institution uses a different opening mix — 45 points private (core/institution.py:39), outside the play policy band. Both are live; see §5.2.

Costs, bands, policy

Pacing

Declared plausibility bands (admin console — priors, not truth)

Asset bands, annualized % over ten years (credibility.py:66): equity −8…14 ret / 12…28 vol · bonds −2…9 / 3…12 · hy 0…11 / 6…18 · commodities −6…16 / 12…28 · reits −8…12 / 12…26 (and pe/pc/re below those lines in the same table).

Programme bands (programme.py:209), all seven: peak_unfunded_ratio 0.25–0.75 · call_rate_y1_3 0.15–0.45 · crossover_years 4–8 · dpi_age9 0.7–2.0 · linkage_bite 0.50–1.20 · linkage_shortfall 0.05–0.35 · forced_secondaries 0.0–1.0.

Battery thresholds

All status: todo — excess_kurtosis, skewness, hill_tail_index, acf_r_lag1, acf_abs_lag1, max_drawdown_median, corr_distance (src/ah/battery/thresholds.yaml). No metric is enforce; the Step-0 battery therefore cannot fail. The Step-2 battery's thresholds are real and live in pre-registration.yaml under the seal.

4. Test estate

Suites

Structural invariants are tests, not conventions: narrative-blindness (tests/test_narrative_blindness.py scans the engine for narrative field access), no-network (pytest-socket, --disable-socket in addopts), leakage guard (tests/test_leakage_guard.py is an import-graph assertion), seal integrity (tests/test_seal_guards.py), the artifacts boundary (tests/test_artifacts_service.py::TestBoundary — core/gen/port never import artifacts), and the programme console's read-only guarantee (tests/test_programme_guard.py).

What CI runs

.github/workflows/ci.yml, on push to main and on every PR, in order:

1. uv sync --frozen --dev
2. uv run ruff check . and uv run ruff format --check .
3. uv run pyright
4. uv run pytest --cov=ah.core --cov-report=term-missing --cov-fail-under=90
5. uv run python -m ah.battery.report
6. uv run pytest tests/test_g0_end_to_end.py -v

Note the coverage gate covers ah.core only (pyproject.toml:69), not the rest of the tree. Two further workflows are scheduled, not gating: data-monthly.yml (a monthly public-source refresh that opens an issue on QC failure) and data-reminders.yml (quarterly manual-intake reminders).

Last-known status, measured here on 2026-08-05

5. Known gaps and discrepancies

Ordered by how much they would surprise someone who read the design documents and then read the code. Everything here was found in this pass; nothing is copied forward from a status table without re-checking.

5.1 The promoted generator cannot reach a player

Step 2's gate promoted hier-flow-v1 (G2-EVIDENCE.md; S2-DEFAULT-GENERATOR), and the WorldSpec enum carries it (core/worldspec.py:62). (That promotion was later reversed — see the standing caveat in §5.9. It is reported here as the position on 2026-08-05, not as a current result.) But every playable path runs the toy engine: ah run calls run_ensemble (cli.py:195), which calls _require_toy (engine.py:151) and raises UnsupportedGeneratorError for any other id. ah/gen/registry.py is never imported by ah/cli.py, ah/bundle.py, ah/serve.py, or ah/play.py. The generator layer is a research artifact; the product runs on toy-v0.3.

Two consequences worth stating separately:

• A fresh clone cannot instantiate the generator stack at all. hier_flow_v1_factory requires experiments/campaign2-flow-s0/checkpoint.pt and refuses anything whose SHA-256 is not c6addb54… (flow.py:551-557); it then loads the L1 and L2 posteriors from experiments/climate-l1-… and experiments/regimes-l2-…, each SHA-pinned (joinery/assemble.py:140-147). experiments/ is gitignored. It resolves on this machine; it would not on a clean checkout. Refitting/retraining is the documented alternative (scripts/fit_climate.py, scripts/fit_regimes.py, scripts/train_flow_final.py) and is GPU-hours of work.
• signature-mmd is a name with nothing behind it. It is in the schema enum and the pydantic mirror, and the registry has no factory for it — by design (worldspec.py:56: "raises UnknownGeneratorError until something registers under it"), but a world declaring it is constructible and unrunnable.

5.2 Two institutions are live at once, and both reach the player

ah/play.py (cash account, calls, forced sales) is what scores. ah/core/institution.py (eight weights, no cash, nothing can be owed) is Step 0's toy. The toy is still wired into surfaces the player sees:

• bundle.py:98 computes the toy twin (hold_course_twin) and :160 writes its final value into summary.twin_final_value (and run puts the same number into summary_stats.final_of_100, cli.py:224);
• bundle.py:158 puts the port twin's ledger into twin_ledger;
• feed.py:175,180 builds the quarterly institution statement and its peer bands from simulate_institution — the toy;
• inspect.py, tournament.py and density.py all run the toy.

This is not theoretical. In the single bundle I built from the stagflation preset (run 03a76c08…, seed 771204, 200 paths):

summary.twin_final_value          = 166.502   (toy institution)
twin_ledger.nav_reported[-1]      =  76.712   (port institution — the one that scores)

The same file states the same institution's terminal value twice, 2.2× apart, and the wire the player reads is narrated off the higher one. The bundle docstring's own framing ("the HOLD-COURSE TWIN's cashflows") is accurate for twin_ledger; summary.twin_final_value is the leftover.

5.3 ER-6: about 29% of every commitment is never called

docs/engine-realism-register.md §ER-6, status open, dated 2026-08-05. The call curve is an ALB-A placeholder that was never fitted; measured at committed = 1.0, paid-in reaches 0.7075 over the contractual life against an allocator's 85–95% expectation. Downstream, peak_unfunded_ratio flags on all four presets (3.018 / 2.445 / 3.308 / 2.453 against a declared 0.25–0.75) and the J-curve crossover lands at ~8.5 years against a declared 4–8. On deflation_bust no year-1 vintage crosses over within the decade at all. The register is explicit that this is a prerequisite for the commitment lever (E1), not a parallel cleanup.

Two further register entries are open: ER-2 (the policy rate is a continuous drift with no meeting calendar and no 25bp quantisation) and ER-5 (the crisis is a rectangular block, giving equity an ACF of 0.364). ER-1 and ER-4 are closed in toy-v0.3; ER-3 is closed in the play surface (wire-play-surface) — a different layer, since ER-3 was the no-cashflow gap and was fixed in ah/port/, not in the return process.

5.4 The leaderboard is unreachable from the dev app

app/src/lib/session.ts:176 fetches /leaderboard/{worldId}. app/vite.config.ts:8 proxies only /sessions. Verified against the running dev stack:

GET http://127.0.0.1:8787/leaderboard/x?seed=1&alpha_version=1.0  → 200 application/json
GET http://localhost:5173/leaderboard/x?seed=1&alpha_version=1.0  → 200 text/html      (vite's SPA fallback)
GET http://localhost:5173/sessions/nope                            → 404 application/json (proxied correctly)

The client parses the SPA's HTML as JSON and throws; Leaderboard.tsx:31 catches it and renders the error string in place of the board. The server side is fine — the same request direct to 8787 returns a well-formed empty board. It is a one-line proxy entry.

5.5 ruff format --check . fails on main today

ruff 0.16 formats Python code blocks inside Markdown, and two committed plan documents contain unformatted blocks:

• docs/superpowers/plans/2026-08-04-private-programme-console-section.md
• docs/superpowers/plans/2026-08-05-world-and-wire-audit.md

uv.lock pins ruff 0.16.0 and CI runs uv sync --frozen, so CI's lint step would take the same view. ruff check . and pyright are both clean; this is the formatter alone, and ruff format on those two files fixes it.

5.6 The battery CI step cannot fail

.github/workflows/ci.yml runs python -m ah.battery.report as a gate, but every threshold in src/ah/battery/thresholds.yaml is status: todo, and BatteryReport.passed is not enforce_failures (battery/report.py:58-64). With no enforce metric the step is green by construction. The file is candid about this ("placeholders documenting intent, not ratified thresholds") — the gap is that CI presents it as a gate.

The CI step is also mislabelled. It is titled "Validation battery (stagflation preset)", but main() runs _stagflation_ensemble() (battery/report.py:189-196), which loads schemas/example-long-stagflation.worldspec.json — a schema example, not src/ah/presets/stagflation.json — forces generator_id = "toy-v0", and runs 64 paths at base_seed=42, not the preset's 1000 at 771204.

Note the Step-2 battery (ah/eval/battery.py) does have real, sealed thresholds and all eight suites registered (_REFERENCE_DEPENDENT_SUITE_BUILDERS, battery.py:955); the two are different objects with similar names.

5.7 Tier-2 authoring passes its bar and still ships nothing

governance/evidence/AUTHORING-REGRESSION.md records the current run as 30/30, 100.0% first-pass, "ships: YES" against the frozen ≥95% bar, and governance/evidence/G4-EVIDENCE.md closes criterion 1 at 96.7%/100.0%. But every bundle still stamps:

meta.artifact_tier = "tier-1 templated wire (build-time, deterministic);
                      tier-2 letters await the frozen >=95% first-pass bar"

(bundle.py:139-142, and I read it back out of a freshly built bundle).

The string is stale, but the deeper point is that it describes the wrong rule. The ratified decision, PD-4 (Instructions/KICKOFF-PRODUCT-SU.md:78), is key-conditional, not bar-conditional: "tier-1 deterministic always; tier-2 letters when a key is present at build, recorded in the bundle either way." bundle.py has no such branch — it never imports ah.artifacts.author, so a key present at build time would change nothing. The stale string is the symptom; the missing PD-4 branch is the gap.

Also unbuilt: G3's leak-checker prompt over the draft, recorded as v1.1 in artifacts/gate.py's docstring, and G6 which warns "not evaluated" unless past artifacts are supplied.

5.8 Documentation that no longer matches the code

5.9 Gate state, carried forward and re-checked

These are the project's own recorded positions, verified against the evidence files rather than restated from the status table:

• Step 3's G1-completion is an honest FAIL — G1-EVIDENCE.md: mark_lag (must-pass) failed at 0.9667 and private_weight_breach failed at 0.0332. Tier 1 beat tier 0 (2 criteria failed vs 3). G3 itself was never taken.
• The holdout is spent — declined at G2 on purpose, spent at WP5.6. RESEARCH-EVIDENCE.md: primary drawdown surprise −0.3952, realized terminal wealth at the 99.6th percentile of the ensemble, cpi band coverage 0.000, and fx_usd went unread because of a reader fallback bug that, per the sealed spec, publishes as a gap rather than re-running.
• The severe test is inconclusive and the L3 leg structurally vacuous (0 blocks dropped) — artifacts/wp211/SEVERE-TEST.md via RESEARCH-EVIDENCE.md.
• The standing caveat — as written on 2026-08-05 this read "hier-flow-v1 beats the benchmark on the sealed criterion". That claim is retracted. It held at G2 and at the second campaign; the third campaign (2026-08-11, on the full 1953–2020 span) reversed it, because the challenger's edge was smaller than its own run-to-run noise, which under the sealed rule is not a win. hier-flow-v1 is shelved and bootstrap-v1 is the generator of record by the challenger's failure rather than by a sealed win (docs/current/METHOD.md, which governs). Neither is a convincing model of history: regime persistence undercalled, drawdowns understated ~2×, the decade tier 73% structurally unavailable (G2-EVIDENCE.md §7–8). Nothing built on either is decision-ready. Do not cite G2 alone as a standing result.

5.10 Named-but-absent, and deferred-by-decision

• ah audit — 2,287 lines of design and implementation plan merged at 545a2ac; no code, no CLI command.
• Drift twin — the third analysis series is a None in the API contract (serve.py:300), by design, so its arrival is a data change.
• E1 commitment lever — Instructions/experience-deltas-register.md marks E1, E2 and E4 PARTIAL, E3 and E5 OPEN, E6 OPEN (I5's ~20-user first-run observation deferred under D-K4-5), E7 and E8 CLOSED.
• Auth — deliberately absent at v0.1 (serve.py:8-10); required before any external user.
• Human-cohort benchmark and the too-rational pathology — deferred by owner decision D-K4-5, recorded in G4-EVIDENCE.md criterion 4/5 rather than proxied with a number.
• Fee/carry on the play path — simulate_play steps cohorts directly (play.py:385) and never calls run_tier1, so the management fee, the European carry, recycling, subscription-line deferral and extension behaviour that cashflow_tier1.py implements are not applied to the scored institution. Whether that is intended is not stated anywhere I found.
• Preset structural coverage — three of four presets carry only parameter_vintage under structural; the toy engine falls back to _DEF (engine.py:89-106) for private-market parameters in those worlds.

5.11 A secondary sale relieves nothing — the obligation survives it

This is on the scored path, and it inverts the sign of the only liquidity lever the player has.

Both the voluntary sale (play.py:298-307) and the forced secondary (port/engine.py:136-140) reduce cohort.nav_true and cohort.nav_reported and credit cash. Neither touches unfunded or committed. Nothing else does: in port/cohort.py, unfunded moves only in step() (:200) and recall() (:233); report() (:207-211) writes the mark and returns. And calls are sized call = min(self.unfunded, call_rate * self.unfunded) (cohort.py:185) — a function of the unfunded balance, independent of NAV.

So you sell an LP interest at a 19% haircut, keep 100% of the future call obligation, and — since coverage is unfunded ÷ what you own — coverage gets strictly worse after the one action that exists to relieve liquidity stress. A cohort driven toward nav_true ≈ 0 by repeated forced secondaries keeps calling at full size. Not in docs/engine-realism-register.md (ER-1…ER-6 do not reach it).

5.12 Two writers to the leaderboard, and a version-label collision

The decision_alpha_version column is the mechanism that is supposed to keep incompatible scores off one board (db.py:92, UNIQUE(world_id, seed, decision_alpha_version, participant)). Three numeric definitions exist and two of them share a label:

tests/test_play.py:167 asserts only that PLAY_ALPHA_VERSION differs from the sealed constant; nothing pins runrecords.py's constant against it. The comment at runrecords.py:22-24 ("inert… Nothing reads them yet") is stale — tournament.py reads it and writes it to the board. Every bundle also stamps meta.decision_stamps.decision_alpha_version = "1.0".

5.13 Named-but-unwired, beyond §5.10

• src/ah/artifacts/chronicle.py — 88 lines implementing artifact publication as type='artifact' chronicle entries, described as the G9 record. Its only importer anywhere is tests/test_artifacts_service.py. §5.7 lists G3 and G6 as the unbuilt gate pieces; G9's record is unwired too.
• scripts/build_artifact.py executes Path("scratch_data.json").read_text(...) at module import (:9). That file is not in the tree and never was (git log --all finds nothing). The script cannot be imported, let alone run — yet CLAUDE.md still lists it under "Regenerating committed fixtures/artifacts (all deterministic)".
• ah data intake apply — "Validate a drop AND apply it: write the vintage, run QC, advance on pass." A state-mutating command with no row in §2, which documents intake validate only.

5.14 Pacing is open-loop, and one fixture governs all three private sleeves

Three separate simplifications, none of them registered:

• The commitment size never responds to anything. amount = START_TARGETS[asset] * _ANNUAL_COMMITMENT_RATE (play.py:213) is a constant. Not a function of current NAV, the world, the coverage ratio, or any decision. A portfolio halved by a crash keeps committing ~6.3 points a year, driving private weight up mechanically at the worst moment.
• PC and RE run buyout's cashflow profile. play.py:334 loads one base document and stamps sleeve_id onto copies (play.py:217); mappings/pacing-parameters-v1.0.yaml carries exactly one sleeve row, pm_buyout. Private credit and real estate therefore inherit buyout's rc_curve, bow 2.5, terminal yield 0.55 and 10-year life. There is no distinct PC or RE cashflow shape in the model at all.
• Equity drawdown drives every sleeve's linkage. play.py:337-339 computes drawdown_depth from q_returns["equity"] alone and feeds the resulting f_call/f_dist to pe, pc and re.

ER-6 covers the shape of rc_curve. None of these three is in the register.

5.15 Smaller things that still change what a reader concludes

• The decision actions are absolute point amounts that can silently truncate. _SHIFT_POINTS = 10.0 and _SECONDARY_POINTS = 8.0 are value units on a book that starts at 100 and ends anywhere; at the ninth window on a book worth 76, "10 points" is 13% of it. Worse, _rebalance returns silently when the source leg is empty (play.py:272-277) and the secondary caps at available NAV (:301) — so a chosen action can execute partially or not at all and surface as a 0.0 window contribution indistinguishable from hold.
• CI never runs the app, and never runs on Windows. ci.yml has no Node step: the React suite (32 tests), its typecheck, its build, and the client-side seal implementation are all ungated. And runs-on: ubuntu-latest only — so the ASCII-CLI rule in §5.8 could not be enforced by CI even in principle.
• CI never validates the promoted checkpoint. Six tests skip when experiments/ is absent — tests/test_blocks_flow.py:689 ("no primary flow checkpoint present (pre-training checkout)"), :806, :832, :855, test_blocks_diffusion.py:481, test_blocks_tuning.py:204. On a clean CI clone all six skip. The "2,226 passed" figure in §4 is this machine's; CI sees fewer, and none of them exercise hier-flow-v1's checkpoint.
• 28 of 157 modules under src/ah/ have no inventory row, including sealed judged code: factors.py and strategies.py (both in pre-registration.lock), eval/reference.py (2,540 lines, sealed in both pre-registration.lock and pre-registration-g3.lock), eval/panel.py, eval/ablation.py, eval/sleevetails.py, and gen/base.py. Worth noting against §5.9's "G3 itself was never taken": G3-pre's judged code exists and is sealed.
• README.md still describes the repository as Step 0. "This repository contains Step 0 — 'lay the rails'… No real market data, no ML training, no UI, and no LLM in the numeric path." Three of those four negations are now false (there is a real vintage store, there are trained models, there is an app); only "no LLM in the numeric path" still holds, and it is a standing invariant. Its architecture tree lists six components and omits data/, gen/, eval/, port/, artifacts/ and app/. It is the repository's front door.
• app/README.md is stale — it describes App.tsx as "the v0.1 shell (browse mode; session binding is su-app-02)" and lists four files, written before su-app-01…05 merged.
• The governance registers are not read by this document. governance/retrofit-register.md is an append-only log of 96 dated deferrals, and neither this summary nor the manual cites it, the decision register, the amendment log, the model inventory, the prompt registry, or the EU AI Act mapping. One row bears directly on §5.9's standing caveat: RFR-66 records that the G2 head-to-head is biased toward promotion by the benchmark's data window — bootstrap-v1 can only resample 1990–2020 while the challenger saw 1929–33, 1937, 1973–74 and 1987, and both are scored against realizations that include all of it. It was sealed into the pre-registration rather than deferred. Anyone auditing §5 for completeness should read that file next; it is the project's own gap register and it is longer than this section.

Evidence for every "working" claim in §2 is either a file:line in this repository or a command I ran on 2026-08-05; the commands and their real output are in docs/USER-MANUAL.md. Anything I could not verify is marked, and the full list of what I could not verify is in the closing note of that document.