mktmind/asx-lake

ASX Document Lake — agent query cookbook

A queryable lake of tables extracted from ASX mining-company PDFs, where every served cell is consensus-confirmed and judge-grounded (DECISIONS #1). Release v2.5: 12,371 served cells / 63 served issuers / 548 pages / 7 served families, Hive-partitioned Parquet, queryable with plain DuckDB — no custom Python.

This README is the agent-facing cookbook: how to pull the lake, the catalog schema, the trust filter, the storage layout, the cross-company query patterns, and a worked section per mining sector skill (producer / developer / explorer) with copy-paste SQL and the measured outcome of each query.

Every classification (Works / Abstains / Served-below-floor) and every issuer count below was measured by running the exact SQL against the published v2.5 lake at /tmp/v2.5-release/. The served catalog is under releases/v2.5/catalogs/.

What's in the lake

family	cells	issuers	pages	concept_id?	family_fidelity
drilling_assays	6699	31	98	—	0.9876 (method-certified)
financial_statements	2294	37	130	32%	0.9695 (method-certified)
jorc_resource_reserve	1939	34	86	—	0.9869 (method-certified)
appendix_5b_cashflow	603	26	67	65%	1.0000 (method-certified)
tenement_schedule	407	24	64	—	0.9630 (method-certified)
production_kpis	243	7	7	—	0.5745 (served, NOT method-certified)
study_economics	186	7	14	—	0.6492 (served, NOT method-certified)
Total	12,371	63	466	—	—

Pages: 548 is the lake's page count (the page catalog / receipt pages_completed). 466 is the count of distinct pages that carry at least one served cell in trusted_table (per-family page counts sum to 466 — no page spans two families). 63 served issuers is distinct ticker in trusted_table; the document/page catalogs cover 64 tickers (one issuer has pages but no served cells).

Two things every agent must know before the first query:

1. Units are partially populated in v2.5. The unit, currency, scale, normalized_value, and normalized_unit columns exist in the schema and are populated on a minority of rows (measured against the published v2.5 catalog): unit 2,357/12,371 (19.1%), currency 2,226 (18.0%), scale 903 (7.3%), normalized_value/normalized_unit 2,846 (23.0%). You can filter by unit/currency/scale where populated (WHERE unit='g/t', WHERE currency='AUD' return rows) but coverage is uneven by family, so most cross-company numeric work still relies on concept_id (same economic concept). See Units status (v2.5).
2. Two families ship below the method-certification floor. production_kpis (0.5745) and study_economics (0.6492) are served — every cell still clears the per-cell consensus+judge guarantee (DECISIONS #1) — but their family-level numeric fidelity is measured below 0.95. WHERE family_fidelity >= 0.95 selects the 5 method-certified families (11,942 rows); the other 429 rows are the two below-floor families. See Querying by trust.

production_kpis is the genuine rule-#1a vendor-arm family (withdrawn from the certify denominator); study_economics is a regular gated family that failed its 0.95 floor (in_denominator=true, the family whose presence drives the combined certify REJECT). No vendor cells ship in v2.5 — arm_kind='consensus' on all 12,371 rows — so both below-floor families are served on the consensus+judge basis only.

How to pull it

# Option A — the asx-lake CLI (pins the revision in lake.ref)
uv run asx-lake lake pull --dir ./lake

# Option B — plain huggingface CLI
hf download mktmind/asx-lake --type dataset --local-dir ./lake

Both write the dataset to ./lake. The served catalog is under releases/v2.5/catalogs/. Verify a release after pulling:

uv run asx-lake release verify --release v2.5 --out ./lake
# ok release=v2.5 sha256=be1dbca3… trusted=12371

The pinned revision is mktmind/asx-lake@b4e40d04… (lake.ref, written last on publish).

Open it in DuckDB (one-liner)

-- Replace <LAKE> with your pull dir. ticker is a hive partition column
-- (pushdown-friendly). document is a DATA column since v2.5 (not a partition);
-- it prunes via row-group statistics.
SELECT * FROM read_parquet(
  '<LAKE>/releases/v2.5/catalogs/trusted_table/**/*.parquet',
  hive_partitioning=true)
WHERE family = 'appendix_5b_cashflow' AND ticker = 'A8G';

Throughout this cookbook <LAKE> means the path you passed to lake pull.

Catalog schema (trusted_table)

Each row is one served cell — a single (family, measure, identity_path) value that survived consensus + judge grounding. The v2.5 catalog has 27 columns (a hard cutover from v2.4's 22: identity JSON-encoded VARCHAR and column were dropped; identity_path, identity_leaf, judge_confidence, family_fidelity, arm_kind, schema_version were added — no aliases).

column	meaning
ticker	ASX code (Hive partition — pushdown-friendly)
document	ASX announcement id (DATA column since v2.5; prunes via row-group stats)
page_id	ticker/document/page
family	one of the 7 served families above
region	in-table location hint
measure	column header text as printed (renamed from column in v2.5 — column is a SQL reserved word)
identity_path	LIST<VARCHAR> row key, e.g. ["current assets","cash and cash equivalents"] (native list since v2.5; replaces the JSON-encoded identity string)
identity_leaf	last element of identity_path, pre-extracted for join-by-leaf patterns
value	DOUBLE — the numeric value (populated on all rows)
unit	parsed unit — partially populated in v2.5: 2,357/12,371 rows (19.1%); concentrated in drilling_assays (g/t, %, m, ppm) and jorc_resource_reserve (g/t, %, tonnes, Mt)
currency	parsed currency — partially populated in v2.5: 2,226/12,371 (18.0%) (AUD 2,190, USD 36); concentrated in financial_statements and appendix_5b_cashflow
scale	parsed scale ('000, M) — partially populated in v2.5: 903/12,371 (7.3%) ('000 781, M 122)
concept_id	normalised cross-company join key — on financial_statements (32%) and appendix_5b_cashflow (65%); NULL on every other family (by design, never invented)
normalized_value / normalized_unit	scale-normalized value + unit — partially populated in v2.5: 2,846/12,371 (23.0%); normalized_value undoes scale (×1e3 for '000, ×1e6 for M) so it is comparable across mixed-scale prints where scale is present
agreement	consensus agreement fraction (trust provenance)
reading_arms	number of pinned arms that read the cell (2/3/4 observed)
arm_kind	consensus (DECISIONS #1) or vendor (DECISIONS #1a) — consensus on all rows in v2.5 (no vendor cells shipped)
judge_confirmed	boolean — judge confirmed the value is printed at its cited region (true on all served cells)
judge_confidence	DOUBLE — per-cell judge confidence in [0,1] from the judge verdict (range 0.72–1.0, avg 0.96–0.98 per family); non-null on every served v2.5 cell; NULL would mean a consensus-only cell (none in this release)
judge_cli_version / judge_model_id	judge provenance (DECISIONS #6)
family_fidelity	DOUBLE — per-family numeric fidelity from the combined certify scorecard; non-null on all 12,371 rows in v2.5 (the publish passed --family-fidelity). Method-certified families carry values ≥ 0.95; the two below-floor families carry their measured values (0.5745, 0.6492). NULL only when no certify measurement was supplied at build time
bundle_sha256 / object_key	content-addressed evidence backref
trust_decision	the served/trusted decision
schema_version	catalog schema tag (trusted_table/v2.5)

Two notes:

1. The v2.5 schema is a hard cutover: column → measure, and identity (JSON-encoded string) → identity_path (native LIST<VARCHAR>) + identity_leaf. There are no backward-compat aliases. Pinned v2.4 consumers must update.
2. The agent join key is (family, measure, identity_path) where identity_path is a native LIST<VARCHAR> (since v2.5). identity_leaf is the pre-extracted last element for join-by-leaf patterns. Verbatim keying is correct but fragments: on the standardised Appendix 5B, "interest received" appears as multiple identity paths. That is what concept_id fixes — join on concept_id for cross-company work (financials + 5B only).

---

Storage layout

The served lake's Parquet catalogs are laid out bench-driven, not by preference. The full design record is docs/storage-layout.md in the source repo; this is the agent-facing summary.

• Format: Parquet, SNAPPY compression, write_statistics=True (per-column min/max/null_count on every column of every row group).
• Partitioning: Hive, single column ticker — directories are ticker=<ASX code>/. document is a data column since v2.5 (queryable but not a path segment); WHERE document=... prunes via row-group statistics.
• File sizing: max_rows_per_file = 50_000, max_rows_per_group = 50_000 (one row group per file). At v2.5's scale this is one file per ticker → 63 shards, down from v2.4's 240 ((ticker, document) layout, one ~7 KB file per document).
• Bloom filters: deferred (measured zero benefit — DuckDB 1.5.x does not consume Parquet bloom filters; ~90× write cost). Revisit conditions in the doc.

Why document is no longer a partition. The v2.4 (ticker, document) layout was the natural write grain but the wrong scan grain: at 12,371 rows it produced 240 shards averaging ~7 KB each, and the cookbook's dominant queries (WHERE family=..., WHERE concept_id=..., WHERE family_fidelity >= 0.95) don't filter on document, so they opened all 240. A per-document partition cannot be consolidated, so document became a data column to allow multiple documents per file.

Bench result. Four strategies were measured on five cookbook queries at 1× / 13× / 100× scale (the 13× point is the realistic Phase-8 cohort target, ~806 issuers). Winner: right-sized files (S3) — ticker partition + max_rows_per_file consolidation. At the 13× decision scale it took the geometric mean from 57.95 ms (S1 baseline) to 15.25 ms — a 3.8× win — and the per-query wins are uniform. The win is almost entirely file-open tax: at 13×, the v2.4 layout opens 3,120 shards where the v2.5 layout opens 819.

scale	rows	files (this layout)	per-query band
1× (v2.5)	12,371	~63	1–2 ms
13× (Phase-8 cohort)	~160 k	~820	7–26 ms
100× (stress)	~1.2 M	~6,300	60–270 ms

For the full bench table, the rejected strategies (family-first, bloom filters), and the trigger to move to Iceberg/Delta, see docs/storage-layout.md.

---

Querying by trust

Every served cell is consensus-confirmed + judge-grounded (DECISIONS #1), but the families are NOT all measured the same way. The method-certified families (JORC / drilling / financials / 5B / tenement) clear a 0.95 numeric-fidelity floor; production_kpis (0.5745) and study_economics (0.6492) are served below it. family_fidelity and judge_confidence make that split queryable.

-- Method-certified families only (fidelity >= 0.95, the highest-trust subset).
-- Returns 11,942 rows: the 5 method-certified families. Drops the 429
-- below-floor rows (production_kpis + study_economics) entirely.
SELECT * FROM read_parquet(
  '<LAKE>/releases/v2.5/catalogs/trusted_table/**/*.parquet',
  hive_partitioning=true)
WHERE family_fidelity >= 0.95;

-- High-confidence judge verdicts only (per-cell grounding confidence).
-- judge_confidence ranges 0.72-1.0 in v2.5 (avg 0.96-0.98 per family).
SELECT * FROM read_parquet(
  '<LAKE>/releases/v2.5/catalogs/trusted_table/**/*.parquet',
  hive_partitioning=true)
WHERE judge_confidence >= 0.8;

-- The two served-but-below-floor families (the 429 rows the >= 0.95 filter drops).
SELECT family, count(*) FROM read_parquet(
  '<LAKE>/releases/v2.5/catalogs/trusted_table/**/*.parquet',
  hive_partitioning=true)
WHERE family_fidelity < 0.95
GROUP BY 1 ORDER BY 1;
-- production_kpis 243, study_economics 186

NULL semantics — read before you filter:

• judge_confidence is non-null on every served v2.5 cell (all 12,371 are judge-confirmed). It would be NULL only for a consensus-only cell (a path that exists in the system but produced zero cells in this release). NULL here means "no judge looked at this cell", not "the judge was unsure" — use judge_confidence IS NOT NULL then a threshold. Note: judge_confidence is the judge's certainty of grounding (DECISIONS #6), not a per-cell fidelity measurement.
• family_fidelity is non-null on all 12,371 rows in v2.5 because the publish passed --family-fidelity. Method-certified families carry their measured value (≥ 0.95); the two below-floor families carry theirs (0.5745, 0.6492). NULL would mean "no combined-certify measurement was supplied at build time" — that is the omitted---family-fidelity case, not this release. WHERE family_fidelity >= 0.95 therefore correctly selects the method-certified subset.

---

Cross-company query patterns

These patterns are the "fast comparison" enablers. All measured on v2.5. The concept_id patterns work as-is; the units-based patterns partially work too — units are populated on 19.1% of rows, so WHERE unit='g/t' (369 rows / 24 issuers) and WHERE currency='AUD' (2,190 / 44 issuers) now return rows. See Units status for the per-family coverage that bounds each pattern.

Pattern 1 — join on concept_id (the canonical win)

concept_id collapses printed-text variants of one economic concept to one id. "Interest received" — multiple raw identities — becomes one concept spanning 15 issuers.

-- 5B interest received across issuers, one concept_id
SELECT ticker, round(max(value),0) AS value
FROM read_parquet(
  '<LAKE>/releases/v2.5/catalogs/trusted_table/**/*.parquet',
  hive_partitioning=true)
WHERE concept_id = '5b.1.4.interest_received'
GROUP BY ticker ORDER BY ticker;
-- 15 issuers reached (Works).

The deepest financials join keys (measured on v2.5):

concept_id	issuers	cells
fin.bs.current_assets.cash	27	63
fin.bs.non_current_assets.ppe	27	61
fin.bs.non_current_assets.total	27	54
fin.bs.total_liabilities	26	56
fin.bs.current_liabilities.trade_payables	25	51
fin.bs.total_assets	25	49
fin.bs.net_assets	25	47
5b.1.9.net_operating	22	50
5b.3.1.proceeds_equity	13	37
5b.1.4.interest_received	15	32

Pattern 2 — filter by family + identity_leaf

concept_id is NULL on the non-financial families, so for JORC / tenement / drilling the cross-company surface is family + identity_leaf (the pre-extracted last row-key element). The JORC category labels are shared across issuers, so category split works even without concept_id:

-- JORC resource category split across issuers (category labels ARE shared)
SELECT identity_leaf, count(*) AS cells, count(DISTINCT ticker) AS issuers
FROM read_parquet(
  '<LAKE>/releases/v2.5/catalogs/trusted_table/**/*.parquet',
  hive_partitioning=true)
WHERE family = 'jorc_resource_reserve'
  AND lower(identity_leaf) IN ('total','inferred','indicated','measured')
GROUP BY 1 ORDER BY issuers DESC;
-- total=17iss/88cells, inferred=10/37, indicated=10/39, measured=5/19
-- (183 cells across the 4 categories).

Pattern 3 — drilling by aggregation/benchmark (NOT identity join)

drilling_assays keys on (Hole ID, Depth From) — holes are company-specific, so identity keys are not shared across issuers. In v2.5 the unit column is populated on 15.3% of drilling rows (1,023/6,699), so a strict grade benchmark WHERE unit='g/t' is available: 81 values / 11 issuers, 65 ≥ 1 g/t, range 0.066–310.89 g/t (measured). Other populated drilling units: m (577 rows / 15 issuers — mostly hole depth/collar), % (176 / 10), ppm (176 / 7). The family is still dominated by hole-collar/survey coordinates (dip, azimuth, easting, northing, rl, depth), and a full grade×thickness intercept still needs an interval-thickness join the lake does not key cleanly — but the grade-value benchmark itself now Works:

-- Gold grade values, unit-confirmed (Works: 81 values / 11 issuers / 65 >= 1 g/t)
SELECT count(*) AS total, count(DISTINCT ticker) AS issuers,
       count(CASE WHEN value >= 1 THEN 1 END) AS ge_1,
       round(min(value),3) AS mn_gt, round(max(value),2) AS mx_gt
FROM read_parquet(
  '<LAKE>/releases/v2.5/catalogs/trusted_table/**/*.parquet',
  hive_partitioning=true)
WHERE family = 'drilling_assays'
  AND unit = 'g/t'
  AND value IS NOT NULL;
-- 81 / 11 / 65 / 0.066 / 310.89  (Works — reproducible grade-value benchmark).

-- Measure-text gold proxy as a cross-check (no unit constraint): 75 values /
-- 12 issuers / 59 >= 1; of these, 58 values / 10 issuers are confirmed g/t.
SELECT count(*) AS total, count(DISTINCT ticker) AS issuers,
       count(CASE WHEN value >= 1 THEN 1 END) AS ge_1
FROM read_parquet(
  '<LAKE>/releases/v2.5/catalogs/trusted_table/**/*.parquet',
  hive_partitioning=true)
WHERE family = 'drilling_assays'
  AND (measure ILIKE '%au%' OR measure ILIKE '%gold%')
  AND value IS NOT NULL;
-- 75 / 12 / 59  (cross-check; the unit='g/t' filter above is the stricter view).

---

Units status (v2.5)

The value column is a DOUBLE and is populated on every row, and the parsed unit columns are partially populated in v2.5 (measured against the published v2.5 catalog, 12,371 rows):

column	non-null	coverage	where it lives
unit	2,357	19.1%	jorc_resource_reserve (60.9%), tenement_schedule (19.4%), drilling_assays (15.3%); near-zero in financials
currency	2,226	18.0%	financial_statements (69.0%), appendix_5b_cashflow (63.2%); AUD 2,190, USD 36
scale	903	7.3%	appendix_5b_cashflow (32.7%), financial_statements (17.9%); '000 781, M 122
normalized_value / normalized_unit	2,846	23.0%	jorc_resource_reserve (56.4%), financial_statements/appendix_5b_cashflow (18–33%)

Top unit values (global): m 610, % 564, g/t 369, ppm 243, tonnes 165, Mt 132, koz 117, kt 85, oz 56, moz 16. This means:

• You can filter by unit/currency where populated — WHERE currency='AUD' returns 2,190 rows / 44 issuers; WHERE unit='g/t' returns 369 rows / 24 issuers (81 of them in drilling_assays across 11 issuers, 275 in jorc_resource_reserve across 14).
• You can scale-normalise where scale is present — normalized_value undoes scale (×1e3 for '000, ×1e6 for M, verified on the 903 scaled rows). The CASE scale WHEN '''000' THEN value*1e3 … pattern from prior releases works on those rows; normalized_value is the pre-computed equivalent.
• Coverage is uneven by family, so for cross-company numeric work concept_id (same economic concept) is still the most reliable surface; unit/scale filtering is a strong secondary constraint within the families that carry it (JORC grade vs tonnage, drilling grades).

Honest caveat on the deepest join key. fin.bs.current_assets.cash (27 issuers) is still not safely averageable: of its 63 cells, only 11 carry scale='000 across 5 issuers (and thus a normalized_value); the other 52 cells (24 issuers) have scale NULL, so normalized_value does not reconcile the whole concept. Across concept_id joins, normalized_value covers only the '000/M sub-print — confirm scale IS NOT NULL (or restrict to one scale) before comparing magnitudes across issuers. Values are otherwise as printed.

This is not a regression from v2.1/v2.4: the v2.5 coverage matches v2.4 exactly (the prior v2.5 cookbook over-stated the gap by describing an early broken publish; the corrected publish repopulates these columns to their v2.4 levels).

---

Worked examples by sector skill

Outcome legend — Works: ≥5 issuers reached via concept_id/identity_leaf/ unit filters. Served-below-floor: family is served and queryable but family_fidelity < 0.95 (apply the trust caveat from Querying by trust). Abstains: partial (few issuers, or the needed column is too sparse — the gap is named).

skill	Works	Served-below-floor	Abstains
producer	3	5	2
developer	1	1	1
explorer	3	0	2
cross-company	3	0	0

Producer (mining-producer)

driver	outcome	issuers	SQL / concept_id
realised price / AISC / production qty / head grade / recovery / unit cost	Served-below-floor	7	family='production_kpis' → 243 cells / 7 issuers (ff 0.5745). KPIs present: grade, tonnes mined, contained gold, mining cost, copper produced/sold/recovery, underlying EBITDA. Apply trust caveat.
reserve life	Abstains	3 reserve / 7 prod	JORC Proved/Probable tonnage present; production divisor now exists in production_kpis but reconciliation is analyst synthesis over two below-floor families.
resource conversion	Abstains	17	JORC Inferred/Indicated/Measured present (183 cells); tonnage columns are per-company/sparse.
cash on hand	Works	27	concept_id IN ('fin.bs.current_assets.cash','5b.4.5.ending_cash')
debt / net-debt	Abstains	5	concept_id IN ('fin.bs.current_liabilities.borrowings','fin.bs.non_current_liabilities.borrowings') — thin (6 cells / 5 issuers).
sustaining capex	Works	7	concept_id='5b.2.1.c.ppe' (5B payments for PP&E; 21 cells; does not split sustaining vs growth).
dividends	Abstains	0	concept_id='fin.bs.current_liabilities.dividends_payable' — below 5-issuer threshold.
5B operating CF narrative	Works	22	net operating 5b.1.9 (22), equity proceeds 5b.3.1 (13), receipts 5b.1.1 (3), admin 5b.1.2.e (7).

-- Producer cash on hand across the peer set (Works, 27 issuers)
SELECT ticker, round(max(value),0) AS value
FROM read_parquet(
  '<LAKE>/releases/v2.5/catalogs/trusted_table/**/*.parquet',
  hive_partitioning=true)
WHERE concept_id IN ('fin.bs.current_assets.cash','5b.4.5.ending_cash')
GROUP BY ticker ORDER BY value DESC;

-- 5B operating cash-flow narrative (Works; net operating reaches 22 issuers)
SELECT concept_id, count(DISTINCT ticker) AS issuers
FROM read_parquet(
  '<LAKE>/releases/v2.5/catalogs/trusted_table/**/*.parquet',
  hive_partitioning=true)
WHERE concept_id IN ('5b.1.1.receipts_from_customers','5b.1.2.e.admin_corporate',
                     '5b.1.2.d.staff_costs','5b.1.2.c.production','5b.1.2.b.development',
                     '5b.1.9.net_operating','5b.3.1.proceeds_equity')
GROUP BY 1 ORDER BY 2 DESC;

-- Producer headline KPIs (Served-below-floor; apply family_fidelity caveat)
SELECT identity_leaf, count(DISTINCT ticker) AS issuers, count(*) AS cells
FROM read_parquet(
  '<LAKE>/releases/v2.5/catalogs/trusted_table/**/*.parquet',
  hive_partitioning=true)
WHERE family = 'production_kpis' AND family_fidelity < 0.95
GROUP BY 1 ORDER BY issuers DESC, cells DESC LIMIT 15;

Developer (mining-developer)

driver	outcome	issuers	SQL / family
NPV / IRR / payback / initial capex / opex / mine life / strip ratio / recovery	Served-below-floor	7	family='study_economics' → 186 cells / 7 issuers (ff 0.6492). Headline metrics present: pre/post-tax IRR, NPV (8%), payback years, LOM sustaining capital, margin % LOM.
tenement position	Abstains	24	family='tenement_schedule' — 0 shared identity keys; per-issuer count + area only.
funding path	Works	27	cash fin.bs.current_assets.cash (27) + equity proceeds 5b.3.1 (13) + net financing 5b.3.10 (10).

-- Developer funding path (Works; cash reaches 27 issuers)
SELECT concept_id, count(DISTINCT ticker) AS issuers
FROM read_parquet(
  '<LAKE>/releases/v2.5/catalogs/trusted_table/**/*.parquet',
  hive_partitioning=true)
WHERE concept_id IN ('fin.bs.current_assets.cash','5b.3.1.proceeds_equity',
                     '5b.3.10.net_financing','5b.4.5.ending_cash')
GROUP BY 1 ORDER BY 2 DESC;

-- Developer DFS economics (Served-below-floor; apply family_fidelity caveat)
SELECT identity_leaf, count(DISTINCT ticker) AS issuers, count(*) AS cells
FROM read_parquet(
  '<LAKE>/releases/v2.5/catalogs/trusted_table/**/*.parquet',
  hive_partitioning=true)
WHERE family = 'study_economics' AND family_fidelity < 0.95
GROUP BY 1 ORDER BY issuers DESC, cells DESC LIMIT 15;

Explorer (mining-explorer)

driver	outcome	issuers	SQL
drill results (grade×thickness)	Works	11	family='drilling_assays' AND unit='g/t': 81 values / 11 issuers, 65 ≥1 g/t (80%), 34 ≥2 g/t (42%), range 0.066–310.89 g/t. Grade values reproducible; grade×thickness intercept still needs an interval-thickness join the lake does not key cleanly.
JORC resource (category split)	Works	17	JORC category labels ARE shared: total 17, inferred 10, indicated 10, measured 5 issuers (183 cells).
JORC grade/tonnage/contained metal	Abstains	19 (per-issuer)	unit now separates grade (g/t 275/14 iss) from tonnage (tonnes/Mt 292/19 iss), so rows are filterable; the gap is per-company verbose identity with no clean cross-company block join (no grade×tonnage pairing across issuers).
cash runway	Works	27	cash fin.bs.current_assets.cash (27) + burn 5b.1.2.e.admin_corporate (7) + 5b.1.2.a.exploration_evaluation_expensed (5).
tenement position	Abstains	24	per-company only (0 shared keys).

-- Explorer: JORC category split across issuers (Works; category labels are shared)
SELECT identity_leaf, count(*) AS cells, count(DISTINCT ticker) AS issuers
FROM read_parquet(
  '<LAKE>/releases/v2.5/catalogs/trusted_table/**/*.parquet',
  hive_partitioning=true)
WHERE family = 'jorc_resource_reserve'
  AND lower(identity_leaf) IN ('total','inferred','indicated','measured')
GROUP BY 1 ORDER BY issuers DESC;
-- total=17iss, inferred=10, indicated=10, measured=5 (183 cells across the 4 categories).

-- Explorer: cash runway inputs (Works; cash strong, burn partial)
SELECT 'cash'            AS k, count(DISTINCT ticker) FROM read_parquet('<LAKE>/releases/v2.5/catalogs/trusted_table/**/*.parquet', hive_partitioning=true) WHERE concept_id='fin.bs.current_assets.cash'
UNION ALL SELECT 'ending_cash_5b', count(DISTINCT ticker) FROM read_parquet('<LAKE>/releases/v2.5/catalogs/trusted_table/**/*.parquet', hive_partitioning=true) WHERE concept_id='5b.4.5.ending_cash'
UNION ALL SELECT 'burn_admin',     count(DISTINCT ticker) FROM read_parquet('<LAKE>/releases/v2.5/catalogs/trusted_table/**/*.parquet', hive_partitioning=true) WHERE concept_id='5b.1.2.e.admin_corporate'
UNION ALL SELECT 'burn_expl',      count(DISTINCT ticker) FROM read_parquet('<LAKE>/releases/v2.5/catalogs/trusted_table/**/*.parquet', hive_partitioning=true) WHERE concept_id='5b.1.2.a.exploration_evaluation_expensed';
-- cash=27, ending_cash_5b=10, admin_burn=7, exploration_burn=5.

Note: cash-runway and reserve-life ratios are analyst synthesis over joinable inputs — the lake gives you numerator and denominator, not the derived ratio.

---

Phase-8 gap list

Both Phase-8 families now ship (no longer 0 rows), but not method-certified, and several gaps remain:

-- The two below-floor families (served, NOT method-certified)
SELECT family, count(*) AS cells, count(DISTINCT ticker) AS issuers,
       round(max(family_fidelity),4) AS fidelity
FROM read_parquet(
  '<LAKE>/releases/v2.5/catalogs/trusted_table/**/*.parquet',
  hive_partitioning=true)
WHERE family_fidelity < 0.95
GROUP BY 1 ORDER BY fidelity;
-- production_kpis 243/7/0.5745 ; study_economics 186/7/0.6492

• Method-certification gap. production_kpis (0.5745) and study_economics (0.6492) are served on the per-cell consensus+judge guarantee (DECISIONS #1) but their family-level numeric fidelity is below the 0.95 floor; the combined certify decision is REJECT. Agents should down-weight or filter them via WHERE family_fidelity >= 0.95 for the highest-trust subset.
• Units are partially populated. unit 19.1% / currency 18.0% / scale 7.3% / normalized_value/normalized_unit 23.0% of 12,371 rows. Unit filtering (WHERE unit='g/t', WHERE currency='AUD') and scale-normalisation (normalized_value, or CASE scale …) work where populated; coverage is uneven by family — see Units status.
• Drilling grade benchmark (partial). With unit populated on 15.3% of drilling rows, the unit='g/t' grade-value benchmark is reproducible (81 values / 11 issuers / 65 ≥1 g/t); the full grade×thickness intercept still needs an interval-thickness join the lake does not key cleanly, and the drilling_assays family is still dominated by hole-collar/survey coordinates.
• study_economics REO-keyed converter. The REO-keyed study-economics converter remains a known defect (carry-over); the served study_economics cells are the non-REO subset.
• Rule-#1a vendor cells. No vendor cells ship in v2.5 (arm_kind='consensus' on all 12,371 rows). The rule-#1a infrastructure (arm_kind column, vendor-grounded policy) is in place but produced no cells this release; production_kpis is the family that would carry them when they arrive.

Until the method-certification gap closes (units are already partial), the lake serves balance-sheet, cash-flow, JORC-category, drilling-grade, and tenement comparison well (financials + 5B via concept_id; JORC by category split and by unit-distinguished grade/tonnage; drilling grades via unit='g/t'), serves producer/developer headline metrics below the trust floor, and treats unit-mediated compares as partial — they work within the ~19% of rows that carry a unit (and the ~7% that carry a scale), and abstain elsewhere.

Concept-id reference (the joinable surface)

The full deterministic maps live in src/asx_lake/concepts.py. Appendix 5B ids follow the ASIC item taxonomy (5b.<section>.<item>.<slug>); financials ids are fin.bs.<section>.<line> / fin.pl.<line>. Only appendix_5b_cashflow (65%) and financial_statements (32%) carry a concept_id today (1,127 rows total); every other family abstains (NULL) — by design, never invented. For the cross-company keying analysis (units, keying, the unkeyed leak), see docs/edd-cross-company-keying.md.

Release history

release	cells	families	pages	date	what changed
v2.5	12,371	7	548	2026-08-05	Schema cutover (identity_path/identity_leaf/measure/arm_kind/schema_version; identity+column dropped, hard cutover) + trust granularity (judge_confidence/family_fidelity) + bench-driven storage layout (ticker-only partition, 240→63 shards, 3.8× speedup at 13× scale). Offline re-release of v2.4 bundles; no cell added/dropped/re-judged. Units partially populated (unit 19.1% / currency 18.0% / scale 7.3% / normalized_* 23.0%) — matches v2.4 exactly.
v2.4	12,371	7	548	2026-08-04	Row-dimension Unit column propagation.
v2.3	12,371	7	548	(prior)	Phase 8: production_kpis + study_economics served.
v2.2	11,942	5	527	(prior)	normalized_value/normalized_unit added.
v2.1	11,942	5	527	(prior)	concept_id + units added.
v2	10,118	5	527	(prior)	Initial 4-arm cohort release.

Pointers

• Source repo + decisions/rules: mktmind/asx-document-lake (DECISIONS.md)
• Storage design: docs/storage-layout.md (300 lines, full bench + Iceberg/Delta trigger)
• Cross-company keying analysis: docs/edd-cross-company-keying.md
• Row-key convention: docs/row-key-convention.md
• Release pointer: pointers/v2.5.json (release sha256 be1dbca3…); lake.ref pins mktmind/asx-lake@b4e40d04…
• Release receipt (source of truth for v2.5 facts): evidence/adl-v2.5/release-receipt.json