How DrillScore works

DrillScore turns every price-sensitive ASX gold and copper announcement into the numbers that matter, with one fixed method for every company. This page explains exactly how - so you can judge the output for yourself.

1. Which announcements

Every ten minutes the DrillScore Autopilot checks for new announcements from the companies it follows. It takes only those the company itself marked price sensitive - the ones the ASX considers able to move the share price - and skips administrative notices even when they carry the flag: trading halts, reinstatements, Appendix 2A/3B/3Y forms, notices of meeting, director changes and substantial holder notices. The AI reading the document can also decide an announcement carries no market-moving information, and it's then skipped.

2. Who reads it

Each announcement PDF is read - text, tables, JORC Table 1 and any Appendix 5B (very long reports as described below) - by Claude Haiku 5.5, an AI model made by Anthropic, at a high reasoning setting. It's given the company's profile, its earlier intercepts, cash positions, raisings and past price reactions from DrillScore's database, and the market at the time: gold in US$ and A$, copper, silver, AUD/USD, the US dollar index, the US 10-year yield, oil, the XGD gold index and the GDX, GDXJ and COPX miner funds.

The model is instructed to extract numbers exactly as the company states them, to leave anything not stated blank rather than guess, and to judge disclosure against the ASX's published focus areas. It writes in plain English and gives general information only - it never says buy, sell or hold.

Long reports. To keep each reading affordable, no request may exceed 95,000 tokens - about 30 PDF pages, because the AI reads a PDF page as both text and a picture. A drill report longer than that is read in two parts: its first six pages as the PDF itself (the highlights, the long sections, the maps), and the pages after them as the PDF's own text, which costs about a third as much. That text is the characters the company printed - not a scan read by software - laid out so a table still reads as a table. The report's body comes first, then the drill-hole tables, and JORC Table 1 last; anything that still doesn't fit is named to the AI as unread, and the page says how much was read.

3. The intercepts, normalised

DrillScore's own code - not the AI - turns every interval into one comparable number:

Every intercept is checked against the PDF. Before an interval can be scored, DrillScore's code looks it up in the PDF's own text: its width and grade must be printed there together - in the same table row, or the same sentence - and near its hole ID where one is given. An interval the AI read that isn't printed that way - most often a label inside a figure's picture, which software can't check - is left out of the DrillScore, the leaderboard and the comparisons, and the announcement's page says so.

The gold bands are Stockhead's widely used rule of thumb. The copper bands are DrillScore's own. Equivalent grades (AuEq, CuEq) are rated too, but always marked as equivalents.

The smearing check: for every interval with "including" sub-intervals, DrillScore works out how much of the interval's metal-metres come from its richest one - for example, "1m at 40g/t contributes 85% of this interval's gram-metres." Above 70%, it's flagged. Why that matters.

4. The disclosure checks

These follow what the ASX itself focuses on when it reviews exploration results:

• Missing tables - a drill-hole table with locations and intercepts, and sections 1 and 2 of JORC Table 1, for exploration results on a material project.
• Metal equivalents without metallurgical recovery information (the ASX assumes 100% recovery when none is given).
• Cherry-picking - holes drilled compared with holes reported, and whether low and nil results are shown.
• Visual estimates of mineralisation without assays.
• Down-hole widths with no true-width statement, and a missing competent person statement.

Flags are graded serious, caution or note. A flag is a question worth asking, not an accusation.

5. The money

Cash, operating cash flow, exploration spend, total outgoings and the estimated quarters of funding come from the Appendix 5B (items 1 to 8.7). Capital raisings record the amount, price, discount and dilution. Fewer than 4 quarters of funding is a caution; fewer than 2 is serious. A raising adding 15% or more to the share count is flagged as heavy dilution.

6. The DrillScore

The DrillScore is only given to drill results - announcements that report new assayed intercepts. A production update, a takeover, a resource statement or a quarterly that only recaps earlier holes has no DrillScore; it shows its news signal instead (strongly negative, negative, mixed, positive or strongly positive - how the substance reads for shareholders), with the same disclosure checks. Intercepts quoted in a takeover, acquisition, raising, guidance, financial-results, resource or reserve, study or production release are shown but not scored - they are usually old holes quoted again.

For drill results, one formula, applied the same way every time:

1. The best hit. The best intercept's metal-metres on a smooth scale through the rating bands, so 210 and 480 gram-metres don't score the same. Gold: 10 g-m is 35 points, 50 is 58, 100 is 70, 200 is 80, 500 is 90, 1,500 is 98. Copper: 3 %-m is 35, 10 is 58, 30 is 70, 60 is 80, 150 is 90, 450 is 98.
2. The breadth. The average points of the best hit in each of the top five holes. One lucky hole among forty misses scores below a program where hole after hole hits. The base is 75% the best hit and 25% the breadth.
3. Deductions. 8 points for every serious flag and 4 for every caution, up to 40 points in total.
4. The band. 80+ Standout, 65-79 Strong, 50-64 Solid, 35-49 Weak, under 35 Red flags.

Every drill result also gets three pictures: the best hit on the rating scale, every hole as one block coloured by its best hit (red low, grey modest, green good to excellent, gold stellar), and a schematic section with the best holes side by side and each interval at its down-hole depth. The section is not a map - collar positions aren't in the data - so it compares holes, it doesn't place them.

The score measures the quality of the drill result and how well it's disclosed. It is not a price target and not a recommendation.

7. The price reaction

From daily closing prices DrillScore measures the move from the last close before the market could trade on the news (an announcement before 4:10pm Sydney time counts from the previous close) to the close on trading days 1, 5 and 20 - and the same move less the S&P/ASX All Ordinaries Gold index (XGD) and less gold or copper, to separate company news from sector mood. Day-1 volume is compared with the 20-day average before.

7b. The comparisons

Every drill result is ranked against everything else DrillScore has decoded (Compare):

• Against every hit - the share of other ASX gold (or copper) headline intercepts its best hit beats, over the past 12 months when there are enough.
• At its depth - the same within its depth band, by the down-hole depth where the interval starts: shallow 0-150 m (open-pit reach), mid 150-400 m, deep 400-800 m, very deep 800 m+. Deeper ore costs more to mine, so a hit is fairest judged against hits at a similar depth.
• Against the company's own record - its best hit against the average of the company's previous three drill results (the same project when there are some): improving at 1.5 times or more, fading under two-thirds.
• Against the market - the DrillScore (or, for other news, the signal) against the move relative to the ASX gold index on day 1, with day 5 and 20 alongside. A move of 2% or more either way counts. The backtest averages these by DrillScore band and shows how many results each average rests on - it means little until there are many more drill results.

7c. The historical pattern

Each announcement page shows what share prices did after past results like it (and each company page does the same for the company's latest scored or signalled announcement), in three separate strands: other results in the same DrillScore band (or other news with the same signal), the company's own earlier announcements (both as average moves against the XGD on day 1, 5 and 20), and the market it landed in - the XGD's move over the 20 trading days before the release (copper's, for a copper company), rising (+3% or more), falling (-3% or more) or steady (on a company page, the last 20 trading days up to now), with the raw moves after past results released into the same kind of market. A strand leans up or down when its day-5 average is 1% or more either way, and says "too few to say" under five results (three for the company's own record).

The strands describe the past. They are never combined into a forecast, a price target or a probability, and they are not advice: small samples swing, markets change, and no two drill results are alike.

8. The data and its limits

• Announcements: the companies' own price-sensitive ASX announcements. DrillScore links to every original.
• Market data: Yahoo Finance (delayed), updated every ten minutes.
• AI can misread a table, especially in scanned PDFs. Each analysis shows the model's own confidence, and you can report a mistake - corrected analyses are re-run.

See also the disclaimer and the glossary.