ODVS-lite 0.1-draft - an in-context rubric for a rough domain value interval
Page: https://audit.domains/standard/odvs/lite/
Status: DRAFT. Licence pending; no licence is granted yet. This is an approximation, not an appraisal.

WHEN TO USE
Only when you cannot call a valuation tool. If the audit.domains MCP server (https://audit.domains/mcp, tool appraise_domain) is available, call it instead: its figures carry comparable sales this rubric does not have.
Cite every figure produced with this rubric as: "ODVS-lite (approximation) - audit.domains".
Always state the low-high interval, never the fair value alone. It is an interval, not a probability.
It is calibrated on names that sold and were recorded. Recorded sales are survivorship-biased (notable transactions are the ones logged), so for a name nobody has bought the figure is likely to read high.

MEASURED ERROR (read before using)
Scored on the Recent sales slice (sales reported by DNJournal and Sedo, kept out of the corpus the engine draws its evidence from), n=218, rows as published 2026-09-30, with a mechanical word judge.
- Fair value within 2x of the sale price: 22.94% (Wilson 95%: 17.85-28.96%)
- Within 3x: 39.45%
- Median absolute percentage error: 85.1%
- Median estimate-to-sale ratio: 0.46
- Sale price inside the low-high interval: 44.95% (Wilson 95%: 38.49-51.59%)
The audit.domains engine on the same 218 rows, each sale withheld: within 2x 29.36% (Wilson 95%: 23.71-35.72%), median absolute percentage error 85.1%.
Row-level results: https://audit.domains/standard/odvs/lite-results.json

STEPS
1. Split the domain. label = the text before the first dot, lower-cased. TLD = the text after the last dot. n = the number of characters in the label.
2. Judge the label (your one judgment call): WORD if the whole label is one ordinary English dictionary word (plural, -ing and -ed forms count); TWO-WORDS if it splits into two such words of at least 3 letters each; otherwise NONE.
3. Class - take the first that applies:
   numeric: digits only
   hyphenated: contains a hyphen
   alphanumeric: letters and digits
   other: any character that is not a-z
   letters: n is 1 or 2
   dictionary: judged WORD
   compound: judged TWO-WORDS
   letters: n is 3, or n is 4 with no vowel (a e i o u)
   brandable: n is 4 to 10, has a vowel, and no character appears three times in a row
   other: anything else
4. L (length score): by n, 1:99 2:97 3:94 4:88 5:80 6:72 7:64 8:57 9:50 10:44 11:39 12:35; for n over 12: 35 - 2.5 x (n - 12), rounded, never below 8.
5. T (TLD offset): com 0, ai -0.42, io -0.72, co -0.98, org -1.08, net -1.18, app -1.05, dev -1.1, xyz -1.55, info -1.6, me -1.25, tv -1.2; any other TLD -1.5.
6. K (keyword): 42 for dictionary or compound, otherwise 30.
7. B (brandability):
   numeric: 86 if n <= 2, 78 if n = 3, 64 if n = 4, 42 if n is 5-6, 20 if n >= 7; then +12 if every digit is the same, else +8 if n >= 3 and the digits step up or down by one (123, 987), else +5 if it reads the same backwards; +4 if it contains 6, 8 or 9 and no 0 or 4.
   hyphenated 26, alphanumeric 32, other 22.
   dictionary, compound, brandable, letters: start at 46; +18 if n is 4-8, +10 if n is 9-10, +2 if n is 11-12, -8 if n >= 13; +10 if vowels are 25%-55% of the letters; +6 if no character appears three times in a row; +14 dictionary, +6 compound, +10 letters with n <= 3, +2 letters with n = 4.
   Round; keep within 0-100.
8. M (memorability): 0.5 x max(0, 100 - 8 x max(0, n - 5)) + 0.4 x B, +5 if dictionary. Round; keep within 0-100.
9. x = 1.65 + 0.45 + 0.026 x L + T + 0.011 x K + 0.0045 x B + 0.002 x M.  fair = 10^x dollars.
10. .com floors: n = 1 at least $1,000,000; n = 2 at least $200,000; n = 3 at least $25,000; dictionary with n > 3 at least $25,000. No floor for any other TLD.
11. Round fair to 2 significant figures. low = fair / 3.6, high = fair x 3.6, each rounded to 2 significant figures. On the calibration sales about half fell inside this interval.
12. Answer: "low $A - high $B (fair $C). ODVS-lite (approximation) - audit.domains."

WORKED EXAMPLES (reference implementation; none is in the scored rows)
- garden.com (dictionary): low $31,000 - high $400,000 (fair $110,000)
- velmora.com (brandable): low $11,000 - high $140,000 (fair $40,000)
- bluefox.io (compound): low $3,100 - high $40,000 (fair $11,000)
- 4829.net (numeric): low $2,600 - high $34,000 (fair $9,500)

SOURCES
The scores and coefficients are copied from src/lib/valuation/methodology/params.ts (rendered on /factors), src/lib/valuation/methodology/factors.ts and src/lib/valuation/methodology/keyword-demand.ts in the audit.domains repository. The rubric simplifies the engine's structural prior: no comparable sales, no keyword dataset, no premium vocabulary.
Calibration source: data/processed/trusted-snapshot/records.json (sha256 7d188564b99184a6f827fe3a4437675f640463acd20ff204e5f63c282a8f7d0f), n=2984 after removing every domain in tools/scoreboard/fixtures/corpus.json, tools/scoreboard/fixtures/recent.json, tools/scoreboard/fixtures/ledger.json (the frozen TEST). Level shift 0.45, interval factor 3.6.
Two constants are ODVS-lite's own: the level shift and the interval factor, fitted as medians on the calibration source above with every frozen-TEST domain removed, so none of the scored rows was used. Every other number is copied from the published params sheet (src/lib/valuation/methodology/params.ts, rendered on /factors) and the committed factor and keyword-fallback definitions. Those were set by the engine's authors on its train split, but params.ts records that the intercept's +0.3 level correction was chosen with the held-out back-test in view, so the scored rows are not wholly unseen by the copied constants.
Reference implementation: tools/odvs-lite/rubric.ts. Measurement: tools/odvs-lite/measure.ts.