What's new

Mr. Deeds ships quickly — this page is where returning users catch up. Newest first.

2026-09-08 — Rare T5 calculation and evidence consistency

Fixed component allowances now keep their correct market behavior through every tier and resource comparison. The site and Ask share that calculation; See the math shows the saved land, package, ordering and market-response amounts.
Rare-T5 comparable asks remain visible even when well above their own estimate. Receipt reconciliation preserves new counterparty evidence and rejects confirmed self-transfers. A broader stacked-sale land fit is saved under validation: historical checks were mixed, so its proposed land-price changes are not active. The current method and validation status are preserved with each release.

2026-09-07 — Rare T5 estimates: longer sales history

Rare tier-5 estimates now search up to a year of sales when recent matches are scarce. The calculation separates land from resources and artifacts, adjusts old receipts for their market level, and shows the dated evidence under See the math. Acid, Luster and Botanical T5 parcels have estimates again. Chaos remains outside model coverage.
These sales-guided figures follow the shared market reference between daily builds. The old borrowed T5 ranges have been removed; supporting receipts are evidence, not a prediction interval.

2026-09-07 — Valuation fixes for loaded deeds (methodology Revision 11)

Valuable resources and artifacts now survive the land-price cap. For deeds with several rare resources, the model compares complete calculations before choosing the result.
Sales-based segment adjustments use the calculation before the asking-price cap. The build checks that adding a valuable trait does not lower an estimate. What changed and how it was checked.

2026-09-06 — Coverage: no estimate for Chaos or for tier-5 rare-environment parcels, on every surface (methodology Revision 10)

Two kinds of parcel now carry no model estimate anywhere on the site or in Ask Mr. Deeds. Chaos (111 parcels, all tier 5) sits outside model coverage — collector prices, not trait math — as the oracle's rules have said since the summer and the pages had never applied: the page says "Collector piece · no model estimate" and shows the parcel's asks, recorded sales, traits and Resource Rank as facts. Tier-5 parcels of Acid, Luster and Botanical (446) show Recent ETH sales — the segment's own dated prints with each print's resources and artifact, and its live asks — instead of the range the site had shown since this morning: the model's nine held-out prints in that segment fell within 25% of it two times, in different directions by environment, and a borrowed width was not an honest number for them. Neither kind of parcel has a gap, a place in a value ranking, or a share of a portfolio total; the CSV exports say why the estimate is blank; the estimate record publishes 0 with a coverage flag, and the internal figures move to a research file the daily archive keeps for the walk-forward.
The same coverage rule is enforced everywhere the review of 6 September found a hidden figure leaking: the trait-math panel no longer ends in a total on such a page, the ask block and the last-sale line show no "vs estimate" for them (first paint and live), the compare dialog never picks one as the winner, My Assets shows a model value of covered holdings with a count of the uncovered ones, and Ask Mr. Deeds is told, in the same words, never to state or sum a figure for them.
Ranges under a level move: a range is the estimate times its calibrated multipliers, so when the live level moves the range now follows the point (low × new ÷ old) on every surface and in Ask — the old land-only split of the endpoints moved the low end the wrong way on 2,464 parcels whose low sat below their non-land component, and anchored pages did not move their range at all. Gaps on the sales page move with the same point the row shows.
Sales are identified by transaction: every row on the sales page carries its transaction hash where the ledger has one, and otherwise deed + day + currency + price, so two same-day sales at different prices are two rows (29 pairs had collapsed into one) and the live feed — which now carries its newest sale events, not only the latest sale per deed — never duplicates a row or drops the second of two sales between two polls. One 29 August accepted bid on #31472 (0.185 WETH) is struck from every measurement as an erroneous transaction, by that event alone; the row stays in the history and the parcel's 31 August ETH sale is untouched.
Corrections found by the review: the market-anchor panel names the measured ratio and the half-strength one it applies (it had shown the shrunk figure as the measurement) and no longer calls the calibrated range a print spread; a deed with a tier-1 rare deposit beside a deeper one no longer claims the tier-1 deposit "stays in the multiplier" (no separate premium is fitted for it — the label now says so); the phone Filters button lifts above the compare tray instead of covering Compare; Escape closes a menu for keyboards as well as screen readers; the "Watch specific deeds" link lands on the right field; a parcel that is not for sale keeps its last recorded sale on the page when the live feed refreshes; and Ask Mr. Deeds' instructions no longer carry the pre-Revision-8 "compute fair value from the factors" block — the record is the number, the factors explain it.
Same evening: the rule is called model coverage on every page and in Ask Mr. Deeds — the site says estimate, never valuation, and the release lint caught the phrase on the Chaos page; the methodology keeps its own title. A parcel with no estimate no longer carries the internal "few sales" mark (it put a few-sales pill beside "no model estimate" on the browse cards); the price block's freshness stamp says "prices live" on a parcel with no ask instead of "ask is live"; and the nightly's mirror-sample check reads Ask Mr. Deeds' figure in its own order — it had read one parcel's figure before the tier-order raise and reported a mismatch on the afternoon's release (#66933) that was not one. No number changed.
Later the same evening, from a second independent check of the afternoon's release: the sales page's live layer was doubling sales. The transaction identity of that release seeded only the hash, so the daemon's hashless latest-sale map re-inserted sales the page already held — 18 duplicates and 76 older sales re-listed at the top as "just sold" on the first screen — and a hash without the deed collapsed a three-deed sweep to one row. The page now reconciles every live record the way the build merges its sources: transaction + deed is the identity; a record without a hash is the same sale as a same-day row of that deed and currency within 3%; a hashed event for a hashless row gives the row its hash instead of a twin; two different hashes are two sales at any price. Also: a segment's asks on a parcel page are labelled "at the last update" and refreshed from the live book (a moved price is rewritten, a parcel no longer listed is marked); the receipts pages say why in plain words (the held-out record is in the methodology); and the mirror-sample check reports an unusable answer or an empty sample as "not proven" instead of a pass.
Night of 6–7 September, from a third independent check (the deep valuation audit): the live layer of the parcel and Koda pages was made to agree with itself. A Koda estimate now moves with the live floor on the detail, browse and value pages from one shared factor, and a feed that returns to its baseline or goes stale restores the build figure everywhere — the detail page used to keep the last scaled value on screen beside a gap computed from the build figure. On a parcel page every "vs estimate" — the last sale, the sale history — is recomputed from the estimate on screen on every update; a parcel whose figure is held at a segment's cheapest ask says so in words that stay true when the level moves or that ask changes or leaves the book (the figure is re-read at the next build). Comparable listings are chosen by the traits that drive price — plot size, then the rare resources by name and tier, then the artifact, then density — and each row says what differs. The audit's findings about the calculation order itself (a cap that can hide a parcel's resources; segment evidence measured against a capped figure) are being repaired behind switches with the backtest, not in this update.

2026-09-06 — Model: three tested changes, honest ranges, and range-only for the hardest parcels (methodology Revision 9)

Three model candidates won their backtest and shipped; three lost and did not. Each was run alone on the review's 13-window test (2,107 sales the model had not seen) against the tree that went live this morning, with the same gate declared before the runs: not worse overall, no population of fifty or more sales worse by more than two points. Tier order: a deed can no longer price below its own twin with one deposit a tier shallower (the review's first finding — 1,951 deeds were, and are raised; on the test, 47 sales moved, 59.6% → 63.8% within 25%). Gradual pooling: a named rare resource with one to four recent prints is no longer thrown back to the supply curve — its prints are blended with the curve, weighted by how many there are (50 sales moved, 58.0% → 66.0%). Anchor shrinkage: the nine segment anchors that scale plain land (Acid tier 5 was ×1.92) now apply at half strength (×1.46), the point between dropping them (which hurt the pricier prints) and keeping them whole (12 sales moved, 33.3% → 50.0%; anchored parcels 46.7% → 60.0%). Together: 102 of 2,107 sales moved, 53.9% → 61.8% on them, all sales 73.4% → 73.8%, median error 14.0% → 13.9%, no population worse. Full tables in the methodology, Table 3c. On today's book 7,252 of the 100,000 published estimates move (median +0.6%, tenth to ninetieth percentile −11% to +14%).
Not shipped, measured: admitting accepted bids at their ask-equivalent instead of the 85% rule (4 sales moved, no gain), dropping bids from component measurement entirely (2 moved, no gain), and lifting the thin-market ask cap (4 moved, every one worse — the capped parcels cleared below the model). The cap and the bid rule stay as they were, and the numbers are in the methodology.
Ranges now mean something: every range on the site is set so that 80% of the held-out sales on that calculation path and plot size fell inside it (the old rule widths covered 57%). Plain parcels read ×0.84 to ×1.46 of the estimate, artifact parcels ×0.78 to ×1.37, named-rare parcels ×0.71 to ×1.43, thin segments ×0.75 to ×2.86 — the market pays above the model more often than below it, and the ranges now say so. The typical range is about twice as wide as before. These are widths set on past sales, not guarantees; the dated archive will test them on sales that come after.
Superseded the same day by the entry above — these parcels now show their segment's recent sales and asks, and no range. Tier-5 rare-environment parcels (446 deeds; the one Club Island parcel among them keeps its island anchor) show a range and no point estimate anywhere — the deed page, the cards, the sales rows, My Assets, the CSV exports, the value ranking and Ask Mr. Deeds. Nine such sales in six months, one in nine within 25% of the model: the review's boundary case, and the point is withdrawn. (The "few comparable sales" label these pages were meant to carry since Batch 1 never rendered — a capital letter against a lowercase list; fixed, and replaced by "range only".)
Housekeeping with teeth: from the next refit every recorded sale carries its two counterparties, and any print where the operator's trading name mrdeeds.eth was a party leaves every measurement (the operator's separate bidding wallet is deliberately not on the list; the disclosures and the methodology say so). The Resource Rank order test the ranking page describes now exists as research: the ranking's order tracks the order of sale prices at ρ ≈ 0.45 on the 2,107 held-out prints, the summed index orders them better than the best single slot (0.45 vs 0.42), and across the twelve near-grid settings the ranking page names (scarcity exponent 0.75–1.25, yield step 2–4) the summed index holds ρ 0.41–0.46 and beats the single slot in every cell — filed with the batch, not yet on the page. (Corrected the same day: the first wording said "stable across the published grid" from seven of the twelve cells.)

2026-09-06 — One calculation record: the pages and Ask Mr. Deeds now quote the same file

Every build now publishes its estimate record — estimates.json: for each of the 100,000 deeds the estimate, its low and high, the part of the estimate that is not land (a Koda, an artifact, a named-resource component) and the calculation path it took. The deed pages and the browse index are baked from it, and Ask Mr. Deeds quotes the estimate, low and high from the same file instead of recomputing them. No number changed: the review's pinned harness reproduces the 5 September in-sample file byte for byte.
The intraday rescale moves only the land. When the live market level moves during the day, every surface (deed page, browse grid, sales rows, My Assets, the share card) rescales the land term and holds the non-land value fixed — the rule the oracle already used. Until today the pages scaled the whole figure, so a deed with a Koda or a rare component could read differently on its page and in an Ask answer after a level move; Club Island deeds never rescale.
Parity is now tested, with teeth: before a release swaps in, a fixture set compares the pages and the browse index with the record on every calculation path (plain, cell anchor, artifact anchor, rare component, multi-rare, thin, thin held at an ask, attested, Club Island, Koda-holding, tier-5 rare environment) — a disagreement holds the release. After the swap the oracle must load that release's record and agree with it at three level positions, or the release rolls back to the previous one automatically. Writing the fixtures found one divergence, now fixed: the oracle treated any cell with a segment statistic as a thin candidate, while the pages require the segment's supply to be under 250 parcels.
Versioned inputs: the Resource Rank extract is pinned by a manifest (v6.2, its hash, row count, ringed count and known answers) that both the pages and the oracle verify — a regenerated or truncated file cannot ship — and the Koda estimate coefficients carry a manifest naming the fitter, its settings, the exam they were graded by and the sale ledger they were fitted on; a tree whose fitter does not match its coefficients builds without Koda estimates rather than publish numbers it cannot reproduce. Details in the methodology, Section 10 (Revision 8, amended 6 September).

2026-09-05 — Model: the first changes after the independent review (methodology Revision 8)

Medium and common artifacts price by the fitted surface again — the separate "isolation" values the September review found did not earn their keep are gone from every published number. On the review's own 13-window test (2,107 sales the model had not seen), the 337 prints this touched went from 69.7% to 72.7% within 25% of the estimate; all prints 72.9% → 73.4%. Before-and-after by cohort in the methodology, Table 3b; the wins and losses by window are stated there too.
One level convention: every sale that feeds a rare-resource or artifact component is now normalised by the same print-derived market level, whatever its age. 49 test prints moved, 69.4% → 77.6% within 25%.
The thin-market label now applies to the land under an artifact: 338 more deeds carry "thin" (3,748 in all) — their land value is capped by the segment's live asks with the artifact's value added on top. None lost the label; no test print moved.
The market level for a nightly build no longer depends on enough sales landing inside 48 hours: when the live window is starved, the level comes from the same widening-window print benchmark the backtest uses — available on 176 of the last 186 days against 43 for the old rule, with the same accuracy. The build gate against unexplained level moves is unchanged, and the how-the-base-works page (linked from every deed page's base row) says which source set the level.
Visible housekeeping: "Kodas traded · 24h" counts Kodas, not sale events (a Koda sold twice counts once), on the baked and the live number alike. Behind the scenes: every newly recorded sale carries its transaction hash and two records of one sale are counted once, each build writes a model record (switches, level source, exclusions, a tier-order property test), and a dated archive of every build's inputs and published estimates has started — the true walk-forward backtest the review asked for reads from it in about a month.

2026-09-05 — Faster pages: real thumbnails, listings-first browsing, a phone filter drawer

Every deed card, sales row and comparison thumbnail loads a small WebP thumbnail instead of the full artwork — the homepage's images went from about 60 MB to under a megabyte. Deed pages keep the full artwork one click away.
The homepage and browser load the listed deeds first (a 33 KB file) and fetch the full 100,000-deed index only when a filter or search needs it.
On phones the filters live in a bottom sheet with a "Show more" card list; the sales page shows 100 rows and appends the rest on request; the nav, the search box and the map dialog work with a keyboard and a screen reader (combobox roles, focus handling, larger touch targets).
"My Otherside" is now "My Assets"; the Ask page says up front what it can answer and what it never asks for (no account, no wallet connection, no token approvals, no custody). The privacy page, reviewed 2026-09-05, lists exactly what the site keeps: anonymous product-use counts, nothing else.

2026-09-05 — Estimate wording, honest freshness, release safety

Every surface that used to say "model floor" now says estimate; deed pages in the cohorts the review flagged carry review chips ("Leans on an attested sale", "Mostly a named-resource value", …) with a one-line "why this result?" under them, linked to what the review found.
The live price strip is honest about freshness: baked numbers stay labelled by their build stamp until the live feed is confirmed fresh, and the Koda cells are judged on their own clock.
Release safety on the nightly build: two generated-output checks and a page-count gate run before the swap and refuse it; the last success and the last failure are recorded separately.

2026-09-03 — The independent review is published

The methodology reached Revision 7, aligned line by line to an independent model audit and a reproduced 13-window backtest (73% of 2,107 unseen sales within 25% of the estimate, 13.9% median error). What was reviewed, what held and what did not is in its Section 13.
Ranges are labelled indicative — a rule-based band, not a probability interval — and deed pages in the cohorts the review flagged (rare-environment tier 5, very-rare deposits, thin or ask-capped, attested, artifact-anchored) carry a caution note.

2026-08-30 — Resource Rank: plainer labels, a market check that stays a diagnostic

The rank card's "Very robust / Robust / …" chip is now Assumption sensitivity: low / low–moderate / moderate / high — same thresholds, honest name: it describes how much the rank depends on α and y, not how good the deed is. Empty deeds no longer carry a label, and "52.1st percentile" became "top 47.9%".
The Top 100 now shows each deed's near and stress ranges in the row, and its footer joined the independent-reference family (it never fed the estimate; the footer said otherwise).
The method page gained a "what current sales say" section: which (α, y) combinations order recorded sales most like the market — a ridge from (1, 1.5) to (2, 3), with the baseline on its edge. Published as a diagnostic; baseline and bands are unchanged.
Estimate cards now name what actually sets the number, largest component first — "driven by Nucleod T2 +0.36 · Rugged Hammer +0.04 (19 sales)" instead of "anchored on 19 Rugged Hammer sales" when the hammer is a tenth of the deposit. "Market-anchored" is used only when a sales-backed component is the largest driver, and the anchor drawer says which slice its prints set.
Housekeeping: the band builder now lives in the site tree (tools/resource-rank/) instead of an archive folder due for deletion.

2026-08-29 — Model refresh, first weekly review cycle, favicon

Deed model coefficients refit on the rolled 180-day window (2,131 ETH sales, effective n 635): rare tier-3 +9%, tier-4 +8%, tier-2 +5%, artifacts +4–5%, the within-rare scarcity gradient −5%. Deed pages read "coefficients dated 2026-08-29"; the sales proxy moved 0.0771 → 0.0738.
Koda estimate v2 re-ran on the nightly: level 0.663 ETH, weapon ×1.444, pendant ×1.175 — no term moved more than half a percent. Koda pages now label the ±4.5% as the backtest's historical median miss, not a confidence interval.
The Weekly Market Update for the week ending August 29: 134 Expanded sales split 34 asks taken vs 100 bids accepted, site-level estimate comparisons by subgroup, and the week's largest prints.
Methodology: the paper is now the Estimate Methodology; §7 says plainly that the 08-14 calibration table is the latest formally scored production-path set and why the 08-18 and 08-29 surfaces have not been re-scored yet. Ask Mr. Deeds copy caught up with the Koda estimate launch.
The site has a favicon. It did not, for six weeks.

2026-08-23 — Koda model estimates go live

Kodas now carry a model estimate — where, and only where, the sale record supports one: level × Mega-trait rung × KodaPendant, fitted to recorded ETH sales with a walk-forward backtest published on the methodology page.
The refusals are part of the product: Megas never get a point estimate, thin cohorts stay receipts-only, locked Kodas show why they can't be priced. 8,133 of 10,000 Kodas carry an estimate at launch.
The Koda browser gained estimate and ask-vs-estimate sorting, listed Kodas show their gap to the estimate, and each estimate rescales with the live Koda floor between nightly rebuilds.
A new Koda market page ranks every live listing against its estimate — deepest below first, re-ranked every 30 seconds — and the estimate itself now refits automatically every night, guarded by its own backtest.
Model v2, same day: carrying a Weapon joined the priced traits (~×1.44 — the strongest, stable in every month on record; it had been hiding inside the old reference class). Median holdout miss improved to 4.5% vs 5.5% for market level alone; mean miss halved to 6.7% vs 13.4%.

2026-08-23 — Learn: Kodas in depth + the Weekly Market Update

Five new Learn articles cover the koda side properly: the 99 Megas, the full trait-and-ability system, the KodaPendant, locked Kodas, and a koda buying guide with the five verification checks.
A new Weekly Market Update section: a short, dated read on the deed and Koda markets, written from this site's own recorded data each time fresh sales land.

2026-08-22 — The site now refreshes itself nightly

A full rebuild now runs automatically every night: fresh listings, koda-holding sync, and every page rebaked — the "priced" stamp should never again be days old.
The market bar's "listed below model" count went live — it now updates every 30 seconds with the rest of the bar instead of freezing between rebuilds.
Koda holdings sync daily against the live collection, so a deed whose Koda was detached shows it within a day (61 stale deeds corrected in the first pass).
All 10,000 koda pages joined the sitemap, so they are now discoverable by search engines.

2026-08-17 — Deed watchlists

Watch specific deeds: paste up to 20 deed numbers into the alert form (or hit "👁 Watch this deed" on any deed page) and get pinged the moment one lists — with a max price, or at any price.

2026-08-17 — Koda: Similar Kodas, Compare, trait intersections, methodology

Every Koda page now shows its 6 closest trait profiles — shared values, not price comps.
Koda Compare puts 2–3 trait sheets side by side with shared values highlighted.
The Koda landing gained trait-intersection facts, each one click from its proof in the browser.
The Koda market page now links a full methodology: what counts as a sale, what is excluded, what the data cannot see.

2026-08-17 — Supply audited on-chain: 1,688 locked, 8,312 tradable

Every latent Koda was checked against the contract; two had migrated since our snapshot. Listed counts now use the tradable denominator everywhere.

2026-08-17 — Richer alerts

Koda alerts can now match "any Mega Eyes / Weapon / Clothing value", and the Discord setup guide lives on the Koda alert page itself.
Alert status pages got a fix — /q/ links now resolve for everyone.

2026-08-17 — Ask Mr. Deeds covers Kodas

The AI can pull any Koda's full trait sheet, search Kodas by traits and abilities, and quote the measured Koda market summary — recorded facts only, never estimates.

2026-08-17 — My Otherside

The wallet view became an Otherside asset view: every deed plus its Kodas — standalone and still fused inside unmigrated deeds. Paste a wallet or ENS; nothing to connect.

2026-08-16 — The Koda section launches

Every Koda: traits, the 99 Megas (with the full audit of why 99, not 100), Mega-trait sharing, KodaPendant coverage, live asks and recorded sales. No scores, no estimates — by design.

2026-08-15 — Resource Rank robustness + free deed alerts

Every deed's resource rank carries a percentile and a stress-tested robustness label.
Deed alerts became free: trait profile + price, Discord or a private status page, 7 days.

Methodology changes are always documented on their own pages — deeds · Kodas · Resource Rank. Disclosures.