Revision 13 · published 2026-09-07 · amended 2026-09-08 · by Mr. Deeds · independently audited, September 2026 (Section 13) · change log · corrections policy.

The Mr. Deeds Valuation Methodology

Revision 11 · September 2026 · Model and evidence cutoff: 7 September 2026

Architecture, evidence, governance, and limits of the public Otherdeed valuation model. Incorporates the independent model audit, the corrected v3 historical simulation, the v3 look-ahead boundary test, the post-review model changes of 5 and 6 September 2026 with their pre-declared tests, calibrated ranges, and the valuation-coverage rule of Revision 10 — no published estimate for Chaos parcels or for tier-5 parcels of the three rare environments.

Contents: 1. Abstract and reader guidance · 2. Market setting · 3. Data and admission governance · 4. Model architecture · 5. Estimation and recalibration · 6. Thin segments, anchors, and caps · 7. Validation design and results · 8. Uncertainty and interpretation · 9. What the estimate claims · 10. Governance and change management · 11. Disclosure policy · 12. Limitations and conflicts · 13. Independent review and v3 confirmation · 14. Status of this document · Appendix A. Evidence and reproducibility record · Appendix B. Glossary

How to read this paper

Sections 3–6 describe the production logic. Section 7 separates reproduced evidence from internal experiments. Sections 8 and 12 state where the estimate and displayed band should not be used as decision-grade outputs.

1. Abstract and reader guidance

Mr. Deeds publishes an estimated fair value in ETH for approximately 100,000 Otherdeed land parcels. The model estimates the center of observed clearing behavior; it is a market-description system, not a rarity score and not a claim about intrinsic value. It combines a transaction-level benchmark, a multiplicative hedonic surface for pervasive traits, and additive or anchored treatments for scarce items and thin segments.

The corrected v3 historical simulation covers 2,107 clean ETH prints across thirteen fortnightly as-of dates from 8 March through 23 August 2026. In that partial simulation, 73% of prints fell within 25% of the earlier estimate and the median error relative to the estimate (the measure Table 3 labels MdAPE; Appendix B) was 13.9%. Plain parcels performed better than scarce tails: 76% of 1,527 plain-parcel prints were within 25%, compared with 63% of 248 rare-resource prints and 11% of nine rare-environment tier-5 prints.

Revision 8 (5 September 2026) records the first model changes shipped after the review: medium and common artifacts price by the fitted surface again (§4.4), component measurement uses one level convention (§5), the thin-segment rule is applied to the land term (§6), and a live build takes its level from the print-derived benchmark when the daemon’s window is starved (§3.2). Measured on the same 2,107 prints against the production tree as it stood before these changes (commit 5cc0265, live as b331), the share within 25% moved from 72.9% to 73.4% and the median error relative to the estimate stayed at 14.0%; on the 339 prints the changes actually moved, 69.6% to 72.6% and 16.2% to 15.4% (Table 3b). Sale records now carry a transaction identity (§3.1), every build writes a model record and a tier-order property test (§4.2, §10), and a dated input archive captures every build from this release (§10). The amendment of 6 September 2026 records the calculation record that the parcel pages and the companion oracle now share, the release-time parity test with automatic rollback, and the versioned manifests for the Resource Rank extract and the Koda coefficients (§10); it changes no published number. Revision 10 (6 September 2026, the same afternoon) records the findings of the release review of that day and the corrections made for them: a valuation-coverage rule (§8.2) under which the 111 Chaos parcels — a collector segment the model never described — and the 446 tier-5 parcels of Acid, Luster and Botanical carry no published estimate and no range on any surface (the latter show their segment’s own recent ETH sales and asks in place of the range Revision 9 had shown for a few hours); the rule that a range follows its point under a level move (§10); sale identity by transaction on the public sales pages (§3.1); one struck event (§3.3); and corrections to the explanation copy — the anchor panel, the mixed-tier label, the order-test wording. No estimate, range or model rule changed for any covered parcel: the pinned harness file differs from Revision 9’s only in the struck flag of one row and one corrected label (Appendix A).

Revision 9 (6 September 2026) records the second set of model changes, each run alone as a candidate on the same 2,107 prints against the tree that carried Revision 8, under the same pre-declared gate, and adopted only when it won: the depth-preserving tier order (§4.2), gradual pooling for named rare-resource components (§4.3), and cell anchors applied at half strength (§6). Together they move 102 of the 2,107 simulated prints, from 53.9% to 61.8% within 25% on those prints and from 73.4% to 73.8% on all of them, with no population of fifty or more prints worse (Table 3c). Three candidates were measured and not adopted — bid-hits admitted at their ask-equivalent, bid-hits excluded, and the thin-segment ask cap lifted — and their numbers are stated (§7.1). The displayed ranges are now calibrated to a stated coverage target on the chronological residuals (§8), the counterparty rule for operator prints is in place (§3.3, §11), and the Resource Rank order test exists as research (§10).

Those results are useful but limited. The backtest reconstructs date-sensitive premiums and withholds post-as-of sale dates, while retaining the current fitted coefficients and current listing-derived structures. It therefore validates the layers above the regression more directly than the full product. Sales have a calendar day but no transaction time, so the test is end-of-day rather than intraday.

Reader guidance

The independent review supports using the estimate as a rough comparison point for ordinary, liquid parcels. It does not support using it as the deciding bid, ask, or reserve for rare-environment tier-5 parcels, very-rare deposits, thin or ask-capped parcels, operator-attested parcels, or parcels dominated by an artifact anchor or named rare-resource component. The displayed band is an indicative rule-based range, not a probability interval.

2. Market setting

Three features of the market shape the model.

Thin and heterogeneous. A six-month window contains only a few thousand admissible observations spread across 100,000 distinct assets and many sparse trait combinations. The model must pool evidence across parcels, and it must disclose when a value rests on a direct print, a segment-level estimate, or structural inference.

Asks and clearing prices are different series. The lowest visible listing can move without a transaction. The model uses realized ETH prints to set the transaction benchmark and treats listing information as a separate market fact. This choice makes the model slower to recognize a genuine repricing, but less sensitive to unsupported ask movements.

Venue and settlement matter. ETH ask-lifts, WETH bid-hits, aggregator prints, and private transactions can represent different execution regimes. The production system tags them separately. WETH observations do not set the core market level, and the review recommends removing the current one-sided WETH screen from component estimation until it is rebuilt as a signed, held-out venue effect.

The central design problem is therefore evidence allocation: decide which observations can set the market level, which traits can borrow strength from the pooled fit, and which scarce cohorts require explicit caution because the tape is too thin for a precise point estimate.

3. Data and admission governance

3.1 Sources and time resolution

Sales enter through a live marketplace feed, periodic secondary-market reconciliation, and a rolling repair pass. The reviewed dataset contained 3,571 sale rows. The v3 partial historical evaluation scored 2,107 clean ETH prints across its forward windows; a separate in-sample diagnostic used 2,112 clean ETH rows. Source rows are retained, while model eligibility is applied through filters and admission records.

Table 1. Input classes and their role

Each sale carries a day key, not a transaction timestamp or block. The system can make a coherent end-of-day cutoff, but it cannot establish whether a same-day observation occurred before or after an intraday model snapshot. Any future intraday claim requires event timestamps and an event-time cutoff.

Sale identity (5 September 2026). Each sale record now carries the transaction hash where its source provides one and the event time where it is known; older rows keep only a day. Two records are the same sale when their hashes match or, where either side has no hash, when they fall on the same parcel-day within a small price tolerance (two venues can book one sale with slightly different price accounting); two records with different hashes are two sales at any price. Since Revision 10 the public sales page keys every row the same way — by transaction hash and parcel where the ledger has one, otherwise by parcel, day, currency and price — so two same-day sales at different prices are two rows (twenty-nine pairs had been shown as one) and two parcels sold in one transaction are two rows, and the live feed carries its newest sale events beside the latest-sale map so that two sales between two polls both reach the page. Every live record is reconciled against the rows the page holds by the rule above: a record without a hash is the same sale as a same-day row of that parcel and currency within the tolerance, a hashed event for a hashless row gives that row its hash rather than a twin, and two different hashes are two sales. (Amended the evening of 6 September 2026: the first Revision 10 release seeded the page’s index with the hash alone, so the hashless latest-sale map re-inserted sales the page already held — eighteen duplicates and seventy-six older sales re-listed as new on the first screen — and a hash without the parcel collapsed a three-parcel sweep to one row; the second independent check of that day found it, and it was corrected the same evening.) The live feed keeps every sale event in its window rather than the latest sale per parcel, so a parcel that trades twice in a day is two events and one asset, and the site’s “traded” counts count assets. The rule is applied wherever the ledger and the live feed are merged: the benchmark, the component measurements, the market checks and the historical simulation, which it left byte-identical.

3.2 Transaction benchmark

The day-level benchmark is reconstructed from realized clean ETH prints. For each day, the model takes the tenth percentile of eligible prices in a ±2-day window, widening through ±4, ±7, ±10, and ±14 days as needed until the window contains at least twenty prints. This replaced a recorded floor series that had frozen from early May into June while clearing prices fell, causing historical May observations to enter downstream calculations at the wrong level.

For a backtest as-of date, v3 removes every sale dated after that day before it constructs the starting benchmark and related level inputs. Earlier dates inside the historical training series still use a symmetric local window bounded by the overall as-of date. That is acceptable as retrospective normalization of the training set; it is not a daily index that could have been published causally on every earlier day.

Level for a live build (5 September 2026). The benchmark that rebases a published build is taken, in order, from: the price daemon’s confirmed level (the tenth percentile of plain ETH prints in its trailing 48 hours, when that window holds its minimum count); failing that, the print-derived estimator above computed on the merged tape of the ledger, the off-venue additions and the daemon’s sale events, its window widened until it holds twenty prints; failing that, the daemon’s last recorded day level while it is recent; failing that, the level frozen at the last refit. Before 5 September the second step did not exist, so on most days the published level rode the weekly refit and the release gate refused whenever a fed window emptied. Evaluated offline on the 186 days to 2 September, the 48-hour rule was available on 43 days and the print-derived series on 176, with the same accuracy against the following three days of prints (median |log(print ÷ level)| 0.131 against 0.134) and smaller day-to-day steps (90th percentile 6.8% against 9.4%). The release gate is unchanged: a level move beyond its tolerance refuses the build unless the level comes from a fed window of the market’s own prints, and a larger move needs a human. The “how the base works” page linked from every parcel page states which source set the level of that build.

3.3 Exclusions and manual review

A print enters the model only when it plausibly represents an arm’s-length opinion about the specific parcel. The principal categories are:

• Distress screening uses two level references. The site screens prints against the corrected print-derived benchmark (a fire-sale cut at 55% of the day level, with ratio trimming inside anchor measurement). The fitter screens against its own reconstruction of the production floor estimator — a 48-hour plain-ETH tenth percentile rolled forward on thin days, 126 of 180 days in the current fit — which in the reviewed history froze at the same value over the same 5 May–6 June stretch as the recorded field. Moving the fitter’s screen to the print-derived benchmark is scheduled (refit runbook step A2).

• Settlement-currency separation so WETH bid-hits do not set the core ETH market level.

• Basket and sweep exclusions when identical prices across unlike parcels indicate package economics rather than parcel-level prices.

• Structural exclusions for parcels whose price is dominated by an attached asset or a special-purpose segment.

• A recorded human review for expensive or context-dependent observations that mechanical low-price filters cannot identify.

Quick flips are tagged in the display. Contrary to an earlier methodology statement, the reviewed fitter did not down-weight them. Revision 7 states the implemented behavior and treats any future flip weighting as a proposed change that must be tested before release.

Operator prints (Revision 9). From the 7 September 2026 refit every recorded sale carries its two counterparties. A print in which the operator’s trading name, mrdeeds.eth, was the buyer or the seller is struck from every measurement — the level series, the component and anchor measurements, the drift checks and the historical simulation — while the public sales pages keep showing it as a fact. The list holds that one name by decision; the operator’s separate bidding wallet, which places standing bids across the collection, is not on it, and the operator states this here rather than leave it implicit. The ledger rows recorded before 7 September carry no counterparties and cannot be tested; the rule is inert on them. Struck events (Revision 10). A single recorded event can be struck from every measurement by its identity — parcel, day, currency and amount, reconciled to the transaction hash where the ledger carries one — without touching any other sale of the parcel; the raw row stays in the public history. One such strike is in force: the accepted-bid print of 29 August 2026 on parcel #31472 (0.185 WETH), an erroneous transaction identified by the operator; the parcel’s ETH sale of 31 August is untouched. Struck events are listed in the model record.

3.4 Operator-attested admissions

An attested admission is a dated exception for a transaction the operator represents as an arm’s-length negotiated trade or as a real trade absent from the ordinary feeds. In the pooled fit, an attested observation carries the weight of one ordinary observation. It may have greater influence in parcel or scarce-component layers: the current system can anchor the sold parcel to its level-adjusted print and can use a very-rare attested print to set a component applied to other carriers. These are judgment-based overrides and require stronger public evidence than ordinary feed observations.

The reviewed admissions identify deeds #75181, #6362, #79234, and #7258. The operator states that neither the operator nor any wallet or account they control was a buyer, seller, bidder, broker, or otherwise financially involved in those four prints. The independent review accepted that statement as a response to the conflict question, while noting that the package still lacked transaction hashes, complete counterparty evidence, and a third-party verification record.

4. Model architecture

The production estimate can be summarized as a land term plus item components:

Vᵢ = Lᵢ + Σₖ cᵢₖ  where  Lᵢ = F × β₀ × Πⱼ mⱼ(i)

F is the displayed collection floor, β₀ is a rebased intercept, mⱼ(i) are multiplicative trait adjustments, and cᵢₖ are fixed-ETH components for selected scarce items. Anchors, thin rules, live-ask substitutions, and attested blends can modify this basic path. The sequence matters; the public method should therefore disclose the actual trace used for each parcel, not only the formula topology.

4.1 Transaction level and floor cancellation

The pooled fit models price relative to the day’s transaction benchmark. At publication time the intercept is rebased against the displayed floor. Algebraically, the floor cancels from the plain-parcel level, so an unsupported ask-floor move should not change a plain estimate until realized prints move the transaction benchmark. Listing information can still affect later caps and substitutions.

V(plain) = F × [fitted ratio × benchmark ÷ F] = fitted ratio × benchmark

4.2 Multiplicative hedonic surface

Pervasive traits enter as jointly estimated multiplicative effects on the land term. They include environment and tier, sediment, resource-slot count, deposit density, composition, and selected interactions. Joint estimation is intended to separate partial effects where traits are correlated. Fit-time monotonicity constraints prevent selected density, composition, rarity, and tier ladders from decreasing before later layers are applied.

The independent review found that monotonicity is not guaranteed after the full stack. A tier-1 very-rare path can exceed tier 2, and rare-artifact multipliers can stack with tier-5 land and additive resource values without enough joint tail evidence. The production trace and final value therefore need post-stack property tests, not only monotone fitted coefficients.

Since 5 September 2026 every build runs one: 7,237 twin pairs (every twentieth parcel against a copy with one deposit a tier deeper) are priced through the full stack and the share of pairs where the deeper deposit is worth less is recorded at each stage — 1.7% on the fitted surface and 2.1% after the layers in the Revision 8 build, concentrated in common tier-2→3 (5.1% of 2,383 pairs) and rare tier-2→3 (5.4% of 350). Since Revision 9 the order is enforced at the final figure: a parcel never prices below any twin of itself with one deposit a tier shallower, recursively along the tier lattice, and when it would, the published figure is raised to the highest shallower twin and the parcel page names the twin. On the 6 September book 1,951 of the 100,000 parcels are raised (median +0.6%); on the historical simulation the rule moved 47 prints, from 59.6% to 63.8% within 25%, and the twelve very-rare prints from 50.0% to 58.3% (Table 3c). The test still records the inversion rate on the fitted surface (1.7%) and after the layers (2.0%) — those are the defects a refit should remove; the final-figure rate is now 0% by construction.

4.3 Rare-resource components

Selected named rare resources enter as fixed-ETH components measured as the residual between a carrier print and the estimated value of the parcel without that component. Recency weighting, robust trimming, pooled class values, and supply-value curves allocate authority when direct observations are scarce. This is a design choice: it treats the scarce item as a separable collectible rather than a percentage uplift on unrelated land traits.

The tier rule as implemented (stated plainly here since 5 September 2026): deposit tiers 2 and 3 of a rare resource price additively; a tier-1 deposit of a rare resource on a parcel that carries no deeper rare deposit prices multiplicatively; and a tier-1 rare deposit on a parcel that also carries a tier-2 or tier-3 rare deposit currently contributes no separate premium on the additive path. Earlier revisions described that last case as a hybrid that kept a distinct tier-1 contribution; the audit found (deed #75181) that no such contribution exists, and an independent population test confirmed it across all 154 non-attested parcels on that path. Whether a validated tier-1 contribution should be added is a scheduled candidate, to be tested against the historical simulation before release. Very-rare resources measure from one print upward and shrink toward a supply-curve reference. Until Revision 9 a mid-rare name needed five kept prints before its own prints counted at all — below that it was priced from the supply curve alone, a cliff the review named (finding 9). Since Revision 9 a mid-rare component with one to four kept prints is shrunk toward its supply-curve reference with weight n ÷ (n + 4) — one print counts for a fifth, four for half — and the parcel page says how many prints stand behind it; the five-print rule and the pool shrinkage above it are unchanged. On the historical simulation the change moved 50 prints, from 58.0% to 66.0% within 25% (rare-component parcels 58.6% → 62.2% alone, 64.0% with the other Revision 9 changes). The cases the review reproduced where a deeper deposit lowered the estimate are now caught by the tier-order rule of §4.2; the representation defect behind them (a tier-2 factor dropping with nothing replacing it) remains a refit item.

Ablation results support a split policy. The supply curve carries useful signal for very-rare parcels: removing it materially worsened that cohort. Named components were neutral to harmful for mid-rare parcels. The next design should preserve and recalibrate the very-rare prior while shrinking or removing mid-rare named components unless a predeclared held-out test supports them.

4.4 Artifact treatment

Until 5 September 2026 the production model used additive isolation values for medium and common artifacts and multiplicative or anchored treatment for selected rare-artifact cohorts. The separation is economically plausible, but the medium/common isolation layer did not earn its complexity in the reproduced ablation: on 339 affected prints, removing it improved the share within 25% from 69.0% to 72.3% and reduced the median error relative to the estimate from 16.2% to 15.4%. Revision 7 labelled it as scheduled for reversion or evidence-weighted shrinkage.

Reverted 5 September 2026. Medium and common artifacts now price by the fitted artifact factors; rare and ultra artifacts keep their thin-segment anchors, which are measured, not fitted. The pre-declared test repeated the ablation on the production tree as it stood before the change (commit 5cc0265): on the 337 prints the layer touched, 69.7% to 72.7% within 25% and 16.1% to 15.4% median error; all 2,107 prints 72.9% to 73.4%. By as-of window the reversion was better in six of the thirteen windows, unchanged in three and worse in four, the largest loss three prints of 109 in the 31 May window, where artifact parcels were clearing below the fitted factors in early June and the lower isolation values had happened to sit closer. The isolation components remain in the code behind a measurement switch and are not used by any published number.

4.5 One-way market rules

Several post-fit rules move the estimate in only one direction. Live asks cap or substitute downward. The WETH component screen admits observations only when they clear a model-relative threshold, allowing positive evidence without a symmetric negative counterpart. Multi-rare isolation is used only when it raises the estimate above the multiplicative path. These rules may express defensible product priors, but they also limit the claim that the point estimate is a neutral description of clearing behavior. Each should be evaluated on later data with signed effects.

5. Estimation and recalibration

Core coefficients are estimated on log price ratios over a rolling 180-day window after the data-governance rules are applied. Observations receive recency weights. Correlated feature families are fitted jointly, and selected monotonicity constraints are applied after fitting. The current model has enough observations to support pooled effects, but many individual intersections remain too sparse for unrestricted cell estimates.

Density and composition now sit inside the fitted surface rather than in a separate bounded recalibration patch. This simplifies the displayed trace and forces shape changes to wait for a refit. Level changes continue between refits through the transaction-benchmark rebase.

Until 5 September 2026 component measurement used raw prices for very recent observations and day-normalized prices for older ones. That branch was introduced to avoid mixing the former ask-derived level with the corrected print-derived level, but it created a discontinuity around the recency boundary; Revision 7 described it as a normalization heuristic. Since 5 September every component observation is normalized by the print-derived benchmark, whatever its age — one convention. The pre-declared test moved 49 of the 2,107 simulated prints: 69.4% to 77.6% within 25% and 15.2% to 12.9% median error on those rows, nothing else touched. The core surface fit still has its own level reconstruction (§3.3) and is scheduled to move to the same benchmark at the next refit.

Any refit or layer change should be evaluated against the same frozen input snapshot and forward windows. Corrections required for causal integrity or false documentation may ship without an accuracy gain, but their release note should distinguish a correctness fix from a model-improvement claim.

6. Thin segments, anchors, and caps

A market anchor substitutes a segment’s observed clearing level when its prints meet the applicable evidence gate and disagree with the generalized surface. Once admitted, an artifact anchor is applied at full strength — it is the artifact’s own measured market. Since Revision 9 a cell anchor (the ratio that scales plain land in an environment × tier segment) is applied at half strength: ratio′ = 1 + (ratio − 1) × 0.5, so Acid tier 5 went from ×1.92 to ×1.46. The candidates measured were full strength (the Revision 8 position), three-quarters, half, a quarter and none: dropping the anchors improved the fifteen anchored prints most (46.7% → 60.0% within 25%) but cost the 114 prints of 0.3 ETH or more 2.6 points and the rare-environment tier-3 cohort 8.3; a quarter kept most of the gain and still cost tier 3 4.2 points; half took the anchored gain in full (46.7% → 60.0%, median error 30.9% → 21.4%) with no cohort worse, and shipped. Shrinkage toward a tier pool or supply-curve reference applies to rare-resource components as before. The affected parcel page should name the segment, the observation count, the anchor date, and whether the value was measured, pooled, inferred, or capped.

The ordinary anchor rule requires at least three qualifying prints. Two prints that agree within 40% qualify any segment, including a cell; the current Luster tier-4 cell anchor is based on two agreeing prints. Two prints that disagree by more than 40% qualify only an artifact anchor, only when both sit above the model band, and the lower print is used. Cells never anchor on disagreeing prints.

A thin gate marks a sparse estimate as a floor or substitutes a qualifying live ask. The September correction applies a thin-segment ask cap to the land term and then adds a qualifying rare-resource component, rather than allowing a plain parcel’s ask to cap the entire rare-resource carrier. The design still has two unresolved limitations: current asks are reused in historical windows, and a downward-only substitution can distort a neutral point estimate. Lifting the cap and showing the ask as evidence only was measured as a Revision 9 candidate: four simulated prints moved, every one for the worse (the capped parcels had cleared below the model, at a median of 0.79 × the estimate), so the cap stays and the page keeps saying when it holds a figure. The candidate remains coded behind a switch and will be re-measured when more capped prints exist.

Across the v3 backtest, 62 scored prints carried the thin-gated label; 48% were within 25% and MdAPE was 25.2%. Only three were classified as thin-capped, so that sub-result is too small to generalize. The thin label is therefore more valuable as a warning about evidence than as proof of predictive improvement.

Thin rule on the land term (5 September 2026). Before this date a parcel with any market anchor was exempt from the thin-segment rule, so a parcel whose land segment was thin but which carried an artifact anchor was published without the thin label or the ask cap on its land. The rule now applies to the land term: a cell anchor (which prices the land itself) still exempts the parcel, an artifact anchor does not, and the artifact’s additive value rides on top of the thin land value and its range. The change moved no simulated print (four more of the 2,107 carry the thin label, 66 against 62) and 338 more of the published parcels carry it (3,748 against 3,410), none lose it. It is a labelling and evidence correction, not an accuracy claim.

Revision 11: calculation order

From 7 September 2026, a thin-segment asking-price cap applies to the land term after dividing out the retained rare-resource and artifact factors. Those factors and any additive components rejoin after the cap. A multi-rare or hybrid deed's multiplicative and component forms both pass through the same anchor and cap stages before the higher finished value is selected. Segment residuals use the value after components and before the cap and attested-deed adjustment; a present asking price no longer supplies that denominator.

The September 7 replay reproduced every published point of b340 before the change. The adopted calculation changed 157 of 99,443 covered figures (85 higher, 72 lower). It removed all 52 current structural failures across 254,708 comparisons and passed the original audit's 51 examples. Controlled low-ask fixtures exercise the cap including synthetic asks below current listings. These checks establish calculation consistency, not new sales evidence for the resulting collector prices. The broader resource-evidence and scarce-T5 studies remain open.

An independent rerun on the September 7 inputs kept the 13-window within-25% rate at 73.8% over 2,107 clean prints; nine historical estimates changed. Five current caps still bind after the correction. Its historical windows do not reconstruct the current asking-price book, so this result does not validate the newly uncapped collector levels. The rare-environment T5 median error worsened on nine observations, including unpublished receipts-only figures; no coverage restriction is relaxed. Existing range multipliers are retained as indicative widths. The previous in-sample file identity changes because the calculation order changes; old identities below remain historical records.

7. Validation design and results

The review used several instruments because no single score can answer every question. The central publication claim comes from the corrected v3 historical simulation. Other instruments diagnose cohorts, layers, presentation, and reproducibility.

Table 2. Validation instruments and scope

7.1 V3 historical simulation

For each as-of date, the v3 harness constructs date-sensitive model layers using sales dated on or before the end of that UTC day, then scores eligible ETH prints dated one to fourteen days later. The evaluation clock is pinned. Same-day sales are deliberately available to an end-of-day snapshot.

The independent rerun reproduced all thirteen backtest files and the in-sample file byte for byte. A boundary test injected forty extreme 0.000001 ETH prints dated one day after the 12 July as-of date. V3 produced the same 357 rows and the same SHA-256 as its clean baseline; v2 changed. This confirms that post-as-of sales no longer contaminate the v3 starting level. It does not establish intraday causality, causal daily normalization for every earlier training date, or a full historical refit.

Table 3. V3 partial historical results

Source: v3 reproduced backtest summary. “Very-rare kind” is the harness classifier and is not the same population as every parcel carrying a resource with supply ≤250. The column headed MdAPE is the median of |print ÷ estimate − 1| — the error relative to the estimate; every within-25% figure in this paper uses the same ratio. Since 5 September 2026 the harness reports that measure as MdAE/est and reserves the name MdAPE for the sale-denominated median of |estimate ÷ print − 1| (the textbook definition and the one the fitter’s own holdout score uses); Table 3b shows both.

Performance varied sharply by market regime. The share within 25% ranged from 46% in the 19 April window to 90% in the 8 March and 9 August windows. The 19 April window had a median print-to-estimate ratio of 1.26 and MdAPE of 27.4%, showing that the pooled headline hides large time variation.

Table 3b. The Revision 8 changes — the same 2,107 prints, each model as it would have stood on the as-of day; “before” is the production tree before the changes (commit 5cc0265, live as b331)

Source: the operator’s 5 September run of the same thirteen-window harness on the pre-change tree and on the release tree; the harness’s in-sample file was byte-identical to the 5 September reference before the model switches were turned. The gate declared before the run was: not worse overall, and no population of fifty or more prints worse by more than two points within 25%. Both held. The pre-change baseline differs from Table 3 in the second decimal (the tree and the ledger moved after the 3 September reproduction); the very-rare row rests on twelve prints and one print is 8.3 points. Sale-denominated error rose on artifact parcels while the estimate-denominated error and the within-25% share improved — the reversion raised estimates that had sat below their prints, and the two denominators weigh that differently.

Table 3c. The Revision 9 changes — the same 2,107 prints; “before” is the Revision 8 tree as it went live on 6 September (build b334), “after” the three adopted candidates together

Source: the operator’s 6 September runs of the same thirteen-window harness — each candidate alone against the Revision 8 tree, then the three adopted candidates together; the harness’s in-sample file was byte-identical to the Revision 8 file before the switches were turned (Appendix A). The gate declared before the runs was the Revision 8 gate: not worse overall, and no population of fifty or more prints worse by more than two points within 25%. It held for each adopted candidate and for the combination. Per candidate: tier order moved 47 prints (59.6% → 63.8%), gradual pooling 50 (58.0% → 66.0%), anchors at half strength 12 (33.3% → 50.0%). Not adopted: bid-hits admitted at their ask-equivalent — the print divided by the median bid discount measured on plain parcels in the same window — moved four prints with no gain (median error on them 6.1% → 10.9%); bid-hits excluded from component measurement moved two with no gain; the thin-segment cap lifted moved four, all worse (25% → 0% within 25%). The anchored and very-rare rows rest on fifteen and twelve prints; the tier-5 rare-environment row on nine.

7.2 What changed from v2

The v3 level-series correction raised the pooled within-25% rate from 71% to 73% and reduced MdAPE from 14.5% to 13.9%. The 12 July window rose from 71% to 89% within 25%. Because the correction removed future sale dates from the as-of level rather than adding a tuned premium, it should be treated primarily as an integrity fix; the accuracy gain is supporting evidence.

7.3 Layer evidence and open dispositions

Table 4. Principal review findings

7.4 Challenger evidence

The operator’s review program also compared the structured model with six internal challengers, including segment medians, comparable-sale matching, gradient-boosted trees, a blend, and a fresh log-linear fit. In the 12-window comparison, scored on the same 1,992 clean ETH prints for every model, the incumbent scored 72% within 25% against 62% for the best challenger, led every challenger on that measure in all twelve windows (nine of twelve on median absolute error), and held a gap of forty-one points on the very-rare cohort (58% against 17%, twelve prints), seventeen on premium parcels and nine on rare-resource carriers overall. Those experiments support keeping structured pooling as the production baseline and using machine-learning disagreement as a discovery tool. They do not prove that the present stack is globally optimal, and Revision 7 does not use them to override the independently reproduced cohort weaknesses.

8. Uncertainty and interpretation

Until Revision 9 the displayed low and high values were generated by fixed model rules that widened in selected thin or special cases. Since Revision 9 they are calibrated widths: for each calculation path (plain, artifact, named rare component, several components, cell anchor, artifact anchor, thin segment, tier-5 rare environment) and environment tier, the low and high multipliers are the 10th and 90th percentiles of print ÷ estimate on the chronological residuals of the thirteen-window simulation — the 2,107 clean ETH prints each scored by the model as it stood on its as-of day — so that 80% of those prints fall inside; a cell with fewer than forty prints borrows its path’s pooled quantiles, a path with fewer than forty borrows every print’s. On the source prints the calibrated widths cover 79.8% against 56.7% for the rules they replace; the typical range is about twice as wide. The multipliers are right-skewed — plain parcels ×0.84 to ×1.46, artifact parcels ×0.78 to ×1.37, named-rare parcels ×0.71 to ×1.43, thin segments ×0.75 to ×2.86 — because the market pays above the model more often than below it. They are still not confidence, credible, or prediction intervals: they were set on the residuals they are measured on (in-sample for the calibration, out-of-sample for the model), the calibration file carries the sample size of every cell and its date, and the dated archive walk-forward is where the target is tested on prints the calibration never saw. Tier-5 rare-environment parcels carry no published range or point at all since Revision 10 (§8.2): the widths available for them were borrowed from other paths, and the segment’s nine held-out prints did not support a centre. The independent review’s out-of-sample range analysis found that the realized print fell inside the displayed band about 50% of the time overall, 47% for anchored parcels, and 14% for anchored parcels with a rare component. A separate in-sample, level-adjusted diagnostic found 70.5% overall coverage, 55.0% for anchors, and 33.3% for both rare-environment tier 5 and the twelve-row very-rare classifier. These rates come from different populations and level conventions and should not be combined as one statistic. They agree that coverage varies sharply by cohort.

Meaning of “indicative range”

The band is a descriptive width around the model’s center. It does not state a probability that a future sale will fall inside it. Readers should use cohort sample size, operative layers, and recent market conditions alongside the point estimate.

The planned interval method is empirical: state a target coverage, estimate held-out residual distributions by final operative layer and market state, include uncertainty from anchors and additive components, and publish the sample size and calibration date. Cohorts without enough evidence should show a broad descriptive range or no precise point estimate.

The out-of-sample result—about 50% overall coverage—described the rule widths that Revision 9 retired; it is kept here as the record of why they were replaced. The calibrated widths above are the current “indicative range”.

8.2 Valuation coverage (Revision 10)

The model publishes an estimate only where it describes the market. Two kinds of parcel are outside that coverage on every surface — the parcel page, the cards, the comparison view, the sales rows, the portfolio view, the exports, the value ranking and the companion oracle — and carry no point, no range, no gap against an ask or a sale, no place in a ranking by value and no share of any total.

Collector pieces. The 111 Chaos parcels (all tier 5) trade at collector prices that trait math does not describe. The operator had excluded them from the oracle’s coverage since the summer; the pages had carried figures for them, and Revision 10 applies the exclusion there too. Their asks, recorded sales, traits and Resource Rank are shown as facts; the page says “Collector piece · no model estimate”.

Receipts-only parcels. The 446 tier-5 parcels of Acid, Luster and Botanical (a Club Island parcel among them keeps its island anchor). The model’s nine held-out prints in this segment fell within 25% of it twice, in different directions by environment — level-adjusted, the two plain Acid prints sat 30–60% above the model and the two plain Botanical prints two to three times below it — and the calibrated widths available for them were borrowed from other calculation paths, so Revision 9’s range-only display was withdrawn the same day it appeared. Their pages show the segment’s own recent ETH sales — dated, linked, with each print’s named resources and artifact — and its live asks, labelled as receipts, and no model figure; a segment with no such sale says so. Obsidian, Shadow and Sludge tier-5 parcels keep their estimates with the few-sales qualification while their own cases are reviewed.

The internal trait-math figures for both kinds are still computed: the historical simulation and the tier-order property test keep scoring them, and each build writes them to a research file (estimates-research.json, not served) that the dated archive keeps for the walk-forward. The public estimate record marks the rows with a coverage flag and publishes zero for them (§10); the oracle applies the same rule by environment and tier as well as by the flag, so an older record cannot restore an excluded figure. Coverage is a statement about evidence, not about value: an uncovered parcel may be worth a great deal; the model does not say how much.

9. What the estimate claims

The estimate is positive rather than normative: it describes observed clearing behavior conditional on recorded traits and visible market evidence. It does not say what a parcel ought to be worth, promise an executable price, or forecast the value of the underlying project.

Large residuals can arise from omitted information, market regime shifts, thin evidence, urgency, buyer identity, bidding competition, bundled economics, or simple noise. Similar error from two model classes can suggest that transaction-specific conditions matter, but it does not establish an irreducible limit on accuracy. Both models may share missing variables or the same historical-data limits.

A widely used estimate may also influence the market it describes. Shared reference points can compress some information differences while creating reflexivity around the published number. Model governance should monitor whether adoption changes residuals, listing behavior, or the distribution of trades admitted to the fit.

Table 5. Intended use by situation

10. Governance and change management

Versioned releases. Each public page should carry the build revision and coefficient date. Every release should publish a stable manifest that identifies the code, input snapshot, parameter set, and evaluator schema used to calculate the number. Since 5 September 2026 each build writes a model record beside its manifest: the switches in force, the level and its source, the exclusions, the manual and attested components, and the tier-order property test result. Since 6 September 2026 the release manifest also names the hash of the estimate record described next, the reviewed version and hash of the Resource Rank extract (a versioned manifest that the pages and the oracle both verify; a regenerated or truncated extract refuses to build), and — for the Koda estimate — the fitter, its settings, the walk-forward exam the coefficients were graded by and the sale ledger they were fitted on; a tree whose fitter does not match its coefficients builds without Koda estimates. Since Revision 9 the range calibration is a versioned data file of its own (schema, target, windows, sample size per cell, date), written by a tool from the simulation rows and required by the build.

One calculation record. The static site and companion oracle should consume the same per-parcel evaluator output. Until 6 September 2026 the deployed build gave the oracle the same cohort caution notes as the parcel pages, but the arithmetic could still diverge after an intraday level move because the page rescaled the stored total while the oracle rescaled the land term and held additive values fixed. Since 6 September 2026 every build publishes its estimate record (estimates.json, schema mrdeeds-estimates/1): for each of the 100,000 parcels the estimate, its low and high, the non-land part of the estimate (Koda, artifact, named-resource component) and the calculation path taken, with the level the record was priced at. The parcel pages and the browse index are baked from the same evaluator results, the oracle quotes estimate, low and high from the record, and every surface applies one intraday rule — the land term scales with the live level, the non-land part holds, Club Island parcels never rescale. 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) before a release swaps in — a disagreement holds the release — and, after the swap, requires the oracle to have loaded that release’s record and to agree with it at three level positions, or the release is returned to the previous one automatically. The oracle keeps its own mirror of the evaluator for its explanation fields and logs any disagreement between mirror and record. Writing the fixtures exposed one divergence, corrected the same day: the oracle treated any cell with a segment statistic as a thin-segment candidate, while the pages require the segment’s supply to be under 250 parcels (§6); parcel #32923 was the case found.

Coverage in the record (Revision 10). A parcel outside valuation coverage (§8.2) publishes zero for estimate, low, high and non-land in the record with a coverage flag — 2048 for a collector piece, 512 for a receipts-only parcel — and the browse index carries the same state in a coverage column; every consumer reads the flag and never a hidden figure. The release fixtures sample both kinds and require, on the page, in the index, in the oracle and in the rendered page at three level positions, that no figure, gap or trait-math total appears. Ranges under a level move. A range is the estimate times its calibrated multipliers, so when the live level moves the range follows the point — low and high are scaled by the ratio of the live estimate to the build estimate — on every surface and in the oracle; the earlier rule scaled each endpoint’s land term separately, which moved a low endpoint downward as the level rose on every parcel whose low sat below its non-land component (2,464 on the 6 September book) and left anchored parcels’ ranges unmoved. Gaps against sales and asks are recomputed from the same rescaled point. Explanation copy. The release review of 6 September found four explanation defects, corrected in Revision 10: the market-anchor panel now names the measured ratio and the half-strength one applied (it had shown the shrunk figure as the measurement) and no longer describes the calibrated range as a print spread; a parcel with a tier-1 rare deposit beside a deeper one no longer claims a tier-1 increment that is not fitted; the Resource Rank order test is described by the twelve near-grid cells it now runs (Appendix A); and the oracle’s instructions no longer carry the pre-Revision-8 direction to compute a value from the factor table — the record is the number, the factors explain it.

Release evidence. Segment diagnostics and historical simulations should be reviewed before a model change ships. A warning is evidence for a release decision, not an automatic guarantee that every deployment was blocked or that only one change shipped.

Dated archives. A full walk-forward claim requires dated fitted coefficients, listings, admissions, purified-sale labels, level series, and code. From the Revision 8 release (5 September 2026) every successful build archives, by day, the coefficients it priced from, the listing books, the sale ledger, the daemon’s snapshot, the model record and every published estimate. The archive is a control, not a result: a walk-forward on it becomes possible once it spans a scoring horizon, and this paper will report one only when it has been run.

Resource Rank order test (Revision 9, research). The ranking page describes an order-based market test; the saved diagnostic was a fitted-sample regression whose preferred setting depended on the scale constant. The order-based test now exists as a filed script: on the 2,107 simulated prints, with each print divided by its day’s level, the rank correlation between the published index and price is 0.45 (0.44 for the published rank itself), the summed index orders prices better than the best single slot (0.45 against 0.42), and the correlation is stable across the published near grid (0.44–0.46). It says the ranking orders the market; it identifies neither the scarcity exponent nor the yield step, and the page’s copy will be revised to say exactly that before any denser grid is published.

Change classification. Release notes should distinguish data repairs, causal-integrity fixes, documentation corrections, and predictive-model changes. Only the last category should claim an accuracy improvement, and only at the strength of its evaluation protocol.

11. Disclosure policy

This paper publishes the architecture, reasoning, evaluation design, aggregate results, and known limits. It does not publish the current fitted parameter table, every admission threshold, the complete purified ledger, or the source code. That policy reduces operational and data-poisoning exposure, but it has a clear consequence: an outside reader cannot reproduce an exact parcel estimate from this paper alone.

Table 6. Disclosure boundary

Outcome testing and process reproduction are different. Anyone can compare a public estimate with a later sale, but that does not independently reproduce the screening, fitting, caps, or overrides that created the estimate. The public claim is therefore transparent methodology with externally testable outputs, not open-source reproducibility.

12. Limitations and conflicts

• The v3 headline is a partial historical simulation. Current fitted coefficients and the current listing-derived structures remain in past windows.

• Sales record a day but no time. The confirmed look-ahead fix is end-of-day; intraday causality is untested.

• Earlier training-day normalization uses symmetric local windows bounded by the overall as-of date. It is retrospective normalization, not a causal daily index.

• The non-backtest scorer pins the clock but should still use the evaluation time as an explicit data upper bound when the dataset expands.

• Thin cohorts, rare-environment tier 5, very-rare resources, and stacked scarce items have small samples and wide residuals.

• OTC and private activity is censored unless admitted manually; the visible tail may differ from the full market.

• The current WETH screen, live-ask substitutions, and multi-rare uplift rule are one-sided and may bias the point estimate.

• Current displayed bands are rule-based and are not probability intervals.

• The operator holds and trades Otherdeeds, including through automated bidding. The code carries a written strike list, which today holds one print (deed #4464, 12 August) struck as a collector’s idiosyncratic premium, not as an operator-involved print; no operator-involved print has been identified or excluded, and there is no wallet-based exclusion. The operator’s automated bids settle as WETH bid-hits, which never set the market level or enter the surface fit but, when they clear 85% of the parcel’s own estimate, do enter rare-resource component measurement — so operator-involved bid-hits can currently enter that layer unidentified. A documented standing rule and wallet-based exclusion remain planned; the model is never adjusted toward the operator’s positions.

• The operator reports no financial involvement in the four attested prints identified above. That statement does not replace transaction-level evidence or independent verification.

• Otherdeeds are illiquid speculative digital assets. The estimates are market descriptions, not investment advice or claims of intrinsic value.

13. Independent review and v3 confirmation

The September review combined a production-path scorer, a thirteen-window historical simulation, segment scorecards, layer ablations, internal challengers, a blind adversarial pass, and an independent package review. The operator-prepared Independent Review Report supplied with Revision 7 summarizes that process. The underlying independent audit and frozen first-pass findings remain the authority for third-party conclusions.

The first review package failed reproduction because its reference outputs did not match the included code snapshot. The reviewer diagnosed the mismatch. Package v2 was regenerated from one snapshot and reproduced byte for byte. Package v3 then applied four audit-prompted corrections: the pinned scorer uses the evaluation clock; post-as-of sales are excluded from the level series; the methodology states attested weight one and accurately describes flip tagging; and stale oracle text was corrected for the intended tier path. The independent v3 rerun reproduced fourteen output files byte for byte.

The negative-control injection test is the strongest evidence for the look-ahead correction. Forty impossible-ID prints at 0.000001 ETH dated 13 July were added to a 12 July as-of run. The v3 12 July output remained byte-identical, while the v2 hash and proxy path changed. The result supports the narrow conclusion that post-as-of calendar dates no longer set the v3 starting level.

Independent review disposition

Use the estimate as a rough comparison point for ordinary, liquid parcels. Do not treat it as the deciding number for the named sparse or layer-dominated cohorts, and do not read the band as a probability interval. Retain the corrected print-derived level history. Address artifact isolation, WETH selection, range calibration, site/oracle parity, attested evidence, and full historical archives before making stronger product claims.

13.1 Publication roadmap

1. Use one versioned evaluator record for the site and oracle, with parity tests across level changes and layer combinations. This comes first because two public numbers for the same parcel is the review’s only critical finding.

2. Calibrate intervals on chronological residuals by operative layer and market state; publish target coverage, sample size, and date.

3. Revert or shrink medium/common artifact isolation and remove or redesign the one-sided WETH screen, each behind a predeclared affected-row test. Status: the artifact reversion shipped 5 September 2026 (§4.4); the WETH screen is open.

4. Split rare-resource policy by supply and tier; enforce post-stack tier order and expose the actual hybrid trace. Status: tier order is measured on every build since 5 September (§4.2), not yet enforced.

5. Archive coefficients, listings, inputs, admissions, and labels by build so a full walk-forward test can replace the partial historical claim. Status: capturing from the Revision 8 release, 5 September 2026 (§10).

6. Publish transaction evidence for attested prints or keep them out of public cross-carrier effects until independently verified.

14. Status of this document

Revision 10 is based on the same production and evidence snapshot of 6 September 2026 as Revision 9 and was published the same afternoon. It records the release review of that day and the corrections made for it: the valuation-coverage rule (§8.2) — no published estimate for Chaos parcels or for tier-5 parcels of the three rare environments — the range rescale rule and sale identity on the public pages (§10), one struck event (§3.3), and the explanation-copy corrections. No estimate, range or model rule changed for any covered parcel.

Revision 9 is based on the production and evidence snapshot dated 6 September 2026. It records the second set of model changes — the depth-preserving tier order, gradual pooling for named components, cell anchors at half strength — with their pre-declared test against the Revision 8 tree (Table 3c) and the three candidates measured and not adopted; the calibrated ranges and the range-only rule for tier-5 rare-environment parcels; the counterparty rule for operator prints; and the Resource Rank order test as research. Revision 8 (5 September, amended 6 September) stands wherever nothing changed.

Revision 8 was based on the production and evidence snapshot dated 5 September 2026. It records the first model changes shipped after the independent review — the artifact-isolation reversion, one level convention for component measurement, the thin rule on the land term and the print-derived level for live builds — with their pre-declared test against the pre-change tree (Table 3b), and the evidence controls that shipped with them: sale-event identity, the model record, the tier-order property test and the dated input archive. It also names the harness’s error measure by its denominator. Revision 7 (3 September, amended 5 September) aligned the paper to the audit evidence and code; its text stands wherever nothing changed.

The paper remains a methodology and limitations disclosure. The parcel pages and the companion oracle carry the same caution labels and the same Revision 8 rules; the calculation-parity issue between them, open at publication, was closed by the amendment of 6 September 2026 (§10). Revision 7 was deployed on the evening of 3 September 2026 (Central) as build b316 (model code, 22:24 UTC) and b319 (the page, 00:02 UTC, 4 September) with oracle build 6b9ee664; tag review-2026-09-final (commit 3d66323, Appendix A) marks that code. Revision 8 was published with the Batch 3 model build and Revision 9 with the Batch 5 build; the change log records the build numbers, and Appendix A the generator identities.

Table 7. Revision record

Appendix A. Evidence and reproducibility record

The following identifiers make the claims in Sections 7 and 13 traceable to the frozen package. They establish identity and reproduction of the supplied files; they do not independently verify transactions or inputs absent from the package.

Table A1. Principal reproducibility identifiers

The deployment candidate differs from the auditor-confirmed v3 build only by the auditor’s post-confirmation evaluation-clock hardening and three reader-facing changes: the “indicative range” label, the cohort review notes, and the corrected link from those notes to the methodology page. Estimate outputs remain byte-identical to v3. The candidate identifiers appear in Table A1; it was deployed as build b316 on 3 September 2026 (22:24 UTC) and carries this revision of the page from build b319 (00:02 UTC, 4 September 2026).

Evidence documents supplied with the review record:

• Independent-Model-Audit-2026-09-03.md — independent findings and disposition on the v2 package.

• First-Pass-Findings-Unedited.md — frozen blind findings before comparison with the operator’s assessment.

• Reproduction-Gate-Report.md — diagnosis of the first package mismatch and the reproduction gate.

• V3-Rerun-Lookahead-Confirmation-2026-09-03.md — independent reproduction, metrics, injection test, and remaining limits.

• Independent-Review-Report-2026-09-DRAFT_1.docx — operator-prepared pre-publication summary used as context for this revision.

• R7-ASSESSMENT-2026-09-03.md and .docx — final line-by-line publication assessment against the committed deployment candidate and evidence record.

• PROVENANCE.md and the v3 output hashes — authoritative code, input, and reference-output identities.

Appendix B. Glossary