# Research notes: Libertad Parametrizada by zJorge Status: draft-for-human-review. No human reviewed, edited or fact-checked this package. ## What was retrieved, and how All live work ran through a browser pane on the operator's computer. The cloud container's egress proxy blocks artblocks.io, the Hasura index, the media proxy and the generator, so nothing in this package was fetched from the container itself. The artist script was retrieved from the index by an HTTP POST filtered on the project id, corroborated against the contract address and project id returned in the same response. It is 10,465 bytes, SHA-256 `5707f7edbdf8bc932e11076e8f5c95e09c302528e02cd85a334eec2e9cf15c35`, script_count 1, and the single scripts row reproduces the script field exactly. Index retrieval, not a direct read of contract storage; no block pinned. The whole script was read. No artist code appears in this package. ## The reimplementation Trait rules were read from the script, reimplemented independently, and run against the hashes of all 243 tokens. Full results are in evidence.json. Six rules reproduce at 243 of 243 (Framing, the static against moving split, the door shape component, Doors facing, Wind, Gravity) and Palette reaches 243 of 243 once the Procedural override is added. Two reach 242 of 243 and one reaches 239 of 243. The named exceptions, reported rather than fitted: - Padding: 92000034, whose padding byte is 128, the exact threshold value and the only token in the edition holding it. - Doors size: one token carrying the larger size multiplier is published as Regular. - Framing shape presence: 92000048, 92000059, 92000098 and 92000198 carry the attribute although their frame flag byte is above the mat threshold. - Glitch eligibility: 92000127 and 92000132 satisfy the predicate and carry neither attribute. In each case a threshold could have been invented that absorbed the exception. None was, because in each case the same threshold would have excluded correctly labelled tokens. ## The glitch attributes This is the strongest result in the package. The project page advertises two attributes, Has glitch and May glitch, that only 22 of 243 tokens carry, 11 each with no overlap, and nothing on the platform explains them. The script contains one branch that fills a single door from a colour snapshotted before the colour strategy cascade rewrites the palette. Reimplementing the branch's eligibility predicate captured all 22 labelled tokens with no false negatives and excluded all 219 unlabelled ones. Adding the branch's central half of canvas requirement split the 22 exactly: 11 of 11 central are Has glitch, 11 of 11 non central are May glitch. Verification then moved from code to canvas. Token 92000147 sits on a colour strategy that desaturates the doors. The reimplementation predicted the anomalous door at 64.4 percent across, 37.3 percent down, colour HSL 24, 99, 59. The rendered pixel at that position measured RGB 254, 130, 47, the exact conversion. Isolating orange pixels across the whole 1324 by 1324 canvas returned 5,647 of them, one door, bounding box centred at 64.4 and 37.2. Predicted position and predicted colour both confirmed independently. ## Defects found in the program Three, described in the essay and recorded in evidence.json. The divisor 265 where 256 was meant, in two places, costing the top of the intended range. The seed slice that includes the `0x` prefix, so fourteen hex digits are parsed into a value beyond exact integer representation; only 35 of 243 seeds are exact. Checked rather than assumed: after truncation all 243 seeds remain distinct, so the edition suffers no collisions. The shuffle index that can reach one past the end of its array, splicing nothing, returning an empty array that is then coerced to zero. Reproducing the placement loop gives 3,587 collapses across the edition, in every single token, up to 26 in one. The affected doors land outside the drawn frame along the left edge and the bottom, which is visible in almost every work in the project. No intention is attributed to the artist for this. ## Where the published metadata and the render disagree Eighty one tokens are published as Filled and are drawn at 60 percent opacity. Rather than rest on a code reading, this was measured: token 92000018 carries 2,874 distinct colours from a five colour palette, and the computed 60 percent composite of its door colour over its mat, RGB 153, 129, 30, is present across 4,580 pixels. Ninety five tokens carry no Framing shape attribute at all. The trait extractor that produced the published features is not part of the retrieved script and was not located, so these are reported as divergences rather than attributed to a particular piece of software. ## Failed attempts, kept Six, listed in full in evidence.json. The ones that would have produced wrong findings if not caught: Physical key presses did not reach the page. After a real arrow press the sketch's key variables still read zero. Reporting the controls as tested on that basis would have been false. The tests were redone with dispatched key events and the method limitation is stated everywhere the results appear. The first read of token 92000018 returned a single colour and a frame counter of zero. Read as a blank or broken render, that would have been wrong: the animation loop was suspended because the browser pane was hidden, and a screenshot forcing a repaint started it. The first uppercase M test showed no change, but the door count was near the program's ceiling, so the test could not distinguish a broken control from a saturated one. Repeated at a lower count with a lowercase control in the same run. The attribute schema was first derived from a single token and missed Has glitch and May glitch entirely. Recensused across all 243. The Density band overlap was hypothesised to come from the 265 divisor. Tested with divisors of 265 and 256 and the raw byte; all three produce the same overlap, so the hypothesis was wrong and was dropped in favour of door count divided by drawing width. The artist website of record 404s. The artist's page for this work was found through the site's own works index instead, which turned out to hold the only first hand account of the project's conception in this batch. ## Frame rates Not reliable in this package. The browser pane was hidden throughout, which throttles the animation loop and in one case suspended it entirely. Where a rate is quoted it is what the environment reported at that moment. Findings resting on whether an image changed between rendered frames are unaffected. ## Rights Artwork licence recorded as category creative-commons with the label preserved exactly and terms url null, with the gap explained. The deed for the named identifier was read directly. No artist published licence instrument was located and the project details panel publishes no licence field at all. Code rights unknown and kept separate. The p5 dependency's own licensing was not inspected and is recorded separately. NFT ownership is not an artwork permission, and readability of on-chain code is not a redistribution right. ## Unresolved leads not converted into sources The trait extractor producing the published feature objects. Upstream p5 licensing. Primary sale receipts for the 0.119 ETH price. The Spanish language versions of the artist's pages, which were not read; the English versions were used throughout. ## Actual model, tools, elapsed time and measured usage Model: claude-opus-5, served through the Claude desktop application in Cowork mode on the Claude Agent SDK. The batch routing label was "claude", a routing label and not a model identifier. The agent is not Astra. Tools actually used: HTTP POST to the Art Blocks public GraphQL index; a browser pane on the operator's computer for page rendering, text extraction, live renderer execution, canvas sampling and dispatched key events; an independent reimplementation of the colour selection, the pseudorandom shuffle, the coordinate remap and eleven trait rules, run against all 243 token hashes; the Web Crypto API for the SHA-256 digest; local shell and file tools for package assembly and validation. No human reviewed or edited the prose. Word count: the coordinator waived the 1,500 to 2,500 word limit for this batch during the session. This essay runs longer than the original cap by that instruction. Elapsed work: produced within a single continuous session on 18 September 2026, Melbourne time. Telemetry: no token usage telemetry is exposed to the agent. All counts are null in usage.json with the reason given. Nothing estimated and labelled measured; batch totals not divided into invented per-essay figures. ## Coordinator revision Complete program read; four original renderers inspected; Down and space tested on#80. Fresh243-token census and complete linked-wallet check retained. Artist project page read in browser after web retrieval failure. Removed unsupported opacity mismatch, intended-divisor allegation and unverified exact statistical claims. Standard artwork deed and p5 licence identified separately. Submitted records preserved for attribution.