Stage 16 GPU evidence and agentic atlas research

Status: active development ledger, opened 2026-08-11. Stage 16I is complete as a validated local assessment draft and Stage 16J is complete as an exploration-only Equal Earth comparison prototype; human review remains pending. No GPU benchmark, promotion, release product, publication, email, push, upload, or external transfer is claimed by this ledger.

Development records · Stage 15 closed ledger · AI workflows and screen products · Marshall Islands speculations v01 · Equal Earth positioning speculations v01

Scope inherited from Stage 15

Stage 15 closed after implementing its approved active subset: the frozen Stage 14 handoff, lossless 2K controls, local CDN/WASM layout, dual-year particulate standard family, six-projection Majuro evidence pass, and bounded PurpleAir-interface and water-debris experiments. Work that Stage 15 explicitly deferred is not a Stage 15 defect or incomplete release claim. It moves here as separately activatable Stage 16 work.

The tracks are independent. A hardware schedule must not block agent-contract work, and an evidence-source gap must not be disguised as a GPU benchmark failure. None enters make all, GitHub release, or UCB AAO/S3 without a later reviewed implementation and promotion decision.

Stage 16I local outcome

The reviewed package is rooted at reports/cartofreako-audit-outcomes-03.pdf. Only that archival PDF boundary object, its predecessor feedback, and its planning brief are admitted as source-controlled research records under the curated reports policy. Report source, HTML, CSS, manifests, metadata, copied plate trees, editor backups, and visual-QA intermediates remain ignored. The report uses the canonical Devastation Pacific v3.9 document_type: assess-outcomes, the randoma11y-accent style, and the required coda-share modifier.

The completed QA record includes:

• canonical assess-outcomes and coda-share source and PDF checks;
• metadata-schema, JSON, and local-link validation covering 34 local files and 79 internal anchors;
• synchronized installable house-style content and 132 passing house-style tests;
• a tagged, 45-page PDF with selectable text, embedded Atkinson typography, and the intended portrait/landscape page geometry; and
• visual inspection of all 45 rendered pages, including twenty uncropped one-primary-plate-per-page appendix images.

The assessment records the Cartofreako × Alpha60 joined workflow as INFERRED, not implemented; stable AAO inventory and integrity evidence do not become measured cache, performance, or power outcomes. Proposed meeting participants and contributors remain suggestions, not invitations or confirmations. Source-control admission records the reviewed PDF as a boundary object; it does not make the package a Stage 16 release or authorize push, publication, email, upload, invitation, or external transfer.

Stage 16J Equal Earth outcome

Stage 16J implements the published spherical Equal Earth forward/reverse equations in standalone C++ and dependency-free JavaScript without extending the closed six-family release runtime. Its neutral fixture bundle contains 30 cases across the canonical Greenwich registration and an explicitly experimental 11.5°E Africa-centered variant. Frozen PROJ 9.6.2 and D3 Geo 2.0.1 observations agree to a maximum raw difference of 7.021666937153402e-16; the finite-difference spherical area audit has a maximum error of 1.203175781228083e-9 over 35 samples.

The one-to-five comparison record renders:

1. a matched Web Mercator and canonical Equal Earth full-world control;
2. canonical and Africa-centered Equal Earth with Tissot diagnostics;
3. one Africa–Europe WGS 84 source-window slice through three registrations;
4. Equal Earth beside AuthaGraph, Dymaxion, and Myriahedral Afro–Eur–Asia; and
5. full-carrier, geographic-preclip, registered-perspective, and Cahill–Keyes native-cell slice roles.

The five 2560 × 1440 PNGs are checked local exploration products and are registered in make all-experiments. Equal Earth is not in make all, the standard artifact catalog, print generation, GitHub release, or UCB AAO/S3. The centered variant is not labeled EPSG:8857. Mathematical implementation does not establish #CorrectTheMap compliance, corrected perception, community authorization, or a decolonial outcome.

Re-frozen benchmark baseline

The Stage 15A benchmark input was re-frozen on 2026-08-13 at the current clean corpus: 211 artifacts across 32 pass IDs, 11 layouts, and 14 approved slices at clean commit 7ba10e6b11475326873b3e206067fecebe7cc1c2, so every retained parent remains buildable from the standard graph.

The handoff records:

• runtime API 3 and geometry ABI 1 manifests;
• canonical SVG/PDF and exact-ratio print geometry;
• every authoritative 3840-pixel parent and SHA-256;
• exact 1920 × 1080 PNG/lossless-WebP baselines, content rectangles, padding, affine transforms, parent identities, and hashes;
• projection, layout, slice, lifecycle, authority, source-period, license, limitation, and interaction metadata; and
• the Stage 15 lossless 2048 × 1024 landscape and 1024 × 2048 portrait controls derived from the same authoritative full PNG parents.

If a parent, transform, or catalog record changes, the new case receives a new identity rather than rewriting the frozen fixture.

GPU benchmark contract

Every 16A–16D run emits a machine-readable result and compact human summary. It records:

• input/product hashes; encoder, decoder, browser/runtime, driver, operating system, dependency versions, exact options, color space, alpha mode, mip policy, and determinism status;
• encoded bytes, compression ratio, encode time, CPU decode time, GPU upload time, time to first usable frame, peak host memory, estimated texture memory, and steady-state sampling behavior;
• cold and warm observations, warm-up and repetition counts, median, and tail measurements rather than one favorable timing;
• pixel and structural error against the lossless reference, transparency, edge behavior, title/legend legibility, thin-line retention, field-opacity preservation, and pass-specific moiré observations;
• Float64-to-Float32 projected and screen-pixel displacement, reverse residual, status, and candidate-set agreement at cuts and overlaps; and
• unsupported operations, failures, warnings, and human visual disposition without silently dropping a product or machine.

PNG remains the visual control, not necessarily the preferred transfer format. Flat water, translucent fields, fine networks, text, boundaries, and high-frequency curve work must be reviewed separately; one aggregate score cannot hide pass-specific damage.

Hardware matrix

The first reproducible baseline is the Framework Desktop with AMD Ryzen AI Max+ 395, integrated Radeon 8060S, 16 physical cores/32 hardware threads, and approximately 125 GiB of reported memory. Each run captures the actual GPU driver, browser/runtime, power profile, display/headless mode, and available memory.

Broader adoption requires at least one discrete-GPU result and one browser GPU path. Results remain per machine. Software rendering may be retained as a correctness fallback but is not GPU-performance evidence.

Reassessment

Cartofreako is already useful to an agent as an atlas selection system and as a bounded, one-off experiment system. It is not yet a general agentic research-iteration system.

The current evidence is substantial:

• a checked catalog exposes 211 standard products across 32 pass IDs;
• deterministic requests and decision receipts select existing artifacts;
• all six projection families have candidate-aware forward and reverse APIs;
• 1080p derivatives support preview-first inspection without opening very large SVGs, while the full PNG, layered SVG, PDF, 44-inch-leading-edge, and A0 products remain authoritative;
• the Marshall Islands v01 workflow turns five projection hypotheses into reproducible PNG comparison plates;
• the Majuro evidence pass exercises six carrier projections without claiming that a planetary carrier is the local analytical grid; and
• the PurpleAir-interface and water-debris experiments demonstrate explicit exploration-only products, missing-evidence labels, source periods, and non-promotion boundaries.

The gap is orchestration. The Marshall Islands renderer hard-codes one useful sequence; the artifact selector chooses only already-known products; and each new research family defines its own manifest, generator, checks, contact sheet, and review language. An agent can assist the present workflow, but cannot yet express and run the whole sequence through a stable, implementation-neutral contract.

Proposed operating model

Use two connected loops rather than one unconstrained “make a map” command.

Loop A — select and compare existing atlas products

1. Normalize the research question into a checked request: geography, interval, evidence themes, audience, allowed lifecycle classes, output budget, and publication boundary.
2. Query the artifact catalog and projection-capability registry.
3. Return one-to-five projection, slice, and plate candidates with explicit reasons, exclusions, source dates, cuts, inverse behavior, and evidence limitations.
4. Render only previews or 1080p comparison sheets first.
5. Record a human review disposition before producing full PNG, SVG, PDF, or print derivatives.

Loop B — develop a bounded new analysis

1. Convert an evidence gap from Loop A into an experiment specification.
2. Inventory candidate sources by evidence role: observation, scenario, interpretation, context, provenance, or UNAVAILABLE.
3. Require an explicit acquisition decision for credentials, terms, large downloads, personal data, or non-redistributable sources. Generation never implies that authorization.
4. Build a small deterministic canary and a contact sheet, then run numeric, source, lifecycle, and visual checks.
5. Let a reviewer accept, reject, or request one controlled revision. Change one named variable per iteration where practical.
6. Preserve an iteration lineage and result receipt. Promotion to optional or standard is a later, separate decision; GitHub release and UCB AAO/S3 deposit remain human-invoked operations.

This division lets an agent recommend and iterate without allowing a fluent answer to become source evidence, community authorization, or release authority.

First Marshall Islands recommendation queue

The initial agent should rank candidates by research purpose, not declare one projection universally “best.” The following queue is the proposed starting fixture.

AuthaGraph is a useful secondary control for full-world distribution and Voronoi is useful for seam and cell-ownership comparison. Neither should displace the Pacific-centered, planetary-water, focused-cell, and local-grid roles above without evidence from a specific question.

For worldwide reef work, begin with Myriahedral Pacific and Dymaxion water carriers, add Cahill–Keyes octants as repeatable basin panels, and use a source-governed local projection for each reef system. A Star-X series may carry an interpretive time sequence, but should not be the measurement grid.

First experimental-analysis queue

The agent should propose small, evidence-bounded increments in this order:

1. Marshall Islands evidence comparison. Join the current Majuro topobathymetry and inundation panels to selected reef, temperature, particulate, and source-timed atmosphere context without merging their dates or evidence roles.
2. Anthropocene water debris v02. Keep the five real 2018 depth stations as observations, then test one newly reviewed debris source at a time. Never infer a garbage-patch boundary or relabel the 2025/2026 atlas edition as the observation year.
3. Atmosphere sequence. Compare the optional JAXA snapshot, standard particulate and temperature editions, and the synthetic PurpleAir interface as separate panels. A real PurpleAir adapter remains blocked on source-use review and observed-data provenance.
4. Worldwide reef sequence. Combine the released 2011 reef-threat geometry with current ocean-heat or bleaching evidence only after the new source’s resolution, interval, uncertainty, redistribution, and missing semantics are recorded.
5. Local atoll evidence tier. Add shoreline, freshwater, infrastructure, current reef, and ocean-heat sources independently. Community or regional review is a promotion gate, not a task the agent can satisfy.

Each experiment should start with PNG-only previews or a contact sheet. Large SVG/PDF/print products are generated only after the visual direction and evidence boundary are reviewed.

Contracts required

Four new checked contracts would generalize the existing canaries.

atlas-research-request-v1

Minimum fields:

• research question, geographic focus and bounding geometry;
• time interval and whether comparison across unlike periods is permitted;
• evidence themes and allowed lifecycle classes;
• intended audience and purpose (planetary-context, regional-comparison, local-analysis, exhibition, or runtime);
• candidate limit from one to five, output budget, and preview-only default;
• allowed local/public inputs and explicit network/acquisition authorization;
• community, privacy, license, publication, and release boundaries; and
• catalog, runtime, source-commit, and house-style identities.

atlas-recommendation-receipt-v1

For every candidate, record its projection, slice, plate IDs, score components, distortion/topology rationale, reverse behavior, source periods, lifecycle, limitations, and rejection reasons. The receipt must distinguish observed metadata from inferred suitability and retain any human override.

atlas-experiment-v1

Record the hypothesis, parent request/receipt hashes, source inventory, evidence roles, transformations, projection/slice matrix, layer defaults, opacity/title rules, render recipes, deterministic seeds, output formats, resource limits, QA fixtures, and an unconditional initial exploration-only lifecycle.

atlas-experiment-result-v1

Record resolved input and toolchain hashes, outputs, warnings, forward/reverse residuals, missing evidence, visual-review status, reproduction command, parent/child iteration IDs, and promotion disposition. A separate human review receipt should name the reviewer action without implying community consent, scientific validation, endorsement, or permission to publish.

Proposed command surface

Keep pure JSON commands beneath thin Make wrappers:

make propose-atlas-research \
  RESEARCH_REQUEST=work/marshall-islands-request.json

make render-atlas-experiment \
  EXPERIMENT_SPEC=work/marshall-islands-v02.json

make check-atlas-experiment \
  EXPERIMENT_RESULT=work/marshall-islands-v02-result.json

make package-atlas-review \
  EXPERIMENT_RESULT=work/marshall-islands-v02-result.json

The proposal target is read-only and may not fetch. The render target may use only inputs authorized by its checked specification. The check target must be offline. The review package contains previews, hashes, source/evidence tables, limitations, and a blank review receipt; it does not publish.

Recommendation logic

The selector needs a checked projection-suitability registry rather than an opaque model ranking. Its features should include:

• geographic containment, relevant seam crossings, visible area, and label budget;
• full carrier versus native-cell or geographic preclip;
• purpose-specific continuity, distance/area limitations, aspect ratio, and portrait/landscape fit;
• forward/reverse availability, ambiguity, component/face qualification, and residual evidence;
• compatible existing passes and source periods;
• evidence lifecycle, license, uncertainty, missing-data semantics, and community-review state; and
• output cost from thumbnail through authoritative print.

The deterministic score should produce an explanation. An agent may summarize or challenge it, but cannot silently replace it with an unrecorded preference.

Safety and authority boundaries

• No target reads secrets from a prompt, writes credentials into a fixture, accepts provider terms, or authorizes an external source.
• No experiment enters make all, a release, or AAO merely because it renders successfully or appears in all-experiments.
• No synthetic PurpleAir marker, inferred debris geometry, interpolated shoreline, false zero, or absent observation may be described as measured.
• No agent claims community consent, local interpretation, scientific validation, endorsement, patentability, or publication authority.
• Current public S3 v13 assets and current local post-v13 products must remain distinguishable; a local result cannot invent a public URL.
• Every authoritative print product retains its projection-specific ratio, 44-inch leading edge, and A0 path. Preview and 1080p products remain derivatives.

Staged implementation proposal

The generic implementation is approximately 10–16 engineering days. A meaningful external collaboration trial likely makes this 3–5 elapsed weeks. New source acquisition, scientific review, and Marshallese/community review are separately scheduled evidence work, not hidden inside that estimate.

Approved direction

The 2026-08-11 planning decision accepts:

1. the two-loop selection/experiment model;
2. Myriahedral Pacific, Dymaxion, Cahill–Keyes, local AEQD, and Star-X as the first Marshall one-to-five recommendation fixture;
3. the source-timed atmosphere comparison as the first generic canary, followed by water-debris v02; and
4. Stage 16 as the distinct program for the generic contracts and workflow.

Approval moves the work into this ledger; it does not claim that the schemas, commands, source adapters, collaboration trial, or GPU continuations have already been implemented.

Combined staging estimate

These estimates are not additive release gates. Run the tracks independently, retain separate receipts, and promote only the evidence that completes its own acceptance path.