Method & provenance · Performance Feedback Atlas

01 / Database & snapshot

The database comes first.

Feedback Atlas is generated from a receipt-backed literature database. Counts, routes, filters, and graphs are projections of the same checked source ledger; the interface is never a second bibliographic authority.

475provisional records

5,977citation edges

7core records

1935–2026year range

Provenance record

Source-snapshot timestamp: 2026-08-16T00:00:00.000Z
Schema version: feedback-atlas.site-corpus/3
Input contract: Database manifest → corpus manifest → sanitized public projection
Launch component: 279 provisional records
Round-1 receipt: 216 candidates = 196 admitted + 19 external-sibling exclusions + 1 held
Connected slice: publication-slice:55080a071a973c7a0757301f9fc843fe
Source mutation: None; inputs were read-only

The round-1 database binds each researched candidate to its source-row hash, identity and deduplication outcome, provider-evidence digest, bibliographic projection, theme decision, and any claim or review ceiling. Its admission unit is a provider bibliographic record or manifestation—not a unique scholarly work. The 19 exclusions hit a private external sibling receipt while remaining absent from the public launch component; that sibling corpus is a dedupe boundary, not a website input.

Identity and claim boundary: a scholarly red-team records 23 explicit adjudications, including document types, preprint/VOR and contribution relations, and 13 corrected projection fields. The superficially similar 2015 call for papers and 2018 review are explicitly distinct, not versions. One possible thesis/seed match remains pending human review and does not alter stable source-record identity. Eighty admitted rows have a lead-only claim ceiling, and 11 carry research-use restrictions; those claim fields remain in the database and are not projected as website findings.

Date distinction: the timestamp above fixes the source snapshot, not the latest build. The round-1 component is receipt-dated 2026-08-16; the 2026-08-18 receipt-dated scholarly corrections are applied to the projection in place without advancing it. The launch component's dependable upstream retrieval time remains unavailable and is not inferred.

02 / Citation graph

Direct edges come from reference lists.

A direct citation edge exists only when one imported source record lists another imported source record among its references.

Record A · citingRecord B · cited

Direction matters. An arrow points from the citing source record to the cited source record.
Edges are local to the snapshot. An absent edge does not show that no citation exists outside this 475-record import.
Local and global counts differ. “Cited by in corpus” counts incoming graph edges; “global cited by” is the OpenAlex snapshot metric.
No edges were inferred. All 5,977 direct edges are present in the source output.

03 / Connected relations

Shared references and shared citers remain distinct.

Feedback Atlas publishes the engine’s bounded, aggregate neighbor rows unchanged. The site validates and displays them; it does not calculate, blend, or rerank relationships.

Bibliographic coupling

Two public records cite some of the same reference targets. Reference degrees count distinct targets represented in the canonical graph index; the raw target identities are not exposed here.

8,884 exported per-seed rows

Co-citation

Two public records are cited together by mapped public citing records. The public metrics are recomputed over that mapped public endpoint set.

4,483 exported per-seed rows

Public selector top_k_per_seed_after_public_endpoint_filter

Minimum shared count: 2 shared references or shared citing records, as applicable, before a pair is eligible.
Per-seed cap: At most 20 neighbors for each record and relation type.
Why-shown metrics: Shared count, both endpoint degrees, cosine, Jaccard, rank, and eligible-neighbor count.

Selection rule

Filter to approved public endpoints before thresholding and top-k selection.
Require at least 2 shared references or shared citing records.
Rank separately for every seed and relation type by shared count, cosine, Jaccard, then stable source-record ID.
Export at most 20 rows for each seed and type, retaining rank and the full eligible-neighbor count.
Publish only aggregate metrics; raw references, raw out-of-corpus reference endpoints, private engine Work IDs, and evidence fingerprints remain outside the site.

These rows are bounded discovery evidence, not direct citations, identity decisions, recommendations, or proof of a complete literature neighborhood. An undirected measure can still appear in only one seed’s exported list because top-k selection is performed per seed.

Whole-corpus delivery

The graph page draws the entire snapshot from one prerendered feedback-atlas.graph/1 dataset at /graph-data/graph.v1.json: 475 nodes and 5,977 directed citation edges, fetched once and never refetched. Node metadata in the dataset is display-minimal — identifier, title, year, first-author surname, venue, theme, access, in-corpus degrees, global citation count, and layout coordinates. The complete provisional record stays on its permanent record page.

The dataset is derived at build time from the same checked-in corpus the rest of the site reads. Its node and edge totals, and every precomputed degree, are recounted from those records during generation, and generation fails rather than publishing if any total or degree drifts from the corpus, if an edge leaves the record set, or if the file would exceed 1,200,000 uncompressed bytes. The browser revalidates the same invariants — schema, ordering, edge closure, recounted degrees, and canonical serialized bytes — before anything is drawn, and refuses to draw a response that fails. The build audit independently regenerates the dataset and byte-compares it to the published file.

The per-seed feedback-atlas.seed-graph/1 route contract remains a stable per-record API. All 475 shards are regenerated from the database projection; a “complete seed shard” means only that its defined snapshot closure was not truncated, not that the source import or literature neighborhood is complete.

Reproducibility gate: every production build reconstructs the merged corpus and public projection from the checked database and component manifests, byte-compares them with the checked-in public artifacts, regenerates every route and export, and fails if any source, count, relationship, privacy boundary, or generated byte drifts.

Canvas layout

Every node position is computed once at build time by a deterministic seeded force pass over identifier-sorted inputs, so the coordinates are identical for every visitor and independent of the order the records arrive in. The browser only reads them. Node area grows with in-corpus citations, node color follows publication year, and seed anchors are drawn as diamonds. Clicking a record raises its incident edges to full opacity with direction arrows and labels its neighbors, while every other record and edge stays drawn; background click or Escape returns to the fitted whole-corpus view. Search and theme, access, and year filters dim non-matching records and never remove them. Focus changes add traversable browser-history entries, and ?focus= restores a record on load.

Position is presentation, not another score. The browser does not infer neighbors, rescore rows, encode cosine or Jaccard as distance, or claim that spatial proximity establishes semantic similarity, quality, canonical identity, recommendation, or completeness. The relationship table beside the canvas remains the explicit explanation of every displayed line.

04 / Themes

Eight lenses, with overlap by design.

Each provisional record has a primary theme, used for record color across the library and for the graph’s theme filter and detail panel, and one or more theme memberships for discovery. Theme counts are not mutually exclusive, and every assignment remains provisional.

Review status: the 279 launch assignments were legacy-migrated. For the 196 round-1 admissions, the database retains 577 panel votes covering 193 records plus three documented engineer assignments. Twenty-one assignments—16 low-confidence, two missing-vote, and three engineer-assigned—remain elevated for domain review. Four later red-team recommendations are recorded as pending and do not overwrite the retained labels. No abstract text is published in this beta.

Foundations 269 records

Aspirations & reference points 213 records

Problemistic search 62 records

Multiple goals 79 records

Innovation & change 143 records

Governance & leadership 71 records

Networks & corporate scope 45 records

Risk & survival 117 records

Because memberships overlap, these counts sum to more than 475. The graph uses primary theme only for node color; filtering by theme uses every provisional membership.

05 / Coverage

One intentionally narrow public indicator.

Access is derived only from the sanitized public graph projection and collapsed to two states.

Open access164

The public projection labels the record open access.

Metadata only311

No public access classification is exposed beyond metadata.

The two states are mutually exclusive and sum to 475 provisional records. Feedback Atlas does not ingest collection-specific access overlays.

06 / Citation and query exports

Portable records, with their limits attached.

Citation files are deterministic presentations of the same site-corpus/3 authority used by the pages. They do not create a new bibliographic or identity authority.

Whole snapshot 475 provisional records · W-ID code-point order

CSL JSON: Generic document entries with literal author display names.
RIS: Generic GEN records; no publication type is inferred.
BibTeX: Generic misc entries with stable W-ID-based keys.
CSV: The same selected public display fields, with formula-leading cells neutralized for spreadsheet safety.

Five static routes. Four whole-snapshot citation files and one companion manifest are prerendered. The manifest binds source lineage, ordered IDs, DOI coverage, byte counts, media types, and SHA-256 hashes.
Explicit scoped preparation. The library prepares one chosen format for the exact current filtered result and visible sort order; each record page prepares one record. Separate prepared links expose the citation file and its exact companion manifest.
Exact graph query manifest. After a seed shard validates, the browser can bind its raw bytes and digest, normalized URL state, selected IDs, source and visible counts, and the current filtered typed relationship rows. Filters do not rerank; render coordinates are excluded.
Observed identifiers only. This snapshot has 467 normalized long-form DOI observations, no ShortDOIs, and 8 records without a DOI. DOI is not used as a canonical-work merge decision.
Missing venue stays missing. The Unknown venue display sentinel is omitted from container-title-style citation fields and is blank in CSV.

Interchange ceiling: these exports are not a lossless source round-trip, a publication-type or canonical-identity decision, a completeness claim, a third-party importer guarantee, or a reference-manager import workflow. Access labels and DOI presence do not grant copyright or reuse permission; verify the underlying publisher or repository record.

07 / Limits

Use the atlas as orientation, not verdict.

Bounded scopeThe graph says what is connected inside the fixed import, not across all scholarship.
Provisional identitySource IDs have not been adjudicated into canonical works; duplicates, reviews, books, and version variants may remain.
Snapshot metricsGlobal citation counts and access URLs can change after generation.
Provisional themesLegacy and panel-derived memberships are navigational labels and may not capture every construct in a record.
Bounded neighborsCoupling and co-citation rows are per-seed top-k selections—not quality, causal relevance, endorsement, or complete neighborhoods.
Metadata variationPublication years, venues, duplicate titles, and source versions can contain upstream inconsistencies.
Coverage ≠ rightsAn open-access label does not make Feedback Atlas a repository or grant reuse rights.

08 / Privacy

Static by design.

Feedback Atlas has no authentication, analytics, cookies, advertising, pixels, or third-party embeds. Search and graph interactions run in the browser and are encoded only in the page URL for shareable state. Filtered-library, single-record, and graph-query downloads are prepared locally only after an explicit action.

✓ No accounts✓ No trackers✓ Public projection only✓ Static output

Deployment boundary: endpoint functions declare intended content type, attachment disposition, cache, and nosniff headers, but a static host may infer, omit, or rewrite headers for emitted files. A deployed host therefore needs its own header and download-behavior verification.