Updates

Public currentness record

Updates revision 2026-09-04 · Last reviewed September 4, 2026.

This is the short, dated record of material status changes. A planning page or small controlled test is not a shipping claim.

Current public state#

2026-09-04#

• Status labels were re-checked against the engine's acceptance ledger and re-bucketed on the homepage, About, the brief, and the evaluation page. Exercised on controlled fixtures: source-linked answers, named gaps and abstention, and repeated-run stability. In validation: time-scoped recall and multi-step paths, alongside the recall engine and the pack pipeline. Nothing was promoted; two items moved from "exercised" to "in validation".
• The positioning no longer calls CleverMemory document AI. The homepage eyebrow and search title state the category as proof-carrying document intelligence; the technical overview already explains why the recall path is not AI in the usual sense.
• The 45-document challenge joins the top navigation, and its page now leads with the download, the archive digest, and the license. Packs and the pack format moved under Learn.
• The corpus matrix's question count was corrected from 40 to 55 in the engine repository and republished here. The archive was regenerated, so its digest on the challenge page changed; the per-document digests did not. Site validation now checks the archive's digest and size against that page and the matrix's count against the suite.
• The use cases now say how their six questions relate to the published suite: five are suite scenarios, shortened where the suite asks for more than one thing, and the clarify question is this site's own. The challenge page's summary was corrected to match, and the homepage walkthrough's row count was corrected: 15 rows sit outside the five cards, not 23.
• Scheduled articles are now held, not failed. An article publishes on its date only when its status review falls within the seven days before that date; otherwise the site builds normally and the article stays a placeholder until it is reviewed. The remaining scheduled articles are no longer drafts; each is held by that rule until its review. Part 10 of Foundations was re-reviewed for its September 7 publication, and part 1's closing list now links parts 6 through 9.
• The comparison adds enterprise search and document extraction services as neighbors. The blog index no longer refers to an industry browser removed on August 1, and the agent-facing site summary no longer names the retired Ardmore case file.
• The site's robots signal now allows answer engines to use the pages (ai-input=yes); training remains disallowed.
• Site repairs: the smallest label sizes were raised to a 0.7 rem floor, the long-page table of contents collapses on narrow screens, the default social card's headline no longer runs to the edge, the homepage's scroll-reveal cards print correctly, and About leads with the work rather than the résumé.

2026-08-24#

• The demonstration corpus is public. The 45-document challenge publishes the MiniWorld fictional enterprise record: 45 sealed PDF documents (115 pages), a 55-question public suite with audited expectations and named failure modes, and a per-document integrity contract, all under CC0. The corpus is an available public artifact. No scored engine run against it has been published, and the flagship case remains withdrawn until it returns judged on this corpus.
• The glossary is now one page per term, so each definition has its own address.
• Every worked example on the site now runs on the published corpus: the home page walkthrough, the seven use-case patterns, the response envelopes in How it works, and the one-page brief all draw on the same 45 documents anyone can download. One example source, and it is checkable. The earlier Ardmore matter file is retired from the pages; dated entries below stay as they were written.

2026-08-23#

• The license rule behind the Packs ledger is now decided rather than queued. A source enters the program only under public-domain, CC0, attribution, permissive, or ShareAlike terms. Non-commercial-licensed sources are excluded in every tier, including internal evaluation. ShareAlike packs will always be packaged and published as separate downloads, and released packs will be hosted on this site with full attribution to the original datasets. Material whose terms allow verbatim use only, such as standards-document text, supports development but will not be hosted.
• The ledger adds Universal Dependencies Russian Taiga (r2.16, CC BY-SA 4.0) as a ShareAlike candidate. UD selection remains treebank by treebank under the per-treebank licenses.
• Controlled validation now exercises two contract behaviors on the controlled fixtures: a genuinely ambiguous question keeps its competing readings rather than having one chosen silently, and a question whose support is absent returns the named gap with its cause. This is controlled validation in a bounded fixture, not a production accuracy result.
• The response contract gains a distinction from the current product plan, decided today. When several readings compete for a single value, that is ambiguity: the readings stay open, and one is eliminated only on citable grounds, never silently. When the answer is legitimately a set (a count, an enumeration, a roll-up), the plurality is the answer's material and passes through untouched. Which kind of plurality a question carries is declared in the data, not guessed at runtime. This is contract language, not yet released behavior.
• Site repairs: the social-card generator no longer renders a card for the blog index itself, a stray the orphan check had caught; the Packs status table now defines the posture labels the inventory actually uses; and articles queued for publication were brought in line with the license rule above.

2026-08-21#

• The build-notes schedule was resequenced. The engineering essays (the round trip back to documents, response outcomes, conflicting records, entity resolution, attribution, controlled test worlds, reproducibility, refusal scoring, and lexical guardrails) now publish first, weekly from September 14. The product introduction follows in November, tied to the demonstration build it will be judged on rather than to a date the build has not yet met; if the build moves, the introduction moves with it, and the change will be recorded here. The executive brief follows the introduction by one week. No published article's text was edited, and the December and January articles keep their dates.
• The unpublished agent and API discovery files (OAuth metadata, an MCP server card, an API catalog, an agent-skills index, and the agent-facing auth page) were retired from the deploy to a non-deployed planned/ directory. Their endpoints do not exist yet, so publishing discovery documents for them implied a surface that is not there. They return with the software they describe. The stray /docs/api header went with them.
• Social cards for unpublished articles were removed from the deploy. The cards render article titles, which the scheduled-publication rule keeps private until the date. Site validation now requires a card to exist exactly when its article is published, and the daily publish workflow generates and commits each card on its article's day.
• Redirects for the retired /reports/ and /sample/ addresses are now real 301s instead of script stubs, anchors preserved.
• Validation gained checks for the voice rules that were previously unenforced: the em-dash ban, the banned marketing-words list, and the frontmatter placement of status_reviewed.
• Small corrections: a malformed structured-data logo URL; the comparison table now includes rows the neighbors win; the homepage names the four outcomes the way the contract does and states status as "in validation"; the pack-boundary diagram no longer draws a padlock over the private layer; and passages duplicated across the docs, use-cases, compare, and glossary pages were compressed or replaced with links.

2026-08-02#

• The site tree was consolidated. Reports and transforms is now a section of How it works, the sample documents are the case file on the use-cases page, and the Foundations series page was retired in favor of the blog index tracks. Old addresses redirect, so no link breaks.
• The blog index is organized into tracks: Foundations in numbered reading order, and dated Build notes as they publish. New series get their own track automatically.
• The top navigation's seven-item Technical menu was split into Learn (how it works, evaluation, glossary, compare, and the one-page brief) and Packs (all packs, sample documents, pack format), and the footer mirrors the same grouping.
• Long pages gained an "On this page" table of contents with a sticky rail and scroll-spy on wide screens (sub-sections stay collapsed except the current branch), section headings gained link anchors, and articles show an estimated reading time, a reading progress bar, and a closing keep-reading rail.
• Every wide diagram was re-laid out to render exactly at the article column width, and fourteen figures got alignment, spacing, and typography repairs. One structural fix: figures that had been silently broken into unpositioned text by a blank line inside their markdown block (the pack format diagram was the visible case). Site validation now fails if either pattern recurs.
• The site palette moved from bright blue/cyan to slate blue and sage; all brand assets were re-rendered (favicons, touch icon, OG cards, LinkedIn banners).
• Site internals: the build moved from Zola 0.19.2 to 0.21.0, the 404 page is now a named gap, Foundations posts show part-progress dots, anchor jumps briefly highlight the target heading, and the selection color matches the palette.

2026-08-01#

• The sample documents page is published: a complete fictional case file (lease, exhibit, amendment, side letter, policy, schedule, rider) with questions to try and one fully worked response envelope. It is the homepage walkthrough opened up, and the use cases page now shows every pattern on the same documents.
• The flagship case page was withdrawn. The acceptance case will return with the demonstration corpus it will be judged on, rather than as a description that drifts while the corpus is built. The dated entries below remain as the record of its earlier form.
• Use cases was rebuilt around the case file: each pattern shows the document excerpts used and set aside. The industry browser was removed; a short statement of what makes a document set a good fit serves instead.
• New reference pages: a glossary defining the terms this site uses, and a comparison with the neighboring approaches (retrieval-augmented generation, knowledge graphs and logic engines, LLM-extracted graphs, agent memory).
• The evaluation page now defines the controlled fixtures the site refers to, and pages and posts link the phrase to that definition.
• The blog gained a series taxonomy and series navigation, and four articles were written and scheduled for December and January: the reading step, the neighboring approaches, the cost of an answer, and corrections to the record. They are held until their publication dates.
• Site internals: one diagram shortcode replaced duplicated per-post figure markup, static asset URLs are now cache-busted by content hash, and OG card coverage is enforced at validation time.
• The response contract gains two qualifications from the current product plan. An answer may be returned narrowed: proven over a stated subset of the material in scope, with the narrowing receipt attached. And every response, even a full answer, carries a counted census of what remains unestablished, so a partial shortfall is reported rather than smoothed over. How it works states both.
• The runtime requirement strengthens: the recall path now also rules out learned weights and adaptive ranking, recorded on the technical overview. Pack builds are deterministic: the same input produces a byte-identical pack, and an answer names the exact pack snapshots it was derived from.
• The evaluation rules now require every release gate to ship with its negative control: a planted violation the gate must catch, shown to fail once before the gate counts.
• The glossary adds certified absence: a negative answer backed by a closure receipt over the extent that was covered.

2026-07-28#

• The product category is now stated as proof-carrying document intelligence, and the site leads with the capability rather than with the components the recall path avoids. The runtime requirement itself has not changed: no LLM, embedding model, vector database, GPU, or hosted inference in the recall path, recorded on the technical overview.
• The representation now has a public name. Temporal Evidence Graph is customer-facing language for an observable set of behaviors, described on the technical overview. It is not a file format, an interface, or a description of the engine.
• A new Reports and transforms page records the output side of the product: evidence-linked reports that preserve how a source represented its information, and document transforms that account for what was preserved, changed, recalculated, or omitted. This is a design target under the current product plan. It is not released behavior, and no public artifact demonstrates it.
• Use cases added two patterns that follow from that plan: reading a figure in the context of its row and column, and reissuing a document with an accounted difference.
• The flagship case acceptance bar now also requires the released artifact to return its result as an evidence-linked report, and the page adds a three-column frame separating what retrieval returns from what the documents establish.
• The pack contract preview added a record's position in its source document to the public contract and to the release checks.
• The evaluation rules now state that when a case expects a formatted report, the expected structure is part of the case and is scored separately from support.
• The blog was restructured around the same story. Foundations now runs to ten parts: document geometry was added as part 8, and a new part 9 covers the seams between the research areas, meaning the places where two of those literatures describe the same phenomenon in terms that do not line up. The series gives each field's primary sources and states what it leaves unsolved; how the pieces assemble is not published.
• Further articles were written and scheduled. They are held until their publication dates and are not listed here. The schedule was resequenced so that the research series completes before the product articles begin.
• One article was retired before publication. The counting-sort piece opened by disclaiming that it described the product and then taught a general algorithm, which does not meet the standard in EDITORIAL.md for writing from a specific decision or tradeoff. It was never published.
• Diagram corrections across the series: one diagram implied that a negated statement was stored as its opposite; three static diagrams rendered as dark panels regardless of theme and are now theme-aware; and the open-data diagram named two sources its article never listed.
• The worked examples were replaced. The hero, the homepage walkthrough, the response envelopes on How it works, and five of the seven use cases previously demonstrated recall over abstract records. They now run on documents, so that the examples exercise the capability the site claims: a figure read from the row that governs it, an amendment that supersedes that row, and a definitions clause that changes the unit of the answer.
• The flagship question was rewritten and its demands went from four to six, adding document geometry and the form of the result. The previous question tested identity, time, connection, and conflict, all of which it tested well, but it could be answered without reading a document structure at all. The acceptance bar now requires selecting the correct row of a tiered table under the version that governed on the day.
• The reports page and its companion article were de-duplicated; the page states the contract and the article argues for it. An article whose URL contradicted its own thesis was renamed before publication.
• About now explains why this became a document problem rather than a text problem, and How it works answers what a document buyer asks: formats, scans, and the boundary against a document management system.
• The earlier positioning below, recorded on 2026-07-14, is superseded rather than removed. This record is additive.

2026-07-14#

• The flagship case now focuses on the four demands of the question (identity, time, connection, and conflict) instead of a broad product comparison table.
• The homepage and Use cases now position CleverMemory as a local-first memory layer for any application, with examples across regulated industries, operations, healthcare, technology, and the public sector.
• The technical overview now explains the machine-checkable mathematical proof carried by a supported answer, along with the limits of what that proof establishes.
• Navigation, page headers, top-bar height, title position, and mobile menu behavior now use one shared layout.
• The About page now includes education, skills, career history, and the project motivation in a scannable layout.
• The site mark was refined to r9: the answer dot is now node-sized at the hex-center of a 2-3-2 grid, surrounded by a dashed cyan ring that marks where the missing piece was. All OG cards, the LinkedIn banner, and the monochrome variant were regenerated.
• Copy across the homepage, specification preview, docs, methodology, and use-cases pages was revised to reduce formulaic hedging and passive constructions.

2026-07-13#

• The Packs ledger now names each selected third-party source and its license or unresolved terms, grouped into friendly, ShareAlike or restricted, and proprietary boundaries. No public pack release is implied.
• The technical preview now records two product requirements: the recall path does not use an LLM, embedding model, vector database, GPU, or hosted inference. The runtime is designed for commodity CPU hardware. Supported configurations and measured limits remain pending release validation.

How changes will appear#

Material changes to availability, validation state, the public contract, pack decisions, or evaluation evidence will be dated here and reflected on the affected page. Dated essays continue to publish through the blog and Atom feed; this record also covers status-page changes that are not blog posts.

No benchmark, cost, latency, scale, accuracy, coverage, or resource number will be listed as a result until its reproducible method and supporting artifacts are ready for review.