🗺 Feature roadmap: what is still open, and what was closed unbuilt

Status: proposal.

Written for someone picking up the next PR, or checking whether an idea has already been weighed. Assumed: ARCHITECTURE.md for the four layers, DRAFT-ITERATION.md for the dossier, and PLAGIARISM.md for the three detection tiers. Not covered here: how the existing detectors work (PLAGIARISM-DESIGN.md), and figure work, which FIGURE-ROADMAP.md owns.

This page lists only work that is not built. A shipped item is deleted from it rather than marked done; the PR that closed it and plans/ record how it was built. A roadmap that keeps its own history stops being a list of what to do next.

🧭 Table of contents

• What exists
• Open items
• Closed without building
• Five constraints every item respects
• What is deliberately not proposed

✅ What exists

FEATURES.md is the capability surface as built, pinned to the code by a test. Everything this roadmap's themes asked for is in it: claim-first evidence (claim:/quote:) and the evidence sidecar, multi-source synthesis and the pre-gate self-feedback step in the genre skills, the review aids (provenance, quotation, uncited, support, union and the rest), the TikZ scaffolds and figure checks, outline-driven drafting and the draft fingerprint, the agenda loop, and topic modelling and discovery (TOPIC-MODELLING.md, TOPIC-DISCOVERY.md).

The synthesis half drew on OpenScholar and the figure half on PaperBanana, both as ideas only. INSPIRATION.md records what was taken and what was refused.

▶ Open items

A new item goes here as one row once it has been proposed and costed. "One PR" is the unit throughout.

Item	Theme	Size	Depends on
Figure work: the remaining figure-drawer items and the parts not yet built	Figures	see there	FIGURE-ROADMAP.md owns the list and the order

Check a new proposal against two rules before costing it. An item that makes something run without a person asking needs the amendment in constraint 2. An item that has a skill repair its own output is refused by the self-marking argument in AUTO-IMPROVEMENT-RATIONALE.md.

🚫 Closed without building

Each was weighed and closed. The reason is kept so the item is not re-proposed as an oversight. "Revisit if" names the evidence that would reopen it.

Item	Closed as	Why	Revisit if
A1b: route verbatim scan findings into agenda-reviser automatically	Declined	A genre skill repairing its own output is marking its own homework, and agenda-reviser runs only when a person asks for it (R11 in AUTO-IMPROVEMENT.md). The scan runs and reports; the person decides whether to invoke the repair	Never on convenience alone; it needs a change to R11
B3: open each section with a one-sentence thesis and a source count	Not a priority (#379)	The design is in the issue	Someone wants it; start from the issue
C4: report a prose numeral with no traceable origin	Deferred as mis-specified	An adjacent citekey is the only traceable origin with instances, so the aid would mark the motivating case (a magnitude beside a real citation) as traced and report its complement. Every dossier here records textbook-chapter, where numerals are invented worked examples, so the base rate is zero. plans/c4-numeral-as-claim.md has the measurements	A survey, thesis chapter or deep-research draft with a dossier exists; or the finding is redefined as the cited case, with a passage narrow enough to check the number in
C6: measure negative rejection (how often a draft asserts what the corpus cannot support)	Skipped by evidence	On a single-subject corpus, neither construction of a negative query set works: unparsed bib entries and held-out shelves still match parsed papers on the same topic. No logged retrieval call has returned fewer results than asked, so the rate would be zero over zero	The corpus spans more than one subject, or a hand-authored negative query set is accepted as ground truth
D4: an optional vision critique of rendered figures	Skipped by evidence	A full visual pass over the 43 figures of this project's drafted book found nothing beyond what review figure already reports. Routing defects are measured by figure check's connector rules	A denser figure set where the checks come back clean and a reader still disagrees
D5: two figure checks computed from source	Closed, not planned (#624)	The point-size half is figure check's type-below-floor rule. Whether two declared colours stay apart in a greyscale print is still a human judgement on TIKZ-STYLE.md's pre-flight list	Someone wants the greyscale half; the design is in the issue
F4: a blocking overlap gate	Declined	Measured on this project's own 178,000-word book, no span-length threshold separated the one genuine violation from correctly quoted passages several corpus papers share	A corpus of real rather than planted reuse, or a version-controlled seed allowlist

🔑 Five constraints every item respects

Each one has already ruled out an obvious design.

1. No LLM output may reach the corpus plane. SOUL.md: the corpus layer "has no LLM and no judgment calls", and the enrichment layer "reads the ledger and never writes it". Extracted claims live in the dossier, and the per-citekey TL;DR (chitragupta draft tldr, FEATURES.md) lives in the drafting layer, in neither corpus nor enrich. Writing either into the ledger would break "same bibliography in, same citekeys out".

2. No second gate. chitragupta draft gate means one thing: a fabricated citekey fails. WRITING-STANDARDS.md §10 says a second meaning "would blunt the first". Everything else is a mandatory step in a skill or an advisory review aid, and the review layer's rule is:

a review finding may be read, may be invoked by a driver, and may never block a draft.

The line is advisory versus blocking, not manual versus automatic. The rule is stated in the review layer and the documents describing it, not in SOUL.md. Making a review aid run without a person asking is a user decision, not an engineering one. AUTO-IMPROVEMENT-RATIONALE.md has the reasoning.

3. Anything with a torch/transformers dependency goes behind an extra. pyproject.toml makes bibtexparser the single core dependency as a point of design, with all ML in the enrich group. The cross-encoder reranker follows this rule.

4. Attribution is owed for the idea, not for the text. Nothing is copied from an upstream, so no licence notice obligations attach. INSPIRATION.md carries the credit. You may read an upstream; you may not paste from it.

5. R1-R12 bind every new review aid and every unattended edit. AUTO-IMPROVEMENT.md states them. Four are easy to breach by accident:

• R2: every finding carries an identity stable across runs, so "this finding is gone" is decidable.
• R3: "An unattended item's check is binary. No continuous score is ever the thing being optimised." Annotate every new aid binary (an agenda may consume it and a loop may act on it) or continuous (a human reads it; nothing acts on it unattended).
• R4: after an accepted edit, every aid re-runs and the result is compared by finding identity as well as by count. The repaired item is resolved, no objective-class finding is new, and the total has not risen, else the edit reverts.
• R10: a new aid is registered in both review.AIDS and review._registry.AIDS, and appears in AGENTS.md, CLI.md, the README tables and mkdocs.yml. review/_registry.py raises RuntimeError when the two dicts disagree, so a half-registered aid fails at import. A missing mkdocs.yml entry is the silent one: missing nav is INFO, not a --strict failure.

One naming rule comes with them: the judgement register belongs to the gate. An advisory aid may not be called audit, reckoning, verdict or ruling. triage is also taken: REJECTION.md records a retrieval stage of that name built and withdrawn.

🚫 What is deliberately not proposed

Recorded so each is not re-proposed as an oversight.

Not proposed	Why
A blocking overlap gate	Declined on measured evidence (F4 above), and a second meaning would blunt the gate's one meaning (WRITING-STANDARDS.md §10)
Claim extraction cached in the corpus or enrichment layer	LLM output on the corpus plane; breaks "same bibliography in, same citekeys out" (SOUL.md)
TL;DR shown in corpus ledger output	Same, plus it inverts the layer order; see FEATURES.md's per-citekey TL;DR section
An LLM-written TL;DR for a paper with no abstract	4.92M input tokens for 160 documents, re-incurred on every re-parse, and SOUL.md requires a human to accept anything abstractive. There is no tldr accept, no review surface, and no answer to who reads 160 machine summaries. Such a paper reports "abstract not available" instead. TLDR.md has the measurements
Abstractive topic summaries	Abstractive models carry factual inconsistencies in up to 30% of outputs. A topic summary asserting a claim no paper made is the same failure class as a fabricated citekey
An LLM transcribing document structure	Span selection (offsets to keep) is safe; span transcription is not, because a transcribed reference can be a fabricated one. GROBID-CITATION-GRAPH.md is the starting point if structural extraction is wanted
Image-generated figures	Non-deterministic, against the byte-identical output rule, and a raster cannot satisfy §10's two-form TikZ/ASCII contract. PaperBanana's own side-by-side shows the image route inverting a relationship, duplicating a category and drawing a bar taller than its gridline where the code route was correct
Any ML dependency in the core	bibtexparser as sole core dependency is a design decision in pyproject.toml, not an accident
Fetching papers from arXiv or anywhere else	Admission is the reference manager's job alone (AGENTS.md)
Copying from zotero-arxiv-daily	AGPLv3, and excluded by the project owner
Positional [n] citation numbering	Upstream's scheme; reordering the passage list silently changes every citation's meaning. This project has real citekeys
OpenScholar's dead code paths	Its LLM reranker (process_ranking_results), final_processing, and --use_abstract are defined and never called; its --norm_cite adds a bounded [0,1] bonus to an unbounded cross-encoder logit