A paper has more than words.
A strong paper can teach how to pose a question, order evidence or make a comparison visible. InkTide records the job each choice performs and the conditions under which it transfers. A new manuscript still begins with the author's own findings, data and intended audience.
The question, evidence, paragraph order and figure all shape what a reader can understand. InkTide draws transferable lessons from inspected papers and applies relevant guidance to new author materials. The output keeps claims tied to those materials.
Evidence-led argument
Connect the question, unresolved obstacle, proposed approach, discriminating evidence and justified implication.
English with a clear purpose
Inspect what each sentence contributes, preserve the force of claims and remove repetitions that add no information.
Figures that expose the comparison
Align panel order, encoding and captions with the scientific question, retaining units, transformations and uncertainty.
Separate evidence from inspiration
Claims point to author evidence; writing and design choices point to inspected examples. A stylistic lesson cannot fill an evidence gap.
Different roles. One evidence trail.
Author evidence defines the task. Inspected lessons inform the choices. Production and independent review meet through an explicit handoff.
Coordinator
Turns the goal and evidence into bounded assignments, manages dependencies and integrates the final result.
Curator
Inspects the argument and figures in source papers, recording useful choices and where they stop applying.
Writer
Builds the section's argument from supplied evidence and adapts relevant lessons into original English.
Figure designer
Makes each panel serve a clear comparison or inference, using actual data and an original composition.
Independent reviewer
Receives frozen outputs and raw evidence, checks fidelity and readability, and returns located findings for repair.
Read carefully. Reuse thoughtfully.
The collection spans machine learning and vision, life sciences, physical sciences, statistics and social sciences, and computing systems. Candidate discovery includes awards, time-tested work, classics and labeled curatorial suggestions. Writing and visual value are assessed separately from recognition.
The current snapshot records 466 main-paper PDFs, 459 distilled reading cards and 416 admitted writing-study references. Reading scope, source identity and transfer limits stay with a lesson. Author annotations are still pending; personal-style calibration depends on real feedback.
Put the writing to the test.
The whole-manuscript study generated 26 complete drafts through real native writer contexts, using fixed synthetic research facts. It recorded 17 primary paired reviews and five position-reversed secondary reviews. The held-out set registered five groups: four were written and reviewed, while one was blocked before writing.
InkTide was preferred in three tested groups, but only two met the strict quality-win criteria. One preferred pair had a shared-input conflict; another passing pair tied. All four English-score pairs tied. These records support specific argument and repair cases rather than a general superiority claim.
Inspect the full comparison and its boundaries →A change worth seeing.
HOLDOUT-01
Separate observed agreement from an exact argument
A synthetic deletion audit asked whether removing copied descendants restored a prototype classifier to a clean-refit reference. Row-only removal disagreed on 12 of 200 test predictions; complete lineage removal had zero disagreements. InkTide made clear that the finite test and the deterministic operation argument do different work: identical remaining rows, arithmetic order and tie rules establish identical class means and prediction rules for this learner. It also retained the incomplete-graph failure and the 72 ms versus 66 ms recomputation cost. The independent reviewer preferred this clearer evidence structure; both drafts passed quality gates and scored 5/5 for English.
Read the case →DEV-04
One missing quantifier can change a theorem
The finite-set case distinguishes selecting a measured winner from certifying the unique true minimizer. In round 2, the guided conclusion reduced the condition to one unspecified competitor gap above 2 epsilon. That loses the requirement that the minimum gap—or every competitor gap—strictly exceed the threshold. The reviewer failed its claim-scope gate. Round 3 restored the minimum/every-competitor condition in the conclusion and made the strict test, uncertified branch and singleton exception explicit in Methods. Both drafts then passed, and the reviewer preferred InkTide because the procedure could be implemented from that section without assembling rules from elsewhere.
Read the case →DEV-01
Fluent English can still make the wrong inference
In the regression-router replay, both round-1 drafts scored 5/5 for English. The guided conclusion nevertheless attributed improved accuracy to more expensive calls and treated an early-versus-late diagnostic as a dependence. The supplied ablation had the same 35% call rate but a different error, and the temporal cause remained unknown. The baseline preserved those limits and was preferred; the guided draft failed causality and claim-scope gates. After two real repair rounds it reached a tie, not a win.
Read the case →From author materials to a reviewable result.
A task begins with drafts, facts, data and the intended audience. Writing produces original prose and an evidence map; figure work produces original visuals, editable files and captions. A separate reviewer reads the frozen result against the evidence and locates repairs.
The project operates as a Codex-native skill and role workflow. Its current scope is source-backed assistance and continuing evaluation. Private paper references and project implementation are retained privately.