Review gates

If you've gone through a run or two, you've already seen both gates in action — the pipeline halts after the filter completes and again before briefing synthesis, each time showing you intermediate results. Both gates are on by default. You can click through any single one without turning it off, and both can be disabled entirely in Settings once you stop catching things at them.

This page covers what each gate shows, what to do there, and when it's worth turning one or the other off.

Where they sit in the pipeline

  1. Fetch papers
      │
  2. Filter papers      (YES / MAYBE / NO triage)
      │
      ⏸   Gate 1 — review the buckets, fix obvious errors
      │
  3. Score abstracts    (0–10 + justification)
  3.5. Post-process     (optional consistency pass)
      │
      ⏸   Gate 2 — review scores, star papers for deep analysis
      │
  4. Analyze PDFs       (full-text read of the top N)
      │
      ⏸   Gate 3 — star / dismiss / comment before synthesis
      │
  5. Briefing           (editorial synthesis + hallucination audit)

All three gates pause a running pipeline until you click a Continue → button. They don't cost tokens — skipping them doesn't save money, only time — but they do catch bad runs early.

Gate 1 — after the quick filter

What it does

The quick filter (Stage 2) is cheap: the filterModel slot (default gemini-3.1-flash-lite) runs YES / MAYBE / NO triage on every paper fetched from arXiv, along with a one-sentence summary and a short justification per paper. Its job is to cut the volume before you spend real tokens on scoring and deep analysis.

When pauseAfterFilter is on (default: on), the pipeline halts immediately after the filter stage. The progress timeline shows "Filter complete — review results before scoring" and the main area fills in with the filter results, grouped into three buckets.

What you see

Each paper in the list shows its title and authors, the filter's one-sentence summary (italic), and a "Verdict reasoning:" line with the model's justification. On the right are three verdict buttons — YES, MAYBE, NO. The current verdict is filled; the other two are outlined. Clicking a different one moves the paper to that bucket and marks the choice with a small ⇄ to indicate it was overridden.

Above the buckets sits Continue to scoring →, which advances the pipeline with whatever verdicts are in effect at that moment, overrides included.

What to do there

If the filter got something right, do nothing. Most verdicts will be fine — the filter model is decent at its job.

When the filter gets something wrong, click a verdict to override. Moving a paper from NO to YES or MAYBE brings it into scoring — do this when you think the filter missed something relevant. Moving a paper from YES to NO excludes it from scoring, which saves you tokens on something you already know isn't a fit.

Every override is logged as a filter-override feedback event. These events act as scope signals downstream: a pattern of NO → YES overrides suggests your profile is too narrow in that area; the reverse suggests it's too broad. The refinement flow sees them. See Giving feedback → The filter-review gate for the full mechanics.

In practice

A few patterns come up repeatedly:

• The filter dismisses a paper adjacent to your interests but not a direct match. Override to MAYBE and see what deep analysis makes of it. This is the most common case.
• The filter accepts a paper whose abstract is mostly marketing. Override to NO to skip spending tokens on a likely dud.
• The filter is confused by a paper straddling multiple fields. Override to match your actual interest and note whether it's a symptom your profile could address.

You don't have to review every paper. Overrides usually happen in the YES and MAYBE buckets; the NO bucket can often be left alone unless something specific catches your eye.

Gate 2 — before PDF analysis (score review)

What it does

After Stage 3.5 (post-processing), the pipeline has a ranked list of scores but hasn't spent any tokens on deep PDF reads yet. That's the cheapest moment to adjust which papers get the full treatment.

When pauseBeforeDeepAnalysis is on (default: on), the pipeline halts here. The timeline shows "Scoring complete — review scores and adjust PDF selections", and the main area shows the full scored list split into three collapsible groups.

What you see

Three groups, each collapsible:

• Will PDF (top-N) — expanded by default. These are the papers that would go to Stage 4 as things stand: the top N by score, where N is your Papers to Analyze setting (default 30).
• Borderline — collapsed by default. The next chunk of papers after the top-N cutoff, up to min(maxDeepAnalysis, 50) papers. These are the candidates most worth inspecting if you think the default cutoff is landing in the wrong place.
• Low score — collapsed. Everything below the borderline band.

Each row in the list shows the paper's title and authors, its score (pre-analysis), the scoring model's justification, and a ⊘ duplicate badge when applicable. On the right of each row sit three controls:

• ★ (star) — guarantees this paper gets PDF-analysed, regardless of where its score sits.
• ⊘ (exclude) — removes this paper from PDF analysis for this run.
• 💬 (comment) — opens a small text box to attach a note to the paper. Does not change whether the paper goes to PDF.

The selection logic

The final set of papers sent to Stage 4 is additive:

PDF set = (top-N by score) ∪ {papers you starred} − {papers you excluded}

Stars can reach into the Borderline or Low score groups and pull papers up. Excludes can remove papers from the top-N. If you star several borderline papers, the final PDF set may exceed maxDeepAnalysis — that's intentional, not a bug: your explicit choices override the automatic cutoff.

Starred papers carry a guarantee all the way through: even if a PDF-analysed paper's finalScore ends up lower than the finalOutputCount cutoff, starred papers always appear in finalRanking and therefore in the briefing.

Round-level note

Below the paper list sits a free-text field labelled "+ feedback on this scoring round". This is for observations about the scoring run as a whole — "scores feel compressed, everything is bunched between 5 and 7", "filter let through a lot of tangential ML papers this time". It isn't attached to any single paper; instead it flows into the next profile-refinement call as a run-level observation. Leave it blank if nothing's worth noting.

Skip remaining gates this run

A small Skip remaining gates this run link appears near the bottom of this gate, and on Gate 1 and Gate 3. Clicking it bypasses all subsequent default-on gates for the rest of this run — useful when you want a single unattended run without changing your persistent settings. It does not modify any setting; the next run will still pause normally.

Click Continue to PDF analysis → when you're done. Stage 4 starts immediately with the adjusted paper set.

Gate 3 — before briefing synthesis

What it does

After Stage 4 (PDF analysis) completes, the pipeline has scores, deep-analysis summaries, and a ranked list of papers. The next step is synthesis: the briefing model reads your profile, the analysed papers, and your feedback, then writes the executive summary, themes, and paper cards.

When pauseBeforeBriefing is on (default: on), the pipeline halts before synthesis runs. The timeline shows "Analysis complete — review results and add stars/dismissals before generating your briefing", and the main area shows the Analysis Results list with every PDF-analysed paper, score, and summary.

What to do there

Star the papers you care about and dismiss the ones you don't. Both signals feed directly into the synthesis prompt: starred papers anchor themes and get the richest "why it matters" treatment; dismissed papers get deprioritised or explained away; everything else is handled according to its relevance score. Leave a + comment on a paper to have your note woven into its paragraph in the briefing.

Stars given here have a different meaning than stars given at Gate 2. A star at Gate 2 says "analyse this paper's PDF"; a star at Gate 3 (or on a rendered briefing) says "I liked this paper". Both are stored as star events — the meaning is contextual based on when you gave it.

The difference between Gate 3 and the same controls on the rendered briefing is timing. Feedback given here shapes the briefing that's about to be written. Feedback given while reading the briefing is recorded against the paper but only affects future runs — via the refinement flow, where paper-grounded signals are among the strongest inputs the profile revision sees.

Click Continue to briefing → when you're done. Synthesis runs, then the hallucination audit runs, then the briefing renders on its own page. The results list stays interactive while synthesis is in flight.

Progression: from on by default to off

All three gates start on because they catch bad runs early — especially useful in the first week or two, when you're still developing a sense of what your profile should say and what the pipeline does with it.

After five or ten successful briefings, the gates can start feeling like more friction than signal. At that point it's worth considering which ones still earn their keep:

• Turn off pauseAfterFilter first. The filter is usually right enough that override-saving is marginal after a while, and scoring is cheap to over-run.
• Keep pauseBeforeDeepAnalysis on while you're still calibrating. This is the best place to rescue papers the scorer ranked low but that you know from the abstract are worth reading in full — and to exclude obvious duds before they consume expensive PDF tokens.
• Keep pauseBeforeBriefing on longer. This is where your stars, dismisses, and comments actively steer the synthesis. If you're not giving feedback at this gate, the briefing tends to drift toward generic.
• All three off once you trust the setup. Runs become fully unattended: click Start Analysis, come back later, read the briefing.

There's no harm in leaving all three on permanently if the workflow suits you. The gates are UI-only — they don't cost tokens, and skipping them doesn't save money.

The Skip remaining gates this run link on each gate offers a middle path: you can engage with one gate and bypass the rest without touching any settings. Useful for days when you just want a quick check of the filter output before letting the rest of the run go unattended.

When to re-enable a gate you'd disabled

Three situations in which flipping a gate back on tends to pay off (applies to all three gates):

• You made a big profile change. A freshly rewritten profile may be poorly calibrated, so the filter may be over- or under-selecting. Turning pauseAfterFilter back on for a few runs catches this early.
• You added new arXiv categories. You haven't built a mental model of what comes through the filter in those categories yet.
• You switched to a much cheaper (or much more capable) model. Model behaviour changes, and the first few runs with a new model are worth watching.

Broadly, the gates are worth having on while you're still calibrating the pipeline to your interests, and worth disabling once a few runs in a row have landed well without intervention.

Where to toggle them

All three gates live in Settings under Review & confirmation, as checkboxes:

• Pause after filter to review overrides — controls Gate 1 (pauseAfterFilter).
• Pause before deep analysis — controls Gate 2 (pauseBeforeDeepAnalysis).
• Pause before briefing to review scores and add feedback — controls Gate 3 (pauseBeforeBriefing).

Two sibling checkboxes nearby — Auto-retry briefing if hallucination check returns YES (on by default) and …MAYBE (off by default) — govern the briefing retry loop, not the pause gates.

Toggles are disabled while a pipeline run is in progress (you can't change gate behaviour mid-run), and changes take effect on the next run.

No per-run toggle — but there's a shortcut

Settings are global, not per-run. For a one-off unattended run without changing settings, use the Skip remaining gates this run link on any gate — it bypasses all subsequent default-on gates for that run only.

Next

Tuning the pipeline → — past the gates, the rest of the knobs sit in Settings: model slots, batch sizes, concurrency, score thresholds, and retry behaviour.

Also worth reading:

• The feedback detail behind overrides, stars, and dismisses. → Giving feedback
• What happens at each stage, not just at the gates. → Pipeline