Skip to content

Convert Claude Code or Codex into the most intelligent Deep Research Agent. Collect, search, and synthesize web research into a persistent, searchable wiki that builds on itself. Hosted API + MCP: hyperresearch.ai

fix: stop the synthesizer trading coverage and mechanism depth for cleaner prose (#56)

Jordan Gibbsca052fadb402
diff --git a/CHANGELOG.md b/CHANGELOG.md
index fd9d8ba..b680069 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -2,6 +2,15 @@
## [Unreleased]
+### Coverage before elegance: the synthesizer stops trading substance for prose
+
+A rerun of a comparison query scored below its own older baseline on comprehensiveness and insight, and reading both reports side by side showed why: same word count, spent differently. The humanization work (primers, calm citations, one committed thesis) had quietly taught the synthesizer to dissolve a systematic ten-axis comparison into an elegant narrative and to compress a fully developed quantitative mechanism down to a one-line mention. Cleaner to read, and worse on exactly the axes a "compare X, Y, Z" prompt is graded on. The prompts were conflating two different edits: cutting prose, which is right, and cutting substance, which is not.
+
+- **The synthesizer now treats coverage and mechanism depth as content, not optional structure.** On a compare or survey task, every decision-relevant dimension the corpus names gets explicit coverage, and a mechanism the sources develop (a named decomposition, a formula with its terms, a causal chain with its numbers) gets developed in the report rather than gestured at. The rule it now follows: elegance is spent on the words between points, never on the number of points.
+- **"Selectivity" is now defined precisely.** The anti-sprawl guidance used to read as license to drop points; it now says selectivity means choosing which sources to cite for a point, not which points to make. Dropping a comparison axis or compressing a mechanism is a coverage gap, not economy.
+- **The length pass cuts prose, never points.** When trimming to the word ceiling, redundancy and filler go first; a comparison dimension, a developed mechanism, a counterargument, and a load-bearing primary source are off limits. If prose cuts do not get under the ceiling, the report has too many words per point, not too many points.
+- **Two adversarial backstops.** The depth critic now flags a developed quantitative mechanism compressed to a bare mention as the highest-insight loss a draft can take. The instruction critic gains a comparison-axis coverage check (register-independent, since a comparison prompt needs its axes covered whatever the register) that names dropped or compressed dimensions.
+
### Run levers: auto-selected register, domain notes, and inference depth
The pipeline had one hard-coded voice (evaluative-argumentative), so a "teach me X" or "survey the landscape" query got an opinionated verdict report. The Q62 register experiment showed the judge prices register heavily (+2.5 RACE from register alone), so register is now a run-time lever instead of a constant.
diff --git a/src/hyperresearch/core/hooks.py b/src/hyperresearch/core/hooks.py
index 2815f50..b180aca 100644
--- a/src/hyperresearch/core/hooks.py
+++ b/src/hyperresearch/core/hooks.py
@@ -773,6 +773,14 @@ prompt. No block = this prompt's defaults apply unchanged.
interim note actually provides
- A named mechanism is mentioned but not explained even though the
interim note explains it
+ - **A developed quantitative mechanism is compressed to a bare mention.**
+ The interim note works a named decomposition (e.g. a factor model with
+ its loadings), a formula with its terms, or a specific causal chain with
+ its numbers — and the draft names it in one clause and moves on. This is
+ the highest-insight loss a draft can take: the mechanism was the whole
+ reason the depth budget was spent, and a gesture at it recovers none of
+ that value. Flag it `major` and tell the revisor to develop it to the
+ depth the interim note supports.
- A comparison between sources is summarized but the actual
disagreement is blanded out
- A citation is dropped where the interim note specifically supports
@@ -1230,6 +1238,23 @@ matters here) before the judgment starts — emit ONE finding:
- `recommendation`: name the sections that need a primer and, for
each, the concept the primer should teach
+**Check R6: Comparison-axis coverage.**
+If the report compares 3+ entities (the same trigger as R4) AND the
+prompt-decomposition's coverage matrix or required items name
+decision-relevant comparison dimensions that the report omits entirely
+or compresses to a passing mention (dissolved into a thesis instead of
+worked explicitly), emit ONE finding:
+ - `failure_mode`: `"missing-comparison-dimensions"`
+ - `severity`: `major`
+ - `recommendation`: name the dropped or compressed axes and suggest
+ giving each explicit coverage (a table row plus a sentence of why it
+ matters). A "compare X, Y, Z" prompt is scored on how many
+ decision-relevant dimensions the report actually works.
+This check is register-INDEPENDENT — a comparison prompt needs its axes
+covered in analyze, survey, and advocate alike — so apply it regardless
+of the Run directives register (the register guard below governs only
+committed-ranking demands, which is a different thing from axis coverage).
+
**Cap:** At most **3** readability-structural findings total. Do not
let these crowd out core instruction-following findings. Use
`"readability-structural"` as the `atomic_item` prefix for these.
@@ -2176,6 +2201,29 @@ permitted to be uneven — pass 2 cleans it up. Goals for pass 1:
primer is not filler; it is the on-ramp that makes the density that
follows legible. Skip it only for the executive summary and for
short connective sections with nothing new to explain.
+12. **Coverage and mechanism depth are load-bearing content, spent before
+ elegance.** Two failure modes cost more points than any prose flaw, and
+ both hide as "tightening":
+ - *The comparison surface.* On a compare/survey task, the systematic
+ multi-dimensional comparison IS content, not optional structure. Every
+ dimension the corpus treats as decision-relevant — read them off
+ `comparisons.md` and `source-tensions.json` — gets explicit coverage:
+ a table row and a sentence of why it matters. Do NOT dissolve a
+ ten-axis comparison into a single elegant thesis that gestures at the
+ axes. A "compare X, Y, Z" prompt is scored on how many decision-relevant
+ dimensions you actually work, so a narrative that absorbs the comparison
+ covers less than a report that lays it out.
+ - *Developed mechanisms.* A quantitative mechanism the sources develop —
+ a named decomposition (e.g. a factor model with its loadings), a formula
+ with its terms, a specific causal chain with its numbers — must be
+ DEVELOPED in the report, not compressed to a one-line mention. A
+ mechanism named but not unpacked has lost the exact insight that made it
+ worth citing; it reads as a gesture, and the insight score reflects that.
+ When an interim note works a mechanism in depth, carry that depth
+ through — that is what the depth budget was spent on.
+ Elegance is spent on the words BETWEEN points, never on the number of
+ points. If you must choose, on these tasks coverage and developed depth
+ win over a cleaner line.
Pass 1 length target: in the response_format range, leaning slightly long
(15-20% over target). Pass 2 cuts.
@@ -2192,7 +2240,12 @@ point is a forced compression rewrite of your own output. Count your words
before finishing pass 2; if you are over the ceiling, cut until you are
under it. A large corpus is never a reason to exceed the ceiling —
selectivity is the skill being graded, and burying the argument under
-every available source scores WORSE on insight, not better.
+every available source scores WORSE on insight, not better. But be exact
+about what selectivity means: it is choosing which SOURCES to cite for a
+point, not which POINTS to make. Cutting a comparison dimension or
+compressing a developed mechanism to a bare mention is not selectivity —
+it is a coverage and insight gap that scores worse, not better. Prune
+redundant citations of the same point; never prune the point.
When pass 1 is done, write it to `pass1_output_path`.
@@ -2238,7 +2291,7 @@ drafts. State the committed position from the synthesis plan.
### Length discipline
-If pass 1 is over the response_format target, CUT. Specifically:
+If pass 1 is over the response_format target, CUT. Cut prose, never points:
- Cut the most redundant sentences first (you've already flagged them above)
- Cut filler ("It is worth noting", "Importantly", "Of note,", "It bears
mentioning")
@@ -2246,6 +2299,12 @@ If pass 1 is over the response_format target, CUT. Specifically:
is restating
- Drop weak adverbs ("really", "quite", "notably" when not load-bearing)
+What you NEVER cut to hit the target: a comparison dimension, a developed
+quantitative mechanism, a counterargument, or a load-bearing primary source.
+Those are points, not prose. If cutting redundancy and filler does not get
+you under the ceiling, the report has too many words per point, not too many
+points — tighten more points, do not amputate one.
+
If pass 1 is under target, EXPAND. Specifically:
- Add interpretive beats where you have factual claims without
conclusions
diff --git a/tests/fixtures/golden_prompts/agents/depth_critic_agent.md b/tests/fixtures/golden_prompts/agents/depth_critic_agent.md
index 08661bf..63a20d3 100644
--- a/tests/fixtures/golden_prompts/agents/depth_critic_agent.md
+++ b/tests/fixtures/golden_prompts/agents/depth_critic_agent.md
@@ -66,6 +66,14 @@ prompt. No block = this prompt's defaults apply unchanged.
interim note actually provides
- A named mechanism is mentioned but not explained even though the
interim note explains it
+ - **A developed quantitative mechanism is compressed to a bare mention.**
+ The interim note works a named decomposition (e.g. a factor model with
+ its loadings), a formula with its terms, or a specific causal chain with
+ its numbers — and the draft names it in one clause and moves on. This is
+ the highest-insight loss a draft can take: the mechanism was the whole
+ reason the depth budget was spent, and a gesture at it recovers none of
+ that value. Flag it `major` and tell the revisor to develop it to the
+ depth the interim note supports.
- A comparison between sources is summarized but the actual
disagreement is blanded out
- A citation is dropped where the interim note specifically supports
diff --git a/tests/fixtures/golden_prompts/agents/instruction_critic_agent.md b/tests/fixtures/golden_prompts/agents/instruction_critic_agent.md
index 81abac6..769ad69 100644
--- a/tests/fixtures/golden_prompts/agents/instruction_critic_agent.md
+++ b/tests/fixtures/golden_prompts/agents/instruction_critic_agent.md
@@ -256,6 +256,23 @@ matters here) before the judgment starts — emit ONE finding:
- `recommendation`: name the sections that need a primer and, for
each, the concept the primer should teach
+**Check R6: Comparison-axis coverage.**
+If the report compares 3+ entities (the same trigger as R4) AND the
+prompt-decomposition's coverage matrix or required items name
+decision-relevant comparison dimensions that the report omits entirely
+or compresses to a passing mention (dissolved into a thesis instead of
+worked explicitly), emit ONE finding:
+ - `failure_mode`: `"missing-comparison-dimensions"`
+ - `severity`: `major`
+ - `recommendation`: name the dropped or compressed axes and suggest
+ giving each explicit coverage (a table row plus a sentence of why it
+ matters). A "compare X, Y, Z" prompt is scored on how many
+ decision-relevant dimensions the report actually works.
+This check is register-INDEPENDENT — a comparison prompt needs its axes
+covered in analyze, survey, and advocate alike — so apply it regardless
+of the Run directives register (the register guard below governs only
+committed-ranking demands, which is a different thing from axis coverage).
+
**Cap:** At most **3** readability-structural findings total. Do not
let these crowd out core instruction-following findings. Use
`"readability-structural"` as the `atomic_item` prefix for these.
diff --git a/tests/test_core/test_prompt_golden.py b/tests/test_core/test_prompt_golden.py
index 51ae221..bdd254e 100644
--- a/tests/test_core/test_prompt_golden.py
+++ b/tests/test_core/test_prompt_golden.py
@@ -73,6 +73,16 @@ Deliberate deviations already folded into the goldens (2026-07-19):
register-conditional standards. The cite-checker and
9-evidence-digest are deliberately untouched (no shim: verification
is register-independent; step 9 spawns nothing).
+ - Coverage before elegance (2026-07-22, after the Q52 rerun scored below
+ baseline on comprehensiveness and insight by trading the systematic
+ comparison surface and a developed factor decomposition for a cleaner
+ thesis, at the same word count): the depth critic gained a shallow-spot
+ bullet for developed quantitative mechanisms compressed to a bare mention
+ (the highest-insight loss), and the instruction critic gained check R6
+ (comparison-axis coverage, register-independent). The synthesizer (not
+ golden-covered) gained pass-1 item 12 (coverage and mechanism depth are
+ load-bearing content), a reframed selectivity paragraph (select sources,
+ not points), and a "never cut a point to hit the ceiling" clause.
"""
from __future__ import annotations
@@ -363,3 +373,35 @@ def test_cite_check_skill_gets_no_shim(ctx):
_read_skill_source("hyperresearch-14-5-cite-check.md"), ctx
)
assert "shims/" not in rendered
+
+
+# ---------------------------------------------------------------------------
+# Coverage before elegance: the synthesizer and the depth/instruction critics
+# must carry the anti-compression guards so a rerun can't silently regress into
+# trading substance for prose (see the 2026-07-22 deviation-log entry).
+# ---------------------------------------------------------------------------
+
+
+def test_synthesizer_carries_coverage_before_elegance_guards(ctx):
+ rendered = render_prompt(hooks.SYNTHESIZER_AGENT, ctx)
+ # pass-1 item 12: coverage + mechanism depth are load-bearing content
+ assert "load-bearing content" in rendered
+ assert "Elegance is spent on the words BETWEEN points" in rendered
+ # reframed selectivity: pick sources, not points
+ assert "which SOURCES to cite for a" in rendered
+ assert "not which POINTS to make" in rendered
+ # length discipline: never cut a point to hit the ceiling
+ assert "Cut prose, never points" in rendered
+
+
+def test_depth_critic_flags_compressed_mechanisms(ctx):
+ rendered = render_prompt(hooks.DEPTH_CRITIC_AGENT, ctx)
+ assert "compressed to a bare mention" in rendered
+ assert "highest-insight loss" in rendered
+
+
+def test_instruction_critic_has_comparison_axis_check(ctx):
+ rendered = render_prompt(hooks.INSTRUCTION_CRITIC_AGENT, ctx)
+ assert "missing-comparison-dimensions" in rendered
+ # must be register-independent (applies in analyze/survey/advocate alike)
+ assert "register-INDEPENDENT" in rendered