Case 03: adversarial loop writing filter

Target note: kb/notes/adversarial-loop-can-reconstruct-the-writing-is-thinking-filter.md

Frozen material

Snapshot hash:

6bddeedeaeecbe3f8b900c7b5c172e9dccdefbd8b2e368295eb5a322cfa34947  baseline-working-tree.md

Experiment log

2026-06-16: prune weak expansions

Instruction under test: case 01 prune instruction.

Report: prune-weak-expansions-report.

Result: strong fit.

What it found:

  • It preserved the note's strongest claim: the writing-is-thinking filter can be relocated into an adversarial human-agent loop.
  • It identified the central precision fix: the architecture cannot "force" human judgment; it can make the condition explicit, routed, and auditable.
  • It treated corpus connection work as a valuable but under-integrated branch.
  • It flagged the final "better thinking than the solo pen" payoff as an empirical bet rather than an established result.

Takeaway: the prune instruction works well on a cohesive note that still has a few overconfident branches. It did not attack the core claim; it narrowed the promises around it.

2026-06-16: split and rehome critique

Instruction under test: case 01 split/rehome instruction.

Report: split-rehome-critique-report.

Result: useful branch diagnosis.

What it found:

  • It kept the reconstructed-filter mechanism as the main note.
  • It identified corpus connection work as the only real rehoming candidate.
  • It recommended deleting or narrowing the "no solo equivalent" and "better than the solo pen" claims.

Takeaway: the split/rehome instruction remains useful when there is only one branch. It does not force a large decomposition, but it names the branch cleanly enough that the applied edit can demote it to a secondary payoff.

2026-06-16: marginal-value redundancy gate

Gate under test: case 02 marginal-value redundancy gate.

Report: marginal-value-redundancy-report.

Result: success.

The gate produced two WARN findings:

  • the standalone corpus-connection paragraph consumes a full chunk on a second thesis;
  • the closing paragraph repeats the division-of-labor condition and adds an unearned comparative payoff.

Takeaway: the new gate catches a different weakness from ordinary semantic review. The issue is not that either passage is false; it is that their current size and placement cost more context than their marginal contribution earns.

2026-06-16: no-DB compression bundle

Instruction under test: run-compression-bundle-on-note.

Bundle under test: compression.

Sub-agent report: compression-bundle-review.

Result: stronger signal than the single-gate pass.

Gate results:

Gate Result Main signal
compression/core-claim-obscured INFO The title and bold sentence expose the core claim, but the corpus-connection branch changes the remembered center of gravity.
compression/branch-bloat WARN The corpus-connection paragraph is a separate defense of agent-operated KB work and should be removed or rehomed.
compression/detail-overhang WARN The opening concession, architecture inventory, and Relevant Notes grounds carry more detail than needed.
compression/marginal-value-redundancy WARN The note restates the human/adversarial condition in several places; the condition can be carried by fewer chunks.

Takeaway: the full compression bundle is more useful than the marginal-value gate alone. It preserves the earlier branch-bloat finding while also catching detail proportion and footer-grounding verbosity.

2026-06-16: applied workshop draft

Applied draft: revised-from-prune-and-gate.

Result: promising local improvement.

Changes:

  • compressed the frontmatter description;
  • rewrote "architecture exists to force the condition" as "make the condition explicit, routed, and auditable";
  • demoted corpus connection work from a separate proof branch to a secondary payoff;
  • narrowed the final payoff from "produces better thinking than the solo pen" to a conditional bet about relocating the stall into review.

2026-06-16: applied compression-bundle draft

Applied draft: revised-from-compression-bundle.

Result: stronger compression than the prune/gate draft.

Changes relative to the baseline:

  • compressed the opening concession to the minimum needed to establish Borretti's critique;
  • removed the standalone corpus-connection paragraph and folded it into the competence-floor paragraph as a secondary payoff;
  • shortened the architecture inventory while preserving the condition that judgment must remain human and adversarial;
  • merged the separate failure section into the closing paragraph;
  • shortened Relevant Notes grounds so they route rather than re-argue.

Interim judgment

This case supports the current workshop hypothesis. The improved critique instructions and the marginal-value redundancy gate can improve a note that is already mostly cohesive by finding where interesting support has grown into a second thesis. For this target, the best review signal is not "split the note" but "keep the reconstructed-filter claim, demote corpus connection work, and narrow architecture guarantees."

Complete file listing (generated at build time)