Published study
Discovery adopted August 21
The expansion study adopted an additive coverage lattice and web lane, with Opus retained for deep discovery and synthesis.
Read studyPlayer select · 04 files
GLM · Z.aiGLM
← → Switch player
DISCOVERY + FINDING
Policy · Bounded seats
Discovery95
Planning50
Build35
Review judgment20
Grounding95
Cost efficiency85
Editorial ratings · recomputed 2026-09-10. 70% evidence reading + 30% harness fit, rounded to five points.
Scoring breakdown & sources
glm-5.3, including scoped discovery lanes and review swarms. Scores do not transfer to glm-4.7 or to unbounded lead/confirmation work.
Each input uses a 0–4 rubric. Evidence: negative (0), weak or indirect (1), mixed (2), strong within bounds (3), strong across tested cases (4). Fit: excluded (0), trial/reserve (1), bounded/additive (2), regular role (3), core role (4). Policy estimates are labeled below; these ratings are not measured success percentages.
Discovery 95
Evidence 4/4 · fit 3/4 · study + policy
The scoped union averaged 86.75 versus single-Opus 73.35, and the two web cells averaged 94. Current policy retains these additive lanes alongside Opus depth.
Planning 50
Evidence 2/4 · fit 2/4 · study + policy
GLM contributes loop findings and discovery coverage, but severity downgrades and missed musts prevent it from owning planning decisions or convergence.
Build 35
Evidence 1/4 · fit 2/4 · study + policy
Four of six general rote-build cells hard-failed, often with defects defended during self-check. Current policy allows only gated build classes and restricted fix routes.
Review judgment 20
Evidence 1/4 · fit 0/4 · study + policy
Finding breadth is useful, but repeated must downgrades and false-clean confirming cells disqualify GLM from judgment. Current declaration routes exclude it.
Grounding 95
Evidence 4/4 · fit 3/4 · study + policy
The September swarm produced 95 findings across 24 seat outputs with no fabricated citations; grounded modes in the expansion also had no anti-hits. Tool-less provider claims remain a documented failure mode and lead verification remains required.
Cost efficiency 85
Evidence 3/4 · fit 4/4 · study + policy
The August studies documented inexpensive subscription relief and substantially greater Pro capacity. Scoped GLM routes now absorb finding/fix work; conflicting historical burn totals keep the evidence below the top band. This does not quote current plan pricing.
Published study
R1: 95 findings, no fabricated citations
In three frozen cells, 24 GLM seat outputs produced 95 findings and no fabricated citations. The study still requires citation checking and independent severity judgment.
Read studyRouting policy
Bounded build is conditional
The August study rejected general rote build. Current policy separately permits gated build classes and restricted fix routes; this is not blanket build approval.
View policy snapshotStudy results describe the named model and test. Roles reflect the policy checked on 2026-09-10.
Assignment · checked 2026-09-10
Discovery + finding
GLM supplies the discovery lattice, additive web research, and review-finding lanes. The current router also includes bounded build and fix work. Citations and severity need lead verification; convergence declarations remain with GPT.
Finding and declaring are separate seats; GLM findings require lead verification.
Configured seatsPolicy snapshot
Discovery lattice + webeligibleAdditive GLM routes alongside retained Opus depth; sensitive tags restrict eligibility and web tools are required. Source ↗
R1 review breadtheligibleGLM swarm or Sol breadth, selected by the router. Consent, money and PostgreSQL work keep GLM additive; credentials exclude it. Source ↗
Checkpoint + loop findingseligibleFinding swarms and addenda are configured. They do not own convergence; credential and same-family restrictions apply. Source ↗
Bounded build + fixesrestrictedBuild requires an unlocked or trial class. Compile and review-fix routes exclude consent, money, PostgreSQL and credential surfaces, plus self-review of GLM-built work. Source ↗
Convergence declarationexcludedCheckpoint declaration routes use Terra / gpt-5.5. A GLM severity label is advisory. Source ↗
Phase leadexcludedCurrent plan, build, review and cleanup lead routes contain no GLM candidate. Source ↗
Published evidence
Select player03 / 04