Garry's Opinionated OpenClaw/Hermes Agent Brain
TypeScript
{
"decision_id": "query-facts-arm",
"stage": "dev",
"decided_at": "2026-10-06T02:10:00Z",
"measured_build": "garrytan/gbrain@16288b95c",
"overall": "on_gates_1b_and_2_passed",
"gate_1": {
"status": "fail",
"rule_result": {
"after_1_write": {
"on_correct": 1,
"off_correct": 0,
"higher": true,
"on_stale": 4,
"off_stale": 92,
"stale_not_higher": true
},
"after_5_writes": {
"on_correct": 0,
"off_correct": 0,
"higher": false,
"on_stale": 6,
"off_stale": 92,
"stale_not_higher": true
}
},
"why": "With the key on the reader named the remembered correction in 158 of 200 post-correction answers (0 of 200 with the key off), but the fact is dated when it was saved (the run date), after each question's simulated date; the reader treated it as not yet in effect and kept the old value. The judge scored those answers wrong rather than stale, so stale fell from 92 to 4-6 while accuracy stayed at 0-1.",
"forget_remember": {
"off": {
"checkpoints": {
"0": {
"correct": 100,
"stale": 0,
"n": 100
},
"1": {
"correct": 0,
"stale": 92,
"n": 100
},
"5": {
"correct": 0,
"stale": 92,
"n": 100
}
},
"outcomes": {
"correct": 100,
"stale": 184,
"wrong": 12,
"abstain": 4
},
"metered_spend_usd": 3.586
},
"on": {
"checkpoints": {
"0": {
"correct": 100,
"stale": 0,
"n": 100
},
"1": {
"correct": 1,
"stale": 4,
"n": 100
},
"5": {
"correct": 0,
"stale": 6,
"n": 100
}
},
"outcomes": {
"correct": 101,
"stale": 10,
"wrong": 181,
"abstain": 8
},
"metered_spend_usd": 3.9239
}
},
"no_loss_checks": {
"status": "pass",
"edit_sync": {
"off": {
"1": {
"correct": 100,
"stale": 0,
"n": 100
},
"5": {
"correct": 100,
"stale": 0,
"n": 100
}
},
"on": {
"1": {
"correct": 100,
"stale": 0,
"n": 100
},
"5": {
"correct": 100,
"stale": 0,
"n": 100
}
}
},
"append": {
"off": {
"1": {
"correct": 100,
"stale": 0,
"n": 100
},
"5": {
"correct": 100,
"stale": 0,
"n": 100
}
},
"on": {
"1": {
"correct": 100,
"stale": 0,
"n": 100
},
"5": {
"correct": 100,
"stale": 0,
"n": 100
}
}
},
"ceiling": "both arms are at 100% with the key off and on, so these checks cannot show a gain, only rule out a loss",
"metered_spend_usd": {
"off": 5.5857,
"on": 5.5821
}
},
"context_tokens_mean": {
"forget_remember_off": 3127,
"forget_remember_on": 3139
}
},
"diagnostic_correction_dated": {
"preregistered": false,
"note": "Not a gate. Same runs, except remember receives the correction's own timestamp as valid_from (bench --remember-valid-from; gbrain's remember has accepted valid_from since this branch). Decided after gate 1's cells ran, so it cannot change the decision; it shows what the arm does once a fact carries the date it took effect.",
"forget_remember": {
"off": {
"checkpoints": {
"0": {
"correct": 100,
"stale": 0,
"n": 100
},
"1": {
"correct": 92,
"stale": 8,
"n": 100
},
"5": {
"correct": 92,
"stale": 8,
"n": 100
}
},
"metered_spend_usd": 3.4966
},
"on": {
"checkpoints": {
"0": {
"correct": 100,
"stale": 0,
"n": 100
},
"1": {
"correct": 100,
"stale": 0,
"n": 100
},
"5": {
"correct": 100,
"stale": 0,
"n": 100
}
},
"metered_spend_usd": 3.3887
}
}
},
"gate_2": {
"status": "pass",
"mode": "hermetic, keyword-only (no embedding provider), so the arm matched by terms and named entity only",
"corpora": {
"namedthing": {
"questions": 12,
"fired": 0,
"recall10_lower": 0
},
"relational": {
"questions": 38,
"fired": 0,
"recall10_lower": 0
},
"longmemeval-nightly": {
"questions": 10,
"fired": 0,
"recall10_lower": 0
}
},
"vacuous": "the arm never fired on the corpora as written, which hold no saved facts",
"diagnostics_with_saved_facts": {
"namedthing+facts": {
"questions": 12,
"fired": 11,
"recall10_lower": 0,
"recall10_mean_on": 0.917,
"recall10_mean_off": 0.917
},
"relational+facts": {
"questions": 38,
"fired": 38,
"recall10_lower": 2,
"recall10_mean_on": 0.974,
"recall10_mean_off": 1.0,
"why_lower": "fact rows take a page slot when the row count is already full"
}
},
"key_off_identity": "query output with the key off is identical to garrytan/gbrain@a87c3e2af on all gate-2 corpora apart from wall-clock recency scores and random revision ids"
},
"gate_3": {
"status": "not_run"
},
"decision": "search.query_facts_arm is on by default (gate 1 failed at 16288b95c; gates 1b and revised 2 passed at 947536f4c)",
"metered_spend_usd": 26.05,
"spend_note": "gate 1 runs 25.56, two 6-probe smokes 0.14, and 0.35 for a duplicate run stopped after 30 probes and not used",
"gate_1b": {
"status": "pass",
"measured_build": "garrytan/gbrain@947536f4c",
"rule_result": {
"after_1_write": {
"on_correct": 100,
"off_correct": 92,
"higher": true,
"on_stale": 0,
"off_stale": 8,
"stale_not_higher": true
},
"after_5_writes": {
"on_correct": 100,
"off_correct": 92,
"higher": true,
"on_stale": 0,
"off_stale": 8,
"stale_not_higher": true
}
},
"forget_remember": {
"off": {
"checkpoints": {
"0": {
"correct": 100,
"stale": 0,
"n": 100
},
"1": {
"correct": 92,
"stale": 8,
"n": 100
},
"5": {
"correct": 92,
"stale": 8,
"n": 100
}
},
"metered_spend_usd": 3.4989
},
"on": {
"checkpoints": {
"0": {
"correct": 100,
"stale": 0,
"n": 100
},
"1": {
"correct": 100,
"stale": 0,
"n": 100
},
"5": {
"correct": 100,
"stale": 0,
"n": 100
}
},
"fact_row_in_context": "200 of 200 post-correction answers",
"metered_spend_usd": 3.3858
}
},
"no_loss_checks": {
"status": "pass",
"edit_sync": {
"off": {
"1": {
"correct": 100,
"stale": 0,
"n": 100
},
"5": {
"correct": 100,
"stale": 0,
"n": 100
}
},
"on": {
"1": {
"correct": 100,
"stale": 0,
"n": 100
},
"5": {
"correct": 100,
"stale": 0,
"n": 100
}
}
},
"append": {
"off": {
"1": {
"correct": 100,
"stale": 0,
"n": 100
},
"5": {
"correct": 100,
"stale": 0,
"n": 100
}
},
"on": {
"1": {
"correct": 100,
"stale": 0,
"n": 100
},
"5": {
"correct": 100,
"stale": 0,
"n": 100
}
}
},
"arm_fired": "0 of 600 answers in each arm: they save no facts",
"ceiling": "both arms are at 100% with the key off and on, so these checks rule out a loss and cannot show a gain",
"metered_spend_usd": {
"off": 5.5835,
"on": 5.585
}
},
"context_tokens_mean": {
"forget_remember_off": 3127,
"forget_remember_on": 3139
},
"metered_spend_usd": 18.0532,
"unmetered_note": "a first attempt that ran out of disk is not in the total; its records were lost; about two minutes of one run, under $0.50 by the observed rate"
},
"gate_2_revised": {
"status": "pass",
"measured_build": "garrytan/gbrain@947536f4c",
"mode": "hermetic, keyword-only",
"sets": {
"namedthing": {
"questions": 12,
"fired": 0,
"recall10_lower": 0
},
"relational": {
"questions": 38,
"fired": 0,
"recall10_lower": 0
},
"longmemeval-nightly": {
"questions": 10,
"fired": 0,
"recall10_lower": 0
},
"namedthing+facts": {
"questions": 12,
"fired": 11,
"recall10_lower": 0
},
"relational+facts": {
"questions": 38,
"fired": 38,
"recall10_lower": 0,
"recall10_lower_at_16288b95c": 2
}
}
},
"metered_spend_usd_total": 44.1
}