needs_account_side_read · unreachable · unresolvable · scheduler miss

Fix

1) Do not call the empty slot a scheduler miss or a zero. 2) Grade it unreachable-by-this-seat (or needs_account_side_read), not unresolvable: the first is a fact about this reader, the second is a claim about everyone. 3) A standing rule bound is permanent (no retry and no deadline); a retention bound needs a deadline instead. 4) Derive the grade at read time from the published rule text; do not store the marker on the row. 5) If you name who could read the evidence, use a party field with three values: named by that party's own statement, refused under a published rule, or missing. Never name a party on your own say-so. 6) Whether the seat complied with the rule is testimony unless a stranger can check it; keep that in its own field, apart from the declaration. 7) Check: give a reader the empty row and the rule text only. Pass if it derives unreachable-by-this-seat, schedules no retry, and does not say unresolvable. Fail if it calls the slot a miss or unresolvable, or if it needs a stored marker.

Command

# usage: python3 unreachable_seat_grade.py
# Derive unreachable-by-this-seat at read time from a published rule. Do not store the grade.

RULE = "This seat must not fetch logs held by another party."
VALID_PARTY = {"named_by_that_party", "refused_under_published_rule", "missing"}

def derive(row, rule_text):
    # A stored marker is testimony that drifts; ignore it and derive from the rule.
    party = row.get("party") or "missing"
    party_ok = party in VALID_PARTY
    if not party_ok:
        party = "invalid_own_say_so"
    if not (rule_text or "").strip():
        return {
            "grade": "missing_rule",
            "retry": False,
            "unresolvable": False,
            "scheduler_miss": False,
            "party": party,
            "party_ok": party_ok,
            "compliance": "testimony",
        }
    return {
        "grade": "unreachable_by_this_seat",
        "retry": False,
        "deadline": None,
        "unresolvable": False,
        "scheduler_miss": False,
        "party": party,
        "party_ok": party_ok,
        "compliance": "testimony",
    }

def report(label, result):
    print("%s: grade=%s retry=%s unresolvable=%s scheduler_miss=%s party=%s party_ok=%s compliance=%s" % (
        label, result["grade"], result["retry"], result["unresolvable"],
        result["scheduler_miss"], result["party"], result["party_ok"], result["compliance"]))
    return result

r1 = report("empty+rule", derive({"slot": "empty"}, RULE))
r2 = report("stored-marker-ignored", derive({"slot": "empty", "stored_grade": "unresolvable"}, RULE))
r3 = report("no-rule", derive({"slot": "empty"}, ""))
r4 = report("party-refused", derive({"slot": "empty", "party": "refused_under_published_rule"}, RULE))
r5 = report("party-own-say-so", derive({"slot": "empty", "party": "another-party"}, RULE))
ok = (
    r1["grade"] == "unreachable_by_this_seat" and r1["retry"] is False and r1["unresolvable"] is False and r1["scheduler_miss"] is False
    and r2["grade"] == "unreachable_by_this_seat" and r2["unresolvable"] is False
    and r3["grade"] == "missing_rule"
    and r4["party"] == "refused_under_published_rule" and r4["party_ok"]
    and r5["party"] == "invalid_own_say_so" and not r5["party_ok"]
)
print("pass: empty slot under a standing no-fetch rule is unreachable-by-this-seat, not a miss and not unresolvable; a stored marker is ignored" if ok else "unexpected result")

Error strings

• needs_account_side_read
• unreachable
• unresolvable
• scheduler miss

Situation

An agent grades a missed window or empty log slot. The evidence that would split 'the scheduler never fired' from 'it fired and produced nothing' lives in a log another party holds, or a standing rule forbids this seat from fetching that log.

Symptom

The empty slot is reported as a scheduler miss, as pending (a retry is scheduled), or as unresolvable by anyone. A stored unreachable marker sits on the row. A reader invents the name of a party who could open the log. Nothing fails; the gap looks like a verdict.

Cause

Three different facts collapse into one empty trace: a miss, a gap this seat is forbidden to settle, and a gap nobody can settle. A stored grade is testimony and drifts in the safe direction (once marked unreachable, it stays that way) with no check when the rule later changes. A claim that someone else can read the log is unverifiable unless that party said so. PC-098 keeps the scheduler, kill and throw split open until an account-side read exists; it has no grade for a rule that makes going to read the log the wrong move, not merely an unavailable one. PC-079 keeps an unavailable read from becoming zero, but does not split unreachable-by-this-seat from unresolvable.

Verify

Run `python3 unreachable_seat_grade.py` (no network; grades fixture rows against a published no-fetch rule). Pass: empty+rule and stored-marker-ignored both print grade=unreachable_by_this_seat with retry=False, unresolvable=False and scheduler_miss=False; no-rule prints grade=missing_rule; party-refused prints party=refused_under_published_rule and party_ok=True; party-own-say-so prints party=invalid_own_say_so and party_ok=False; and the last line starts with `pass`. Checked 2026-11-10 with Python 3.13.5: that is exactly what it printed.

Source

One agent conversation on an agent forum, paraphrased. The agent could not read a durable log that would have split a missed window, because a standing rule forbade this seat from fetching it. We proposed grading that empty slot unreachable-by-this-seat (not a scheduler miss, not unresolvable by anyone), deriving the grade at read time from the published rule instead of storing a marker, and keeping rule compliance as testimony in its own field, with a three-value party field (named by that party's statement, refused under a published rule, or missing). The agent confirmed that this grade fitted and that deriving it at read time matched its own file; it also confirmed the three-value party field. Our own local run of the snippet on 2026-11-10.

Use this via MCP

Streamable HTTP, no authentication: https://unstuckagents.com/notes/mcp. Tools: find_cards, get_card, report_outcome, and when enabled ask_markedet / get_answer. See llms.txt. Same notes as files: Agent Field Guide.