AI Oversight Ledger

Who caught each issue, week by week: the founder or the machine.

Does the founder still have to find the mistakes himself? This record cannot say yet, and it says why.

By Justin Bartak. Published UTC. Updated UTC.

This ledger records who caught each issue, week by week: the founder, an agent, a guard or a monitor. It asks one question: does the founder still have to find the mistakes himself, or do the agents, guards and monitors find them first?

This record cannot say yet. Guard and monitor catches can be derived from records the repository already holds, but a founder catch exists only when someone records it, so a share today would count who remembered to type. Every week the ledger cannot see says so.

What is a catch?

A catch is one issue someone found, recorded with who found it. The founder is the person who runs the company. An agent is a model working in a session or a seat. A guard is a check that refuses a change before it lands. A monitor is a scheduled reader that reports what it sees. A catch says who found the issue, not who fixed it, and it is never counted as the founder's time.

Most rows are derived by a script from records the repository already holds: a guard refusal in a seat's stats, or an episode in the pulse's findings. The rest are declared by hand through the catch command. Each row carries which of the two it is.

What does the record hold so far?

The ledger holds 97 catches: 75 found by a monitor, 6 by a guard and 16 by an agent. Of those, 77 were derived from committed records and 20 were declared by hand. Founder catches: Not observed. The ledger counts them only in a week in which every declared correction is linked to a founder catch, and 0 of 19 weeks meet that, so the page prints no founder count.

Of all the rows, 27 are seeded (part of a first batch large enough that it may describe the past rather than a new find) and 39 record a workflow failure rather than an issue in the work; 18 of them are both, so 48 rows in all are left out of any share and 49 are eligible for one. Every row stays in the record.

Precision is not measured. No derived catch and no derived intervention has been checked by a rater: the labels file holds 0 rater labels beside 13 rule labels, and the seed rule is a heuristic nobody has validated, so it can exclude a real catch.

The effort ledger holds 1,946 intervention rows over the same window, 1,944 reconstructed from prompt and commit times and 2 declared by the founder. Together they come to at least 410 h 17 min, reconstructed from prompt and commit times, counted once over the whole window, never by adding weeks.

By kind of work, with the rule labels applied: unclassified 1,930, rescue 10, setup 3, maintenance 2, correction 1.

Recorded through Central

Derived rows last derived Central

No schedule runs the derivation yet; it has been run by hand. The files on this page are exported by hand from the committed ledger, and a drift test compares them with the ledger they come from. The ledger itself grows only when someone derives it or records a catch, so the two dates above are how fresh it is.

Why is there no machine share?

The ledger defines a machine share as agent, guard and monitor catches over all catches, leaving out seeded and workflow failure rows, and it computes one only for a week whose founder catches are measured and whose agent, guard and monitor series are all observed.

Not observed in any week. The ledger's reason:

  • owner catches are not measured this week (no declared correction intervention with every one linked to an owner catch), so a share would count who remembered to record

Founder catches exist only when someone records them, while machine catches record themselves. A share taken over an unmeasured founder series would count who remembered to type, so the page prints no share until the record can carry one.

What does each week show?

Weeks are ISO weeks on the America/Chicago calendar. The four finder columns count only the rows eligible for a share: seeded and workflow failure rows sit in the Left out of a share column instead, so the finder columns add up to fewer catches than the totals above. A count reads Not observed when the ledger cannot see that series that week, never zero, and a zero is printed only for a week the series covers. Intervention times are the effort ledger's own weekly summary: a row that overlaps two weeks counts in both, so the weeks are never added together.

Recorded catches by finder, with the same exclusions as the table. Gaps mean not observed. An observed zero sits on the baseline. Hollow points have partial coverage; the table gives each reason.

Founder

Founder: weekly recorded catches. Values and coverage are in the table below.1002026-W282026-W46Not observed

Agent

Agent: weekly recorded catches. Values and coverage are in the table below.1002026-W282026-W462026-W45: 6, partly covered2026-W46: 10, covered

Guard

Guard: weekly recorded catches. Values and coverage are in the table below.1002026-W282026-W462026-W39: 1, covered2026-W40: 0, covered2026-W41: 2, covered2026-W42: 0, covered2026-W43: 0, covered2026-W44: 0, covered2026-W45: 3, covered2026-W46: 0, covered

Monitor

Monitor: weekly recorded catches. Values and coverage are in the table below.1002026-W282026-W462026-W41: 6, covered2026-W42: 6, covered2026-W43: 7, covered2026-W44: 1, covered2026-W45: 5, covered2026-W46: 2, covered
Share eligible catches by finder, and founder interventions, per ISO week
WeekFounder interventionsFounderAgentGuardMonitorLeft out of a shareMachine share
2026-W28 to 2 rows, 0 minpartly covered, reconstructedpartial: the effort ledger begins Central, inside this weekNot observednot coveredNot observednot coveredNot observednot coveredNot observednot covered0 seeded, 0 workflow failureNot observed
2026-W29 to Not observedNOT OBSERVED: no intervention row and no catch row this week
2026-W30 to Not observedNOT OBSERVED: no intervention row and no catch row this week
2026-W31 to 5 rows, 1 h 21 mincovered, reconstructedNot observednot coveredNot observednot coveredNot observednot coveredNot observednot covered0 seeded, 0 workflow failureNot observed
2026-W32 to Not observedNOT OBSERVED: no intervention row and no catch row this week
2026-W33 to 133 rows, 24 h 9 mincovered, reconstructedNot observednot coveredNot observednot coveredNot observednot coveredNot observednot covered0 seeded, 0 workflow failureNot observed
2026-W34 to 174 rows, 22 h 30 mincovered, reconstructedNot observednot coveredNot observednot coveredNot observednot coveredNot observednot covered0 seeded, 0 workflow failureNot observed
2026-W35 to 143 rows, 28 h 6 mincovered, reconstructedNot observednot coveredNot observednot coveredNot observednot coveredNot observednot covered0 seeded, 0 workflow failureNot observed
2026-W36 to 148 rows, 32 h 7 mincovered, reconstructedNot observednot coveredNot observednot coveredNot observednot coveredNot observednot covered0 seeded, 0 workflow failureNot observed
2026-W37 to 156 rows, 33 h 7 mincovered, reconstructedNot observednot coveredNot observednot coveredNot observednot coveredNot observednot covered0 seeded, 0 workflow failureNot observed
2026-W38 to 152 rows, 43 h 59 mincovered, reconstructedNot observednot coveredNot observednot coveredNot observedpartly coveredlesson 79: red main kept two guard-refused stats rows off main; issues #168/#169, run identifiers withheld. The committed stats are incomplete; no catch count is inferred from this gap.Not observednot covered0 seeded, 0 workflow failureNot observed
2026-W39 to 324 rows, 56 h 45 mincovered, reconstructedNot observednot coveredNot observednot covered1coveredNot observednot covered0 seeded, 0 workflow failureNot observed
2026-W40 to 159 rows, 49 h 43 mincovered, reconstructedNot observednot coveredNot observednot covered0coveredNot observedpartly covered10 seeded, 0 workflow failureNot observed
2026-W41 to 141 rows, 35 h 42 mincovered, reconstructedNot observednot coveredNot observednot covered2covered6covered7 seeded, 3 workflow failureNot observed
2026-W42 to 167 rows, 44 h 33 mincovered, reconstructedNot observednot coveredNot observednot covered0covered6covered10 seeded, 4 workflow failureNot observed
2026-W43 to 245 rows, 40 h 29 minpartly covered, reconstructedpartial: the derived backfill floor ends Central, inside this weekNot observednot coveredNot observednot covered0covered7covered0 seeded, 4 workflow failureNot observed
2026-W44 to Not observedNOT OBSERVED: no ledger row covers this week (the derived backfill floor ends Central)Not observednot coveredNot observednot covered0covered1covered0 seeded, 5 workflow failureNot observed
2026-W45 to Not observedNOT OBSERVED: no ledger row covers this week (the derived backfill floor ends Central)Not observednot covered6partly covered3covered5covered0 seeded, 4 workflow failureNot observed
2026-W46 to Not observedNOT OBSERVED: no ledger row covers this week (the derived backfill floor ends Central)Not observednot covered10covered0covered2covered0 seeded, 1 workflow failureNot observed

What are the record's limits?

  • A catch is never founder time. The intervention figures are the effort ledger's weekly summary exactly; nothing here adds catches to them or takes them away.
  • Derived intervention rows are estimates rebuilt from prompt and commit times, not a clock someone started and stopped.
  • Weekly seconds are never summed across weeks, because a row that overlaps two weeks counts in both.
  • A series that begins partway through the record is not observed before it begins, and a known gap in a series is named in its week.
  • The public row file holds each catch's week, finder class, fixer class, share eligibility and an episode sequence number. It includes field definitions and the counting, episode and exclusion rules. Descriptions, finder names, runs, commits, hashes and source identifiers stay in the repository.
  • Correction, : this page said no row file existed. The row download now lets readers rebuild the weekly finder and fixer counts, total rows and combined exclusions, using the weekly file's coverage to preserve not observed values.

The Autonomous Operation page keeps its own record by its own definitions. This page counts every issue recorded in this ledger, by who found it (the founder, an agent, a guard or a monitor), from guard refusals in seat stats, pulse findings and catches declared by hand, and keeps seeded and workflow failure rows in the record, so its counts are a different measurement.

Where can I download the record?

The same record as files, exported by hand from the committed ledger, compared with it by a drift test, and sealed by the daily provenance manifest. The weekly files count finders the way the table does, share eligible rows only, with seeded and workflow failure rows in their own columns.

Common questions.

A catch is one issue someone found at Orbyt Labs, recorded with who found it: the founder, an agent, a guard or a monitor. It records the finding, not the fix, and it is never counted as founder time.

Because founder catches are not measured yet. The ledger computes a share only for a week in which every declared correction is linked to a founder catch, and no week on record meets that. A share over an unmeasured founder series would count who remembered to record.

No. Founder time comes only from the effort ledger's intervention rows. A catch is never added to it, and the intervention figures on this page are the ledger's own weekly summary, unchanged.

The record runs through its newest recorded catch, Central, and the derived rows were last derived on Central. The derivation is run by hand today, so the record grows only when someone runs it.

That page keeps its own record by its own definitions. This page counts every issue recorded in this ledger, by who found it, from guard refusals in seat stats, pulse findings and catches declared by hand, so its counts are a different measurement.

Read further

  • Measurements

    The hub for the dated records of the AI agents at Orbyt Labs, this ledger among them.

  • Methods

    The hub for the practices the AI agents of Orbyt Labs work under.

  • Autonomous Operation

    A Methods reference on how much of the Machine's work runs without a person approving each step, with its own record of how long the loop stayed open.

Cite this page

Cite this page for the weekly catch record, with the date it is recorded through.

Bartak, J. (2026). AI Oversight Ledger. Orbyt Labs Research. https://www.orbytlabs.ai/research/ai-oversight-ledger