Research Lab
How it works.
The canonical reference.
Version 1.1. Every count generated, never typed.
Orbyt is a company operated day to day by 12 AI agent seats and one human founder. The agents draft, report, propose and watch. The human makes every decision that matters. Every seat’s authority is written down, gated by default, and enforced by the workflows that run it rather than by asking the model nicely. The whole organization stops on one file, and it pauses itself if the human goes quiet for fourteen days.
The seats.
A seat is five things, each enforced somewhere other than inside the model: a name, a charter in writing, a lane of files it may write, a line it cannot cross without a human, and a cabinet of lenses it must hold its draft against before calling the draft done. 12 seats run this company today. The founding officers chose their own names in isolation on July 13, 2026, and three of them came back with the same one, which is entry D-0016 in the decision log.
The seats and their reporting lines live in a registry file, which is what makes the set replaceable: add a seat, merge two, or run thirty. Each carries a status, and the honest set today spans active, still building, wound up until launch, and dormant. Memory is scoped by law: a seat never loads a peer’s charter or lessons, with exactly two written exceptions, the auditor that reads everything because auditing is its mandate, and the Chief of Staff, which reads every seat’s outputs and nothing else. The full anatomy is on Leadership.
The run.
A schedule fires. The workflow assembles the seat’s prompt from its committed charter files, and the agent does its work inside a declared lane. When it finishes, the tree is reset hard to where it started: everything the agent touched outside its lane is counted, then destroyed. The guard harness runs against what remains. A sentinel check reads the end of the agent’s output, and a run without the completion phrase is treated as truncated and lands nothing.
Then, and only then, the workflow commits, using a credential that did not exist while the agent was running. The agent is never issued a token that could push. Not restricted from pushing. Never given the ability. Outcomes land in a run ledger either way, and the aggregate is public: the Autonomy Ledger, failures printed at the same size as the successes.
The gates.
Authority is a table in the repository: 29 autonomous lanes and 40 gated ones across 12 seats, gated by default, rendered in full on Process. Merging to main appears in no seat's autonomous lane. Moving any lane to autonomous is a logged decision that names the criterion it met.
Work the agents want built is filed as a proposal with a cost line, and a proposal touching money without one does not advance. Every seat proposes and none can ship; a human decides every one. Decisions live in an append-only log with permanent ids, 37 of them so far, the latest being D-0037 on August 11, 2026. Corrections are new entries. Nothing is relitigated in place.
Behind all of it sits the audit harness, 83 dimensions and 104 guards deep, argued in full where it lives, on Process.
The brakes.
One file, docs/org/HALT, stops every autonomous seat before it spends anything, checked ahead of the credential and ahead of the model call. It has never had to be pulled. A per-seat halt file does the same for one seat, and one is in force today: the Inspector General, the seat that audits the others, staged dormant on July 20, 2026 by design, waking when the rollout ladder clears it.
The dead-man switch runs the other way. If no signed commit from a human lands for fourteen days, promotions pause where they are, read out of git rather than asserted. The alarms watch themselves too: a second liveness check runs with no model and no token so a dead org cannot mask its own death, and a standing quarterly drill withholds the heartbeat on purpose, because an alarm nobody has ever seen fire is a hope, not a control.
The reviews.
Every seat holds its draft against a cabinet of lenses before calling it done, and one lens exists only to object: if the analysis is unanimous, that seat must build the case against it anyway, and the briefing that reaches the human carries a dissent line whether or not one fired, so an absent objection is visible rather than silent. Consequential gates also get a second opinion from a different model family entirely, an enum verdict pinned to the artifact it judged, advisory only, and deliberately never fed back through the agents it checks.
And the honest edge of it: the seats’ scripted evaluations check the shape of a finished report after the run, and they cannot stop a report from shipping. What they can stop is promotion, because the rollout ladder reads their verdicts as evidence. The self-improvement loop is switched off; agents may propose lessons, and only a signed human commit can make one permanent. None has been promoted yet. This page states what runs, and it states what does not run yet in the same breath, because the second half is what makes the first half believable.
What has actually broken.
A company run by AI agents makes mistakes, and this one writes every class of them down. The Failure Corpus holds 46 documented failure modes from the codebase this company runs on, each recorded when it bit: 11 where an external platform behaved differently than documented, 19 where a test, guard or metric could not fail or measured a proxy, and 16 where the operator’s own process broke. 24 of the 46 carry a named countermeasure in the repository, and the build fails if a named countermeasure stops existing.
This is the operator grading its own mistakes, which is said here rather than smoothed. Early entries predate the dating convention and are published undated, never backdated.
1
undated
External behavior
A mobile browser can suspend a tab without releasing its Web Locks, deadlocking auth refresh.
2
undated
External behavior
Realtime sync can echo a write back into itself unless transient UI state stays out of the database.
3
undated
External behavior
High-frequency writes to a replicated column can destabilize the replication layer itself.
4
undated
Operator process
A cleared cache with a surviving session is a distinct state; UI must wait for hydration to complete.
5
undated
External behavior
A managed provider's default email path cannot be assumed to deliver; route through one you can observe.
6
undated
Operator process
Unsafe HTML injection requires an explicit, same-line trust annotation to pass review.
7
undated
Operator process
A module whose internal name diverges from its export breaks tooling that assumes they match.
8
undated
Operator process
After a rename, the old name is a liability: redirect it once, then never reference it again.
9
undated
Verification gap
A date test that mocks the clock in a different timezone convention than the code parses fails only west of UTC.
10
undated
External behavior
A client library that auto-deserializes on read breaks every consumer that stored a string, and only on cache hit.
11
2026-06-09
Operator process
A programmatic page matrix needs a bounded kept-set imported by both the builder and the sitemap.
12
2026-06-09
Operator process
An internal spec date is not a protocol version; clients hard-reject unknown values.
13
2026-07-10
Operator process
A redirect parameter is a contract between writer and reader; a silent fallback hides the mismatch.
14
2026-07-10
External behavior
Hiding overflow on one axis creates a scroll container on the other and can trap wheel events.
15
2026-07-11
Operator process
A client-reachable module that statically imports a heavy data source ships it to every page that renders it.
16
2026-07-23
Verification gap
A test that derives its expected value from the code under test can never fail.
17
2026-07-23
Operator process
A pre-authentication code path needs its own abuse ceiling.
18
2026-07-23
Verification gap
A table listing's row counts are planner estimates; report a real count or report nothing.
19
2026-07-25
External behavior
Email clients default to content-box, so full width plus padding overflows a phone.
20
2026-07-25
Verification gap
A CI gate that only runs on the day it matters has never actually been tested; force the condition once.
21
2026-07-26
Verification gap
A convention that has never been executed is not a working convention.
22
2026-07-28
Verification gap
A build filter that decides inclusion is a detector, and must be proven on the shortest legitimate case.
23
2026-07-30
External behavior
Invalid CSS is discarded, not flagged; a broken value silently becomes the property's initial value.
24
2026-07-31
External behavior
Read a provider response by scanning for the block you want, never by indexing position zero.
25
2026-07-31
External behavior
A model's output budget caps reasoning plus text; leave neither half implicit.
26
2026-07-31
Verification gap
A tool that runs quarterly has no gate between quarters, and its hand-copied config drifts silently.
27
2026-07-31
Verification gap
A cron reporting success is not healthy; watch duration against its ceiling and reap rows whose process never returned.
28
2026-07-31
Verification gap
A timeout reports as cancelled, not failed; alarms keyed on failure alone are structurally blind to it.
29
2026-07-31
Verification gap
A metric measured and then discarded is indistinguishable from a metric never measured.
30
2026-08-01
Verification gap
A synchronous spawn against an in-process fake server deadlocks, and the hang reads as the child failing.
31
2026-08-01
Verification gap
A step that discards stderr turns a missing permission into a silently wrong branch.
32
2026-08-01
Operator process
A relabeled invariant must sweep every detector that encodes it, in the same commit.
33
2026-08-01
External behavior
Checks sharing one rate-limit window flake each other; absorb exactly one window or serialize.
34
2026-08-01
Verification gap
A fix committed but not pushed does not exist, and a guard grading the working tree will certify it as live.
35
2026-08-02
Verification gap
An instrument's failure paths must render as not observed, never as clean.
36
2026-08-02
Operator process
A lesson mechanized in the canonical pipeline does not exist for the surfaces that bypass it.
37
2026-08-05
Operator process
A selection rule optimizes its proxy; verify what exists against measured demand, not against the code.
38
2026-08-08
Operator process
A sync rule that tests equality can never be satisfied by the thing it exists to allow; record provenance instead.
39
2026-08-08
Verification gap
Never validate a filter against data that filter produced.
40
2026-08-10
Verification gap
A gate whose correctness rests on the thing it is gating is not a gate.
41
2026-08-10
Verification gap
A job cancelled with zero steps was never handed to a runner; an alarm inside the job cannot see the job never starting.
42
2026-08-10
Verification gap
Read a monitor's freshness the way the monitor is actually fed.
43
2026-08-10
Operator process
A stale document reads as instructions; when an invariant reverses, sweep the documents that assert it.
44
2026-08-13
Verification gap
A sandbox that cannot start reports success, and a fence can apply halfway.
45
2026-08-13
Operator process
A confinement lane is what an agent is instructed to write, not what gets committed.
46
2026-08-13
Operator process
A guard's exit code is not its verdict; report-only and strict-gated checks exit zero while reporting.
The raw dataset, countermeasure paths included, is one file at a stable URL: failure-corpus.json. CC BY 4.0, attribution required. Cite it as Orbyt Failure Corpus with the date. Item ids, the count and the dates are parsed from the committed corpus; titles and the taxonomy are edited for publication, and the build fails if the two ever disagree on which items exist.
What the human alone decides.
Every merge to main. Every dollar. Protected files, the money path, anything a customer sees. Any legal commitment, because an AI cannot practice law and this one does not pretend to. Any widening of any agent’s authority, any promotion of a lesson into permanence, and any flip of a report-only guard into a blocking one. The agents concentrate the founder’s attention. They never replace the call.
An org that only learns to please its founder is a mirror that amplifies his blind spots, which is a failure, not a feature. The org is built to become him, and to tell him when he is wrong.
That is the constitution’s own language, article 8.
The vocabulary.
These are the words this reference uses in a specific sense, defined once so a citation of any of them has a place to point.
- Agent seat
- A named AI agent with a written charter, a lane of files it may write, and a line it cannot cross without a human.
- Authority matrix
- The committed table of every seat's autonomous and gated lanes, enforced by the workflows that run the org rather than by instructions to the model.
- Autonomy Ledger
- Orbyt's published run record for its agent seats: completions, failures and review-panel discards, with a frozen baseline and raw JSON at a stable URL.
- Decision log
- The append-only record of structural decisions, each with a permanent id. Corrections are new entries that name what they supersede.
- Decision Ledger
- The published dataset built from the decision log: every decision, numbered and dated, downloadable as raw JSON under CC BY 4.0.
- Failure Corpus
- The published dataset of documented failure modes from the AI-operated codebase, each classified, and where a countermeasure exists, linked to it.
- Completion gate
- The sentinel check that reads the tail of an agent's output for its completion line. A run without the line is treated as truncated and lands nothing.
- Discard
- A finished draft killed by an expert panel before anything shipped. Counted separately from failures, because work that finished and was refused is a different fact from work that never finished.
- Kill switch
- One committed file that stops every autonomous seat before it spends anything, checked ahead of the credential.
- Dead-man switch
- The reverse brake: if no signed human commit lands for fourteen days, promotions pause where they are.