An AI-enabled workflow, with the oversight built in. 144 errors on the record: the human caught 36, and his controls caught the other 108.
The AI does the volume. The human sets the strategy, writes the controls, and catches what no control can. Every error in this build is logged with what broke, who found it and the rule it earned. The human's catches come first, because no control could have made them. The machine's catches come second, grouped by the control that made each one — every one of them a rule the human wrote into the system.
144errors logged, each with the rule it earned
36caught by the human
108caught by controls the human put in place
36 human · 25%108 his controls · 75%
Part one · 36 human catchesGrouped by the move he made. These are the errors every check passed.
Part two · 108 control catchesGrouped by the control that caught them, and where each control came from. Tap a control to see its entries.
One entry touched what employers received: a resume line sent with the early control-wave applications read stronger than its evidence, and the next build corrected it. Nothing else in this ledger reached an employer, a recruiter or any public surface. One error reached the second model in the stack — two unverified claims inside a working document — and was found by comparing the two documents. That is the honest claim, and unlike a clean zero it can be checked.
Part one
What the human caught
36 errors. In every one, the machine's own checks had passed. They fall into six habits — and the single catch that no check of any kind could have made.
No. 25 · the catch no control would have made
A decision about presentation was carried out as a decision about content, and drove two full rebuilds.
“I'm talking about layout and presentation, not the document's outline or shell.”
RuleA ruling that references another asset is read back naming the dimension — layout, structure, content, voice or length — before any work.
Every guard passed on this one: files current, versions right, format clean, figures sourced. The output was wrong because the instruction was misread, and every later step agreed with the misreading. Only the person who gave the instruction could find it.
He asked one sceptical question
10 catches
The cheapest move and the highest yield. Every time, the machine's answer was fluent, consistent and wrong.
1
A file labelled canonical was stale; the real tracker sat unlabelled in the Drive root.
“Are we sure this is the best to use now?”
RuleVerify by modified time, never by a label or folder.
2
The quality-control document asserted the old scoring weights and cited itself as verified.
RuleRe-verify from the live file at the start of any session that acts on it.
5
Six retries fired against a service that was down and had said so.
“Confirm new credits aren't being added without authorization.”
RuleOne retry, then fall back and report the outage.
6
A workflow was proposed around the surface the machine could update, not the one both models can read.
RuleAnything both models must read lives in Drive.
7
A relay document was invented for the human to write notes into, where a live channel already existed.
RuleA new manual step has to justify itself.
11
Long close-out formatting buried the answers to short questions.
“Too much text. … I didn't read out of frustration.”
RuleMatch format to the question; short questions get the answer first.
32
The headline number on a new dashboard came from the wrong date column and overstated the search by about seventeen days.
RuleA derived figure names its derivation before it can be a headline.
97
The machine said Gemini couldn't read Drive and pasted 11 KB into chat for the human to copy. It could.
RuleBefore saying something can't be done, name the check that established it.
136
Two product figures in the profile draft were credited to a source document that did not exist, and a word crept into their wording that the original never had.
RuleA figure's source is read from the origin document, never from an earlier summary of it.
137
A company-lineage count was carried through nine profile drafts and never counted; the sources supported three names, not five.
RuleAny count inside a claim is counted against its sources before it ships.
He opened the finished thing and looked at it
4 catches
Found in the time it takes to open a file. No check that reads the source instead of the result would have found them.
17
Two resume generations were delivered as unformatted plain text — 297 hard-wrapped paragraphs.
RuleA resume is uploaded with conversion on and verified by counting its structure.
18
Scoring was scoped to the top twenty by fit when the agreed sequence named a different set.
RuleA ruling names the subset; no stage may assume one.
111
The public ledger page still said nine entries five days after the merge made it ninety.
RuleA merge rebuilds every surface that states the old count.
128
A resume called two pages ran to three in his copy of Word; the fit had been judged with no spare lines.
RuleA two-page claim states its measured spare lines; fewer than five is not two pages.
He wouldn't take a document's word for itself
1 catch
22
An item said a dry run was already clean; no log existed and the script was an old version.
Rule"Already clean" is valid only with a log entry or run time cited.
He used the new tool and saw it lose his input
1 catch
30
Recording a ruling hid the row, so his decision looked like it vanished the moment he made it.
RuleRuling and clearing are separate acts; a write is confirmed on screen.
He noticed a decision he'd already made hadn't landed
7 catches
The most expensive kind, because the cost is never the file. It is his confidence in the system, and it compounds.
36
Internal codes put in front of him in a question, a third time after he'd ruled against them.
“You're still using codes I don't have insight into. Third time now.”
RuleQuestions to the human are plain English with the options spelled out.
37
A caveat he had already overtaken was repeated three times in one session.
“I'm wasting time repeating myself.”
RuleA gate is stated once, logged, and not repeated.
42
Four documents each said "read this first", two carried opposite rulings, nothing said which won.
“Things go out of sync as new strategy unfolds.”
RuleOne entry point, later timestamp wins, propagate in the same turn.
43
He was offered two drafts that both broke his ruling on listing client logos.
“Neither has the logos listed out succinctly.”
RuleBefore offering a choice, check whether both options fail a ruling already made.
44
A comment fenced "for discussion" was built anyway.
RuleA fence covers everything after it.
71
A caveat he'd already acknowledged came back as a "new" finding.
“You should know that already.”
RuleA newly measured fact isn't new information if he already knows it.
121
The run sheet dropped the live demo beat he had planned because a prerequisite wasn't installed, instead of flagging the missing install.
RuleA missing prerequisite is raised as a prerequisite, never designed out of the plan.
He judged it by whether he could use it, not whether it was correct
12 catches
The largest group. The machine's work matched its own specification and failed the job. A machine checks against the spec; a person checks against the use.
61
Asked to check the tracker, the machine rebuilt all twelve tabs instead of fixing one defect.
RuleA check reports; rebuilding is a separate ruling.
62
A workbook self-test was validated in one program and failed in the one it lived in.
RuleVerify in the destination, or don't say verified.
63
A delivery problem on the machine's side was handed to the human as "two clicks."
RuleA delivery constraint is the machine's to solve or name as a blocker.
64
The same failing upload method corrupted the payload twice.
RuleStop after the first failure and change approach.
65
Two sparse columns were dropped that held the most valuable research in the system.
RuleJudge a column by what its filled cells are worth.
66
A rebuilt tracker was shipped as unformatted text — usable as data, not as a tool.
RuleName the trade-off before making it.
67
The rebuild committed the defect it had just diagnosed, then wrote it up as advice to the human.
RuleNever phrase a machine failure as guidance to the human.
81
Automated photo selection was built into the workflow on two good categories out of five.
RuleSelection of personal material is not automated.
83
Every site photo was flattened to 900 pixels; invisible on his monitor, obvious on the TV.
RuleUse each source at its own resolution.
84
Build notes described three elements that weren't in the page.
RuleEvery claim in a build note is read back off the page first.
85
A photo refresh renumbered the set and the page kept pointing at the old pictures.
RuleA refresh and a rebuild are the same action.
110
The demo desk looked weak; underneath, its "computed" tiles were typed and contradicted each other.
RuleA demo surface computes every figure from its own data.
Part two
What his controls caught
108 errors caught by the machine before the human saw them. Not luck and not a smarter model: each one was caught by a control he put in place, most of them born from one of his own catches above. The machine runs the checks. The human designed the net.
Read the file, not the label
32 caught
The measurement rule: a claim about a file is only valid if it came from reading the file. It is the first rule in the human's project instructions, and it exists because of his first catches — a file labelled canonical that was stale, a snapshot that cited itself as verified, a script header that said a dry run was clean when it had never run.
Show the 32 ›Hide ‹
3The fuller tracker was 63 bytes smaller; file size was nearly read as completeness.Rule: Diff the content; never infer completeness from size.
4A resume mirror had drifted from the original — stray characters, lost sections.Rule: A mirror is labelled as a mirror and names its original.
10An empty search result nearly became "your company is not registered."Rule: A null result isn't a finding until the query is proven able to find a true one.
12Two pending-writes documents existed; the older one's 23 rows were stranded for eleven hours.Rule: One pending-writes document at a time, ever.
14A formula audit was blind to array formulas and reported two false broken links.Rule: An audit that reports an absence must first prove it can detect a presence.
16Rows found through an aggregator were credited to the channel the human saw, inflating its share.Rule: The repair script reports the mismatch and refuses to write it; the human rules.
23A clean run was reported as a timeout because a leftover dialog box hung it.Rule: Read the log tab, not the execution banner.
34A page limit was assumed at about 2,000 words; the real ceiling was about 1,600.Rule: Render and page-count before staging any cut.
41A change list written from memory missed an edit.Rule: Change lists come from a mechanical comparison of the two texts.
46Two job titles were drafted while the real ones sat in the corpus.Rule: Search the corpus before drafting anything about the human's history.
48A bullet-length ruling was applied to half the page.Rule: Compliance is measured by rendering and counting lines.
49Change notes pointed to a record that had never been recovered.Rule: Anything cut is quoted verbatim in the change notes.
50A logger was shadowed by another script's function, crashing every apply for five days.Rule: One name prefix per script; after a failed apply, read the executions page first.
53A fix was written for an assumed cause while the real one sat in the error trace.Rule: Read the trace in full before patching.
58An item asked the human to paste a version already superseded.Rule: The live instructions field is compared against staged text before any paste is requested.
73Eight top-tier roles were marked closed from aggregator links; three were still open.Rule: Liveness is read from the employer's own board.
76A sweep filtered on the full country name when the data used a two-letter code — 84 rows lost.Rule: Check a filter against the actual values before trusting a low count.
78Intake was diagnosed as a sourcing problem; the emails were arriving and nothing was reading them.Rule: Measure the inflow before blaming the source.
79Three job boards were recorded as live channels when they'd never produced a row.Rule: A channel is live only when it has put a row in the register.
82Album membership was recorded as unreadable when the main library held it.Rule: Test "unreadable" against the primary store first.
87A one-off service error was carried as a standing breakage for two days, holding 77 rows.Rule: Re-measure from the live heartbeat before repeating a failure.
88Values were written into a formula column, silently moving thirteen roles up a tier.Rule: Read a column's anchor cell before staging into it.
91An open item was closed on a title match without opening either file.Rule: Nothing is labelled done without checking the target itself.
96A LinkedIn field was counted without the note that shares it — 216 over, not 8 under.Rule: Count the unit the platform enforces: the whole field.
100A role was carried as a send target after it had closed.Rule: Liveness is checked on the employer's board at send time.
101A run sheet shipped two open items that were both false.Rule: Open items are read from Drive when the sheet is written.
103Two documents said intake couldn't set geography; a formula already did.Rule: A stated mechanism is read, not inferred.
115A LinkedIn field was measured without its shared note and called under the limit; it was 90 over. A false ledger entry was logged against the earlier, correct count.Rule: Measure the exact text that will be pasted into the field.
124The ledger page and the site change notes cited a third-issue ledger document that was never written; entries 114 to 119 existed only on the page.Rule: A page names a source document only after that document is read in Drive.
135A recording date was copied from a memory note instead of the trim log.Rule: A date comes from a dated document, never from a summary or a note.
139A cleaned photo was never used because the rebuild worked from old file lists instead of the folders.Rule: Rebuild from what the folder holds, not from a list of it.
144The firm's legal suffix in saved notes was wrong; the state's own record said otherwise.Rule: An entity fact comes from the registry record.
Read it back before calling it done
18 caught
Nothing is reported as built until it is read back out of the thing itself. Born from the human opening two resume generations and finding them unreadable, and from build notes he caught describing work that wasn't in the file.
Show the 18 ›Hide ‹
15A character was edited after the syntax check, and the upload would have failed on first run.Rule: The syntax check is the last action before upload.
20A newer duplicate resume failed the format guard and sat one click from being sent.Rule: Run the guard; don't assume the newer file is the good one.
27A size field read as upload failure — a repeat of entry 3 forty minutes after citing it.Rule: The read-back is the only valid upload check.
28Tables failed to convert while the format guard passed.Rule: The guard checks tables too; no markdown tables in converted documents.
45The document the human reviewed never carried the ruled layout; only the export did.Rule: The export is read back for font, size and spacing before a version is announced.
52Migrated dates displayed one day early from a time-zone mismatch.Rule: Verify written values against source values, not counts.
60A reviewer brief rendered with escape characters visible.Rule: Anything for an outside reader is read back before it's called delivered.
69A Drive document was created empty as a placeholder and couldn't be filled.Rule: Build the full body first; create once; read back.
70The same placeholder mistake, six days after its rule was written.Rule: Rules go into the preflight procedure, not only the ledger.
86An old page layout was published to the preview.Rule: The file built and the file published are the same path.
89A storyboard was reported built when every text cell was empty.Rule: A read-back must land on a text cell.
95A decision record rendered in Drive as raw markup.Rule: Formatted documents go up as markdown and are read back.
102A document uploaded with its markup escaped; caught on read-back within one action.Rule: The read-back guard worked; no new rule.
107Five LinkedIn documents rendered with visible markup — entry 95 again.Rule: The markup check moved into the preflight's create step.
116A published agent walkthrough carried a half-edited working line in its total.Rule: Read the published page back before announcing it.
122A paste into the demo workbook landed at the wrong row; restored from version history and re-verified cell by cell.Rule: Read the selection back before any paste; version history is the first rollback.
133A storyboard note pointed to the wrong numbered risk after the list was renumbered.Rule: Cross-references are generated from the numbered list, not typed from memory.
143A resume upload failed twice because the file was retyped by hand on the way up; nothing wrong reached Drive.Rule: A file travels by a path that never retypes it.
Count it, don't copy it
11 caught
Every figure is computed from the source at the moment it is written, never carried forward. This ledger exists because the human wanted every error kept, attributed and counted honestly — and the count itself is the thing this control most often catches.
Show the 11 ›Hide ‹
19The ledger's headline didn't match its own table.Rule: Any summary count is regenerated from the entries.
38The ledger undercounted the human's catches by one, and four documents inherited it.Rule: The count is recomputed in the same edit that adds an entry.
39The consolidation miscounted its own new rows in the same edit that logged entry 38.Rule: Count mechanically, last, and paste both lists into the headline.
51The register was counted from a dashboard, twice — 36 rows "stranded" when the cells held 158.Rule: The register is the cells.
68A published page header carried typed counts; two of three were wrong.Rule: Counts on a page are computed from its data at load.
74A decision record said a rule would move about 44 rows; it moved three.Rule: Any count in prose is computed from the cells.
90The human-catch split was undercounted for three days.Rule: The split is computed by script from every entry's caught-by field.
113The to-do said thirteen entries were owed; nineteen were.Rule: The owed count comes from listing the staged files.
125Entry 121's add file said caught by the system; its own text says the human corrected it. The row wins, so the human's total is 33, not 32.Rule: The caught-by field is set from the entry's text, and the split is computed from the rows.
126A planning page's headline hours were never reconciled to its own tables.Rule: A summary figure is computed from the page's own line items in the same build.
129The site kept a register figure one higher than the resume after the resume changed.Rule: A resume figure change reaches the site's claim row in the same turn.
One record, newest wins, Denver clock
24 caught
One entry point; the later timestamp wins; every document carries Denver date and time; nothing exists unless it is in Drive or the project. The human's ruling of 6 September, after he caught four documents each claiming to be the one to read first.
Show the 24 ›Hide ‹
8A script didn't survive the session wipe.Rule: Nothing exists unless it is in Drive, the project, or delivered to the human.
24A resume was named current when a newer one existed seven minutes later.Rule: List the family, sort by modified time, take the newest.
29The project pointer sat two resume versions stale.Rule: An upload isn't complete until the pointer moves in the same action.
40Seven documents were dated in the wrong time zone with no time at all.Rule: Every stamp carries Denver date and time with the zone named.
55Two documents disagreed about which file was the register.Rule: One register, named in the entry point.
56Five documents were stamped at times estimated ahead of the clock.Rule: Read the clock in the same call that writes the stamp.
59A named work thread left nothing in Drive or the project.Rule: A thread with no artifact is reported by name, not summarised from guesswork.
75The entry point was updated in the project but not in Drive, twice in a day.Rule: Both copies are written and read back together.
77The entry point was nearly rewritten from an eight-hour-old copy.Rule: List the folder by modified time before the rewrite.
98Send times were read off a stalled sandbox clock, 45 minutes behind.Rule: Send times come from the confirmation email.
106A working note about a ruling was treated as the ruling.Rule: Rulings are read from the decision record.
108Two copies of the shoot document disagreed on how the video opens.Rule: A document is written once and mirrored by copy.
109The Deal Desk was flagged stale after reading a retired address.Rule: Read the page at the address the entry point names.
114An application record and its Gemini pass were stamped from estimates, not the clock.Rule: Read the clock before stamping.
117A site-opener document was stamped twelve minutes off the clock.Rule: Read the clock in the same step that writes the stamp.
118The site copy's second issue was stamped from an estimate — third time in four days.Rule: Never write a time from memory.
119"Drop in all facets of life" was read as put it back; the human meant remove.Rule: When an instruction can read two ways, take the reading that matches the last ruling and say so.
123Methodology change notes were stamped 23:45 when Drive shows 23:37.Rule: Read the clock in the same action that stamps.
127Resume rulings were written into a review document and never reached the to-do, which listed them as owed for a week.Rule: A ruling recorded anywhere updates the to-do in the same action.
131A ruling was stamped with an estimated time.Rule: A message's time is stated as a bracket between two clock reads.
132A queue document was stamped two minutes ahead of the clock.Rule: Stamp from the clock read immediately before the write.
138A handoff was stamped fourteen minutes ahead of its creation time.Rule: Read the clock before stamping any document.
141A corroboration record was stamped about thirteen minutes ahead of the clock.Rule: Read the clock in the same action that writes the stamp.
142A profile build's title was stamped five minutes ahead of the clock, and fixed within the minute.Rule: Read the clock in the same action that writes the stamp.
The nightly sanity check
6 caught
The end-of-day check the human put on a nightly schedule. It compares the records against the machinery and treats a system going quiet as a finding, not a quiet day.
Show the 6 ›Hide ‹
57The daily intake briefing was switched off and nothing noticed for a day.Rule: The nightly check pairs task health against the newest intake date.
80The automatic write-back crashed and nobody logged it.Rule: The bridge log is read in the nightly check.
92The overnight check invented a geography defect from a value it didn't recognise.Rule: Unrecognised is reported as unrecognised, never as a defect.
93Open questions were given ages in days with no date raised behind them.Rule: Record a date raised before showing an age.
94Two of eight checks never ran while the table implied all eight had.Rule: All checks run from one script that can't skip one.
112The published run sheet page stayed three issues behind its document.Rule: The page and document update together or the page is marked superseded.
Check before it ships
10 caught
The human's standing order: quality-check every response and deliverable before it goes out, and slow down rather than ship something unsynced. These are the catches made mid-task, before anything left the machine.
Show the 10 ›Hide ‹
9Three lookups failed silently in a scheduled run and were reported as no data.Rule: Label the failure in the deliverable; unattended steps use tools that need no permission.
21An item asked the human to "turn on network egress" without naming where the setting was.Rule: Anything asked of the human names the exact path, checked live.
26A client's work was nearly re-credited to another client on an ordering instruction.Rule: A claim's attribution travels with its source, not its position.
31The save button sat above the note field, so a ruling saved with its note empty.Rule: Input first, commit button last.
47The fix for entry 44 reverted the part the human had asked for.Rule: Re-apply the accepted part; list what was kept and reverted.
54A brief said only the human could edit a scheduled task; the session could.Rule: Check the session's own tools before saying "only the human can."
72Search text landed inside live source code; repaired before any save.Rule: Never type into a code editor to read it.
105Automated keystrokes replaced a live document's body; recovered the same minute.Rule: Confirm focus with a screenshot before any shortcut.
134The nightly check reported a check missing that had not fired yet, and built a finding on the absence.Rule: The nightly check looks for the previous night's run, never the same night's.
140A layout item was reported to the human as not done when the change notes recorded it done and measured.Rule: Layout status is read from the change notes or the render, never from page text.
The human holds the keys
4 caught
The machine never sends, never applies, never downloads and never files personal material on its own. The human's rulings gate all of it, so the machine stops itself at that line.
Show the 4 ›Hide ‹
13A batch of photos that were not evidence landed in the evidence folder that would one day be public.Rule: Every evidence batch is triaged on arrival; personal material leaves the tree in the same action.
99An application was attempted from the cloud sandbox instead of the human's browser; nothing was submitted.Rule: Applications run from the human's Chrome only; a human check is his step.
120A spreadsheet export link was opened in the human's Chrome, likely downloading a demo CSV without asking.Rule: Never open an export or download link in his browser.
130A scratch comparison image was written into the human's own Downloads folder.Rule: Scratch work stays in the machine's own space.
Two models checking each other
3 caught
The human runs two models and arbitrates between them. Each reviews the other's work in a fixed handoff format, and a disagreement is his to settle, never theirs.
Show the 3 ›Hide ‹
33Resume text edited inside a paste document reached Gemini with two ungraded claims.Rule: Any resume edit produces a new numbered document.
35Gemini applied a ban to a mention fourteen years outside its scope.Rule: A prohibition names its scope.
104The human's LinkedIn review notes sat uncaptured in a Gemini chat for three days.Rule: A Gemini return is captured the day it lands.
What this is evidence of
The oversight is not the tax on the speed. It is what makes the speed safe to sell.
36 catches is the measured size of the human role in a system that otherwise ran itself — small enough to afford, specific enough to staff. Six habits, not a second full-time reviewer.
And the other 108 are his too, one step removed. Every control that caught them exists because he required it: read the file, not the label; read it back before calling it done; count, don't copy; one record, newest wins; a nightly check; nothing sent, applied or downloaded without him. The machine executes the controls. Designing them is the job.