The real-time shape of my work
Every piece of work from the log, placed in the field it belongs to.
Each point is one piece. It sits with its field, and when it also draws on a second field it hangs between the two. The map grows every morning when the log does, so over months it shows where my skills are widening and where the hours actually go. Drag to turn it. Hover a point to read it.
Work with meSOVRAN Analysis · Latest summary
Updated October 16 · last 30 days · 116 pieces
Gregorio has been splitting his attention between building software and shaping how it reads. Software engineering and product and user experience each made up about 18% of the work he logged in the last 30 days; agents and automation, meaning software that runs on a schedule without anyone watching, about 15%. The mix says he can build a tool, make it clear, and leave it running safely.
Newer is the checking. Data and security and research and evaluation each accounted for about 12%: licence reviews before outside data enters a product, tests against answers known in advance, and reports that state what they do not cover. He has also been adding approval steps before an automated system sends or publishes anything. For anyone hiring him, that means work that ships and can be audited later.
Where the logged work went
- Software engineering18%
- Product & UX18%
- Agents & automation15%
- Data & security12%
- Research & evaluation12%
- AI safety & governance11%
- Strategy & operations7%
- Writing & content7%
Learning lately
- Reading very large files stored online in small pieces, without downloading a copy.
- Running several AI assistants at the same time, each handing back work for him to approve.
- Using an outside compliance checker on his own product, then publishing the gaps it found.
- Repairing scores that counted the same factor twice, and checking the fix against published readings.
- Testing a land change detector against invented scenes whose correct answers are known beforehand.
Where the effort goes
465 pieces of work over 103 days
Bars count the pieces whose home is that field. The line is the running total. A ring with nothing in it is a field the work has not reached yet. Fields are assigned from each entry’s own label, so a few pieces may sit next door.
Read the full log, day by day →
Read the map as a list
Agents & automation · 89
- 2026-10-15A second model provider behind one client, with generated numbers checked against the record
- 2026-10-15Three scheduled jobs silently failing, and a stuck run that blocked every later dispatch
- 2026-10-14Every message from a person gets a drafted reply, approved on a phone before it sends
- 2026-10-13An agent that answers from the record when no model is available, and declines off-topic asks
- 2026-10-09Three agents in separate checkouts in parallel, each returning a pull request for decision
- 2026-10-04A command guard that refuses destructive git operations anywhere in the agent workspace
- 2026-10-02Applying only the audit fixes that cannot break anything else, parking the rest
- 2026-10-02The portfolio's only automated self-check had stopped reporting for thirty-four days
- 2026-10-01A render engine giving every published dispatch its own narrated film
- 2026-09-30Porting an unattended posting pipeline and lowering its autonomy to one approval per post
- 2026-09-25Teaching a web observer to decline consent through a settings layer before reading a page
- 2026-09-21Rehearsing four months of automated entries before one human command published them
- 2026-09-20Hardening a scheduled writer so a missing evidence source stops the run before any model call
- 2026-09-19Drawing an operating org chart and its flow diagrams from live configuration
- 2026-09-19Restarting a scheduled agent that had exited at preflight every run for eleven days
- 2026-09-19Refreshing an executive agent's model of a business line from its current record
- 2026-09-18Publishing a daily work journal automatically behind a mechanical confidentiality gate
- 2026-09-13Grounding a claim in the live records, then correcting the agent's memory so it cannot recur
- 2026-09-12Fanning out a portfolio review over parallel research agents, with paid custom as the validation test
- 2026-09-12Moving a sandboxed agent's network reach into a build job that commits hashed extracts
- 2026-09-10Crediting a run to the identity that signed the commit instead of the files it happened to touch
- 2026-09-10An inventory no agent may edit, and a daily ledger that checks it against the commit record
- 2026-09-10Moving two scheduled agents off a laptop into isolated cloud runs, orders kept in version control
- 2026-09-09Mandating what a generated letter must do instead of the sentences it must contain
- 2026-09-09A four-seat pre-mortem packaged as a skill, with each plan scored as found and as rebuilt
- 2026-09-08Proving an approval gate by addressing its first live card to the operator
- 2026-09-08First live run of draft by agent, dispatch by hand, confirmation by observation
- 2026-09-07An approval gate that queued the request, built the link, then could not deliver it
- 2026-09-06An inventory of every unattended agent, then only the reversible fixes
- 2026-09-06One charter for per-folder agents, with escalation instead of cross-folder reads
- 2026-09-05Two agent sessions sharing one checkout, and the staging discipline that stopped them colliding
- 2026-09-04An operating agreement for a root-level agent, and the first time it wrote unasked
- 2026-09-03Scoring every candidate workflow on impact and risk before an agent is allowed to run it
- 2026-09-03A contract-review agent that prices every finding and what each fix costs the other side
- 2026-09-02Re-grounding the clock and the record before an assistant takes an irreversible action
- 2026-09-02Mapping the existing site with three read-only exploration agents before writing any plan
- 2026-09-01Conditioning image-to-video on two real frames so the model only fills the middle
- 2026-08-31Resuming a multi-agent writing run from cache so only the unfinished agent works again
- 2026-08-30Tracing a silent publisher to a scanner that had no news source for one theme
- 2026-08-29Choosing local scheduling over isolated cloud runs so an evening routine keeps its context
- 2026-08-29Scaffolding a new venture as numbered stages, with owner milestones instead of one big build
- 2026-08-28Encoding a seven-lens landscape framework as a runnable skill with explicit hinges
- 2026-08-28A research loop whose verification gate is tested by making it fail first
- 2026-08-27An approval queue that turns an agent's outbound action into a one-click email decision
- 2026-08-26A doctrine for motion prompts: the still carries the look, the prompt carries only the change
- 2026-08-26Letting an outreach agent send on its own, then verifying against the mail provider's sent folder
- 2026-08-25A publishing routine where the release script refuses any draft without an approval stamp
- 2026-08-25A routing layer over a portfolio of agent projects so a cold session knows where to work
- 2026-08-25An agent stopped at the protected branch and handed over two reviewable proposals
- 2026-08-24A copywriting subagent that can only quote proof from a verified list
- 2026-08-24Pooling many ranked sources after the single upstream feed went quiet
- 2026-08-21A permission gate that stopped a recoverable step after an irreversible one had already run
- 2026-08-21An outreach agent that sends the safe majority and interrupts a person only on exceptions
- 2026-08-20An agent clicked send when it meant to close a draft, and the control that followed
- 2026-08-19First live case showing the new agent retrieved existing plans instead of producing method
- 2026-08-19Designing a sales engine by making three architectures compete under an adversarial judge panel
- 2026-08-18Making an assistant ask for unverifiable facts instead of inventing them on formal documents
- 2026-08-18Driving onboarding from the booking record, with the assistant drafting and a human sending
- 2026-08-16Silencing scheduled-agent notifications unless there is a new draft or a failure
- 2026-08-16Validating generated content before publishing, so a failed generation fails loudly
- 2026-08-13Doubling an advocacy agent's cadence and moving its rotation to a per-run cursor
- 2026-08-04Structuring a research pipeline so numbered folders are the stages and each stage declares its inputs
- 2026-08-03Gating a retired worker cascade behind one opt-in flag so it stops burning tokens
- 2026-07-30Putting a drafting model behind a scheduled job so a content section fills itself on a cadence
- 2026-07-27Silencing four cron-triggered agent workflows while keeping the manual-dispatch one enabled
- 2026-07-21Diagnosing a false completion signal from a backgrounded batch job
- 2026-07-17Screening a shot list against an image-to-video model's fidelity limits before generating anything
- 2026-07-17Declining to fill a party's legal name on a binding document until a person confirmed it
- 2026-07-16Packaging a published agent architecture as a slash-invocable project scaffolding skill
- 2026-07-15Delegating a build to specialist subagents against stated acceptance criteria and an automated done gate
- 2026-07-15Stamping project state at the moment of work instead of inferring it from file timestamps
- 2026-07-12Building a five layer context architecture so an agent loads its identity and rules on every session
- 2026-07-11Recovering a parallel agent fan-out after one worker stopped short of reporting
- 2026-07-11Composing image prompts in layers so the category sets the scene, not only the subject
- 2026-07-10Diagnosing a tool connector that a non-interactive agent session cannot authorise
- 2026-07-07Wiring blocking hooks and scoped subagents into an agentic build loop
- 2026-07-06Drafting outbound supplier mail for a human to send instead of sending it unattended
- 2026-07-04Making output language a parameter of every prompt, including the deterministic fallback text
- 2026-07-04Weaving a new diagnostic lens into existing interview chapters instead of bolting on a questionnaire
- 2026-06-28Fanning a verification task across background agents, and recovering their results when delivery failed
- 2026-06-25A scheduled agent kept firing because its pause was written on a branch the scheduler never reads
- 2026-06-22An agent halting at a declared boundary before promoting a build to production
- 2026-06-18Turning a generation prompt into editable configuration with a save-as-default control
- 2026-06-16Pausing every scheduled agent at once while keeping manual runs available
- 2026-06-15Tracing outreach bounces back to contacts a research agent had invented
- 2026-05-09Draining a batch of chat approvals so an agent sends only what a human tapped
- 2026-05-07Repairing a chat approval queue that dropped late taps and double presses
- 2026-05-07Gating re-automation behind a small live run against real external services
- 2026-05-07Diagnosing why a scheduled outreach agent went silent for two weeks
Product & UX · 85
- 2026-10-15Report plates drawn from the land's own elevation, with the clearest satellite pass picked automatically
- 2026-10-15Framing product photos on the object in the browser, without re-cropping a single file
- 2026-10-14Repairing drag and pinch on a full-screen map, and promoting the output to its own button
- 2026-10-13Carrying one interface language through every entry path, with old screens kept for existing entitlements
- 2026-10-12Splitting one catalogue into two shelves, with a copy per shelf so edits stay independent
- 2026-10-12Reducing a product to one page per place, readable by a person and by an agent
- 2026-10-11Each report section opens with its question, one answer, and one image with findings marked in place
- 2026-10-10Withholding the score on built-up ground and declaring the scope before anyone searches
- 2026-10-08A visual page editor so a non-technical owner can recompose a page without code
- 2026-10-08One availability schedule, intersected with the live calendar, with a weekly cap and a required question
- 2026-10-06A listed lab where each experiment states its question, its start date and what is unfinished
- 2026-10-04Merging a separate water tool into the land reader so one app classifies the boundary
- 2026-10-04A grid-aware cover system so a social profile reads as one composed surface
- 2026-10-03Choosing a capstone a learner can picture, and treating the working circuit as half a device
- 2026-10-03Rewriting lessons so the check asks for a prediction the lesson actually taught
- 2026-10-01Sun at the minute, bare earth under the canopy, and a brief ordered by intent
- 2026-09-30Rebuilding a dense land report answers-first, then restoring the visible evidence behind the score
- 2026-09-25Reading every page with a classifier to find where the site told two different stories
- 2026-09-20Giving an offer page one URL per language so crawler and reader see the same words
- 2026-09-18Cutting a crowded masthead to five items and reordering a front page around one claim
- 2026-09-17Renaming a retainer offer across every surface that names it, in both languages
- 2026-09-14Shipping accounts, pricing and a paywall to production with the gate left switched off
- 2026-09-13A render cap was silently hiding published testimonials; proof moved next to the claims it supports
- 2026-09-08Pinning a link preview to one language because the sender's device builds the card
- 2026-09-04Measuring a crowded layout before redesigning it, then applying every ranked finding
- 2026-09-03Making a three-month minimum unmissable at the three points where a monthly price is misread
- 2026-09-02A comma removed from a headline exposed two stacked lines concatenating without a space
- 2026-09-02Collapsing a multi-page site into one scrolling funnel, with retired routes kept reachable
- 2026-08-31Directing a 3D hero object to a spec, then judging delivery by inspecting the file
- 2026-08-29Retiring monospace from body text and ranking the offer that returns money first
- 2026-08-27Giving a second audience its own page, one heading, and a promise above the fold
- 2026-08-26Reading the platform's published eligibility rules before adding a follow control
- 2026-08-26Unifying four navigations into one, then restoring the page that lost every link to it
- 2026-08-24Geography proposes a language, an explicit choice overrides it
- 2026-08-23An operator console that is its own system of record for a sales pipeline
- 2026-08-23Converting a whole site to one design system, then fixing the collisions it exposed
- 2026-08-19Rebuilding a landing page and a pricing page around two audiences and one low-friction first step
- 2026-08-13Retiring an unused self-serve tool and rebuilding the homepage as one clear offer
- 2026-08-11Shipping one argument in two formats: an editable document and a private web page
- 2026-08-06Selling a multi property, multi night itinerary as one bookable product
- 2026-07-30Codifying a photographer's observed technique from published frames into a generation directive
- 2026-07-28Rebuilding an informational page as a magazine feature, then resyncing its authoring surface
- 2026-07-26Selecting the exact face by index inside a font collection after probing its contents
- 2026-07-24Countdown board computing overdue days across month boundaries, linking out to the system of record
- 2026-07-22One link, three roles: a shared pricing sheet an outside party uses without an account
- 2026-07-21Making a production board reflect finished work that arrived outside the data model
- 2026-07-17Rebuilding a teaching game so the terms decide outcomes instead of decorating the mechanic
- 2026-07-16Turning a reading-only site into an application with a social layer and a persistent shell
- 2026-07-15Codifying a reference site's motion vocabulary as hard limits in a new design system
- 2026-07-14Rebuilding a teaching game around merge mechanics so the curriculum is the thing being played
- 2026-07-13Exporting an on-screen contract to a print faithful PDF with a wet signature route
- 2026-07-13Turning a terminology list into a playable tech tree so the mechanics carry the curriculum
- 2026-07-11Showing the category mix as a gauge so editorial balance is visible while planning
- 2026-07-10Consolidating an application's pages onto one header, navigation and attention pattern
- 2026-07-09Adding a sanitized progress panel so an external reader sees work in motion
- 2026-07-08Surfacing a blocking required action on the landing screen instead of last in navigation
- 2026-07-05Seeding one fully populated example record instead of describing what a page would look like
- 2026-07-05Collapsing every primary call to action onto one colour while keeping per item accents for structure
- 2026-07-05Re-spining a signed-in app so one assessment read is the home of every organisation
- 2026-07-04Cutting a signed-in navigation from three tangled products down to one spine
- 2026-07-03Folding two public identities into one hub where each contact channel appears exactly once
- 2026-07-03Pre-visualising a social feed as a grid before a single post leaves as a draft
- 2026-06-29Choosing a public-domain satellite scene over a generated one for a link preview card
- 2026-06-29Edit-in-place on cards for the three fields changed daily, modal kept for depth
- 2026-06-28Naming homepage sections after their contents instead of reordering the blocks
- 2026-06-27Hand-authoring prop contracts after the extractor read compiled output instead of sources
- 2026-06-26Scoping a portable UI kit out of an application so a design tool can render the real components
- 2026-06-26Turning a desktop-only admin into an installable phone app that syncs when it opens
- 2026-06-23Carrying one visual system from the landing page into the signed-in app and its link preview
- 2026-06-22Matching each rung of a service ladder to the action that fits its level of commitment
- 2026-06-22Showing each offer once and reducing the booking block to a compact selector
- 2026-06-21Repositioning a landing page and its booking page around one outcome-first category claim
- 2026-06-17Replacing a free-text material filter with a derived type taxonomy, keeping old deep links alive
- 2026-06-17Turning a static product demo into an explorable one with per-account data and a first-run tour
- 2026-06-17Fixing a language switcher that opened off-screen and was missing on the entry page
- 2026-06-17Extending one dark glass theme to a second page, with a different crop and parallax
- 2026-06-16Embedding a live tool full-bleed inside a static page, with authorship and sources stated
- 2026-06-16Inverting a page to a dark theme over a fixed photo backdrop without losing legibility
- 2026-06-15Regrouping a flat admin menu by task and promoting a global action out of a catch-all page
- 2026-06-12A public wall of saved readings so the first thing a visitor sees loads instantly
- 2026-06-11Deriving tokens from a live site so a new app reads as the same house
- 2026-05-25Placing the case narrative after the credentials block and before the biography
- 2026-05-24Distributing social proof across a long page so each quote sits beside what it proves
- 2026-05-22Pairing each buyer archetype with an illustrated plate so a typology reads at a glance
- 2026-05-21Building a single-page static site from copy that was already locked
Writing & content · 34
- 2026-10-12One source driving both language cuts of a product explainer, voice and captions included
- 2026-10-01Reissuing nine published compliance guides with dated correction notes in five languages
- 2026-09-22Publishing a findings dispatch in five languages, with every figure re-read from the live data
- 2026-09-21Separating a pain-first offer headline from an identity line about the operator
- 2026-09-20Rewriting a daily content prompt so every post stops making the same rhetorical move
- 2026-09-18Withdrawing a weekly newsletter promise from every surface while the list keeps collecting
- 2026-09-17Keeping an English job title inside the Spanish phrase instead of translating it
- 2026-09-16Rewriting a tool's front page so the first screen states scope and how to run it
- 2026-09-12Publishing the commitment and not the route, when a public roadmap starts as an internal plan
- 2026-09-09Lead copy that still sold a retired offer, and every self-claim of compliance removed in five languages
- 2026-09-08A dated correction on a public changelog instead of an eighth revision
- 2026-09-03Correcting a live article that conflated three separate articles of one regulation
- 2026-08-28A prose checker that fails drafts mechanically, and a gate that downgrades the artifact
- 2026-08-27Rewriting the offer section from the deliverables the work actually produces
- 2026-08-22Rewriting social covers from abstract nouns into concrete images for a wider audience
- 2026-08-21Correcting a headline that claimed the service does the automating instead of the teaching
- 2026-08-04Correcting draft article numbers, deadlines and penalty figures across a published back catalogue
- 2026-07-28Choosing plates for what they prove and rewriting each caption to describe its actual frame
- 2026-07-27Resequencing an About page so the commercial block lands early and the narrative closes on vision
- 2026-07-26Sourcing every claim on a graphic from the strings its destination already publishes
- 2026-07-13Sorting an operator's tool surface by kind before comparing options, then publishing it as a reference
- 2026-07-11Auditing existing drafts against the new standard before paying for a redraft
- 2026-07-11Rewriting the canonical brand text so every downstream generator inherits the new positioning
- 2026-07-08Cutting from a briefing the questions whose answers were already in the intake record
- 2026-07-05Attaching a consequence line to each scored dimension so a number says what it costs
- 2026-07-04Inventorying every string on a long page and testing each claim against the code before rewriting
- 2026-07-03Producing twenty posts in one batch against a declared content mix instead of one at a time
- 2026-06-28Writing an invitation against a checkable voice specification rather than an adjective
- 2026-06-23Reframing low scores as work to be done while leaving the number untouched
- 2026-06-22Rewriting the founding-purpose section to the current mission rather than the previous product's
- 2026-06-16Moving marketing copy out of components into editable content with code defaults
- 2026-05-22Editing a reference wall for attribution discipline and an ending that resolves
- 2026-05-21Enforcing a punctuation rule by audit, replacing each dash by its grammatical role
- 2026-05-21Renaming the entry offer across marketing pages, terms and the refund policy in one pass
Research & evaluation · 37
- 2026-10-15Refusing a language model as the independent check when it would judge from the same inputs
- 2026-10-14A sub-score that tracked how many people had reported species, not what the land holds
- 2026-10-12A scoring ceiling that slides with the dry season, checked against every published reading
- 2026-10-12Taking one variable out of two sub-scores so a dry season is not counted twice
- 2026-10-10Screening where adaptation works belong from terrain derivatives, with earthwork barred under canopy
- 2026-10-09Measuring 2050 climate deltas against each model's own historical run, not an observed archive
- 2026-10-08Telling undisturbed forest from plantation by keeping a class the import had discarded
- 2026-10-06Reading per-post reach and watch time to separate a format problem from an account penalty
- 2026-10-05Testing a land-change detector against synthetic scenes whose answers are known
- 2026-10-01A first-run walkthrough invalidated because its own reset switch stayed in the address
- 2026-09-25Declaring the population before the run, then reporting that nothing was observable
- 2026-09-20Sealing a weekly census's predictions before its first run, then splitting refusal from disappearance
- 2026-09-18Freezing a detector's rules, then measuring it on cases written after the freeze
- 2026-09-18Surveying a field, then choosing between three oversight designs by pre-mortem score
- 2026-09-12Two-pass verification of register candidates, with an adversarial agent trying to refute each entry
- 2026-09-10Running a research beat by hand after the scheduled run missed, with rejections recorded beside additions
- 2026-09-09An observed site can dress for the observer, so each grade now carries the conditions it was measured under
- 2026-09-03Auditing outside artefacts from the public record, and returning insufficient evidence instead of a verdict
- 2026-09-01Checking a generative video tool against the real constraint, then making the archive's inconsistency the concept
- 2026-08-31Testing a scored diagnostic with degenerate strategies until the shortcut stopped passing
- 2026-08-27Reading the operative regulation, then withdrawing an earlier recommendation in writing
- 2026-08-21Closing the loop on a generation pipeline with hook variants, a metrics store and a weekly digest
- 2026-08-16Auditing scheduled agents against their artifacts rather than their own completion reports
- 2026-08-13Auditing published content for mis-cited regulation articles and wrong penalty tiers
- 2026-08-11Checking every quoted figure against the live source and declaring what stayed unverified
- 2026-08-04Auditing how a land scoring model reads night brightness, and finding only a penalty
- 2026-08-04Marking each researched link by how it was verified, and rejecting a quotation spliced from two sources
- 2026-07-31Running a widget-auditing agent against ten live sites and finding every captured reply was interface text
- 2026-07-27Discarding a trigger test after finding the harness could not observe the variable it changed
- 2026-07-22Sampling a selection pool to show a repeated headline was chance, not a broken seed
- 2026-07-21Quality control pass on a bracketed exposure fusion batch, with geometry taken out of scope
- 2026-07-16Proving an image-to-video model preserves the real object before writing the integration around it
- 2026-07-14Catching a fabricated figure before publication and relabelling it as an explicit sample
- 2026-07-06Reading bounce notices and near-name replies as delivery failures rather than answers
- 2026-06-25Evaluating a 3D geospatial platform and finding the capability already present in the current map stack
- 2026-06-23Penalising thin evidence so a partly measured site cannot outrank a fully measured one
- 2026-06-11Checking every dependency and data endpoint in the spec before writing a line of code
Strategy & operations · 29
- 2026-10-14Spending on data licences is a cost; the restriction inside the licence is the barrier
- 2026-10-13Arguing the product could be cloned in a day, then ranking which moats are real
- 2026-10-08Restructuring a services page into three rungs and retiring the paid entry call
- 2026-10-06Designing a one-to-many talk as the sales channel, with the sign-up question feeding the demo
- 2026-10-03Benchmarking enterprise incumbents before writing a three-sentence pitch for a public feature slot
- 2026-09-20Capping live projects at five and registering every unattended job so silence means healthy
- 2026-09-19Finding that the instruments measured activity and the only automated check had died quietly
- 2026-09-19Splitting a portfolio by sales motion before deciding which kind of demand to buy
- 2026-09-09Free to claim your entry, paid to keep it verified, with no promise of compliance attached
- 2026-09-07Withdrawing an unsupported market claim, then writing the rule that would have caught it
- 2026-09-02Pricing a fractional retainer as a share of a finite month, so a discount buys fewer hours
- 2026-09-01A thirteen-week short-video campaign with ten pieces scripted and an executable production kit
- 2026-08-26Benchmarking a four-tier offer ladder against the market with parallel research agents
- 2026-08-25Back-solving the daily outbound volume from a stated close-rate cascade
- 2026-08-24Rewriting a section that addressed a buyer the rest of the site never sells to
- 2026-08-22Writing the success criteria and the test plan before paying for an outbound campaign
- 2026-08-21Rejecting model-written cold email against a conversion bar and scoping the missing craft
- 2026-08-20Letting contact-data coverage choose the outreach channel instead of guessing addresses
- 2026-08-19Sizing a prepaid membership against fixed capacity before setting the member cap
- 2026-08-03Rebuilding a homepage around the value exchange instead of the demo it used to sell
- 2026-07-31Re-ranking detected compliance signals by enforcement deadline and remediation scale, not ease of detection
- 2026-07-17Scaffolding a channel-growth operation as numbered stages, seeded by a cited research pass
- 2026-07-14Finding one architecture behind two service lines and packaging the overlap as a third offer
- 2026-07-06Choosing a payment rail by where the payer sits rather than by provider feature list
- 2026-07-01Splitting one venture into three pitchable stories, each matched to an outlet type
- 2026-06-21Building a value ladder where by-request tiers anchor the ceiling above self-serve bookings
- 2026-06-17Opening a free product beta to public signup while gating paid bookings behind a code
- 2026-06-12Weighing a diagnostic tool as public lead capture against an instrument used inside the engagement
- 2026-05-22Renaming a staged advisory ladder after its deliverables and making the first action a call
Software engineering · 124
- 2026-10-15Serving copies of upstream datasets from the product's own storage instead of calling them live
- 2026-10-13Expensive renders built on first view, capped per day, with earlier failures retried on their own
- 2026-10-12Counting what is actually unmerged across dozens of parallel sessions
- 2026-10-11A retry that never fired, and a dated local copy for when a public register stalls
- 2026-10-09Serving photographs at the size they display, with a cached redirect and a preloaded lead image
- 2026-10-08A read that timed out was recorded as no data, so it now fails visibly
- 2026-10-07A publish that replaces the live catalogue with one device's local copy, caught before it ran
- 2026-10-05Costing a precomputed regional layer set against readings computed on demand
- 2026-10-05A render reported success while writing no file, so verification moved to the artefact
- 2026-10-03A platform used frame one as the thumbnail, so a designed cover became a queue precondition
- 2026-10-03Adding a vertical output path to a film engine and proving the existing output byte-identical
- 2026-09-30First-person walking over true-scale terrain with canopy, water and buildings from open datasets
- 2026-09-25Sixteen deployments that never failed: builds refused because machine commit authors matched no account
- 2026-09-25A spend ceiling that stopped all work because the per-visit reserve equalled the whole ceiling
- 2026-09-22A verification that compared an identifier the framework never emits, caught mid-session
- 2026-09-22Halving deploy builds with a rule that skips pushes changing nothing a reader can see
- 2026-09-21A scheduled job died at launch because the OS checked the wrapper shell, not the interpreter
- 2026-09-20A converter with a contract: fixed output schema where every field cites its source line
- 2026-09-17Checking the content override layer before trusting compiled defaults, then rendering both locales
- 2026-09-16Declining to reassemble a large binary by hand, leaving one named manual step instead
- 2026-09-16Renaming a public artifact to its true scope while a permanent redirect keeps old links working
- 2026-09-14A booking paid outside the checkout rail now confirms with every side effect a capture runs
- 2026-09-09Refusing the sale when a provider quota cannot deliver the schedule being sold
- 2026-09-08Mail that reported success and sent nothing: discarded errors and a credential present but wrong
- 2026-09-06Empty completions traced to a reasoning block sitting where the text was expected
- 2026-09-05Alerting on checkouts that go quiet, added as a stage inside an existing daily job
- 2026-09-02Tracing a rejected scope to one never registered, then retiring a redirect flow the provider removed
- 2026-09-02Proving six components unreachable three ways before deleting them, against a recorded typecheck baseline
- 2026-09-02One module became the source for a price that had drifted across code, interface and published terms
- 2026-09-01Correcting a wrong-baseline branch comparison, then reading the working tree before grouping commits
- 2026-09-01Tracing appointments that confirmed without calendar events to an integration that was never configured
- 2026-08-31Shipping an animated 3D hero that never loads on phones and stays out of first load
- 2026-08-29Extending single-use invite links to accept a bounded number of independent submissions
- 2026-08-27Access rules written into the migration and proved by an executable test suite
- 2026-08-26Extracting the genuinely reusable primitives from an application into a design tool
- 2026-08-25Two daily checks had failed for five days, and neither failure was about what it reported
- 2026-08-25Testing the live product to find what already earns, instead of trusting its own documentation
- 2026-08-24Taking a whole site and its signed-in service into five languages
- 2026-08-23Removing the manual script from a scheduled generator so a missing period builds itself
- 2026-08-23Treating working credentials and an open feature gate as two separate switches
- 2026-08-22Date-only comparisons in UTC made the app roll to tomorrow after six in the evening
- 2026-08-21Moving a daily render and publish job off a laptop into continuous integration
- 2026-08-20Putting a large offline-first sync engine under version control behind a real build gate
- 2026-08-20Finding a funnel with no live payment path and connecting the rail end to end
- 2026-08-19Catching that a feature verified on a local server had never reached production
- 2026-08-19Collecting social proof through single-use invite links that publish on submission
- 2026-08-13Tracing a reported total outage to migration drift rather than a failing database
- 2026-08-13Moving paid bookings to a rail that will actually onboard the business
- 2026-08-13Flipping the default locale to English without breaking the regional translation delta
- 2026-08-11Rendering a programmatically generated document to catch layout defects a script cannot see
- 2026-08-06Exercising every route in a real browser and in a production build before calling it done
- 2026-08-06Replacing hardcoded assumptions with entry surfaces so every figure recomputes from input
- 2026-08-04Comping a paid unlock with a redeemable code held in the database rather than the source
- 2026-08-04Making an elapsed legal deadline render as past instead of counting down to negative days
- 2026-08-04Updating a recurring event's join link in the one row every confirmation email reads
- 2026-08-03Instrumenting a fix that appeared to fail, and finding a stale build under test
- 2026-08-03Automatic two-way sync across phone and laptop with no changes to existing mutators
- 2026-08-03Deriving published prices from the database that actually charges them
- 2026-08-02Checking the deployed revision against the last commit before shipping one feature
- 2026-08-01Generating client-ready PDF documents from a selection interface with the math engine kept separate
- 2026-07-31Diagnosing a deploy run from a laptop, which left the live site weeks ahead of version control
- 2026-07-31Committing one feature out of an entangled working tree by building the index file by file
- 2026-07-31One command that exercises every external dependency and reports which stages of the system actually work
- 2026-07-30Catching a not-found page that answered 200, then reverting the two fixes that changed nothing
- 2026-07-28Verifying a modal's dismiss, snooze and submit paths at the state layer instead of by simulated clicks
- 2026-07-27Tracing content loss to a public page hydrating the shared store with a stripped snapshot
- 2026-07-26Treating a scannable code as a release artefact: destination checked, decode confirmed
- 2026-07-24Seeding recurring obligations through a versioned store migration that never resurrects deleted rows
- 2026-07-22Per-action autosave with polled refresh and an explicit gate between draft and shared
- 2026-07-22Dropping a finished branch when the trunk had been rebuilt, then porting only the new layer
- 2026-07-22Rebuilding a rule-based prose engine so each reading sounds like written counsel
- 2026-07-21Parallel fan-out with polling replaces one-at-a-time clip generation
- 2026-07-17Fusing a day of exposure-bracketed interiors as cheap proofs, then finishing only the selected frames
- 2026-07-17Retiring a public contact value that lived in published data and overrode the shipped default
- 2026-07-16Collapsing a three-stage media pipeline into one action without losing per-step anchoring
- 2026-07-16Running one release gate and fixing two end-to-end tests that raced a shared database
- 2026-07-15Compiling an authored concept document into typed records at build time, with the prose staying canonical
- 2026-07-15Painting a catalogue site from cached local state so first render stops waiting on the network
- 2026-07-14Producing a branded PDF by rendering authored HTML through a headless browser
- 2026-07-14Driving a browser game through every state path in headless tests before shipping it
- 2026-07-13Removing an embedded token from a stored join link so automated confirmations cannot go stale
- 2026-07-13Making a page's encoding an invariant instead of trusting the server's content type
- 2026-07-11Migrating a controlled content vocabulary from four categories to six across the whole stack
- 2026-07-10Tracing live behaviour to deployed commits when the correcting change sat uncommitted locally
- 2026-07-07Designing a draft and approve state for outbound scheduling messages
- 2026-07-07Confirming a corrected record through the read path instead of trusting the write
- 2026-07-06Detaching a long-running local server so a reaped launcher stops taking it down
- 2026-07-05Shipping new seed records through a versioned client store migration and a real deploy
- 2026-07-05Giving a booking page its own share preview after it silently inherited the site's generic card
- 2026-07-04Turning a launch brief's acceptance criteria into one script that passes or fails
- 2026-07-03Auditing every published contact address so nobody writes to an inbox that is never read
- 2026-07-03Letting a session type carry its own availability so a twice-weekly offer escapes global business hours
- 2026-07-01Adding a new console domain by mirroring an existing entity through every layer
- 2026-06-29A blank unfurl traced to unpushed changes and a host variant with no DNS record
- 2026-06-28Replacing clear-then-insert publishing with upsert-then-prune so a failed run cannot empty a site
- 2026-06-27Pre-compiling JSX so a packaging tool stops inheriting the application's preserve setting
- 2026-06-26Assets excluded from version control still shipped, because the deploy step reads a different ignore file
- 2026-06-25Stopping an application crash caused by a browser translation extension rewriting the page
- 2026-06-25Fixing a magic-link login that looped because the session arrived in the URL fragment
- 2026-06-25Two share cards that failed for unrelated reasons: an image renderer and a missing DNS record
- 2026-06-25Making a local-first admin safe to edit from a second device before building the phone interface
- 2026-06-25Auditing a codebase whose build was configured to ignore its own type and lint errors
- 2026-06-23A public gallery stayed empty because serverless writes went to memory that dies with the request
- 2026-06-23Modelling an agreement as states: draft, sent, signed, awaiting payment, upgrade
- 2026-06-22One agreement template, a variable dictionary, and a validated generator producing PDF and Word
- 2026-06-22Diagnosing builds that completed but never went live: commit author verification
- 2026-06-22Confirming a release by what the page renders, not by what the HTML happens to contain
- 2026-06-21Typecheck, production build and a live content probe before calling the deploy done
- 2026-06-21Aligning localized metadata, structured data and the social card with the new positioning
- 2026-06-18Deleting a duplicate subsystem after finding the same capability built in parallel
- 2026-06-18Enforcing push-before-deploy with a single ship command
- 2026-06-17Making shared links render a card by referencing the generated Open Graph image in metadata
- 2026-06-16Separating pushed from live when a production deploy never picked up the branch
- 2026-06-16Building a native booking flow where payment upfront confirms the slot
- 2026-06-15Serving email cover images through a proxy that re-applies the publication check
- 2026-06-15Adversarial review of a new subsystem before deploy, then live checks after it
- 2026-06-12Choosing a background-removal model by its licence before writing the feature
- 2026-06-11Six named checkpoints and a ship gate instead of one large final commit
- 2026-06-11A scaffold that reported success and produced nothing: exit status lost in a pipeline
- 2026-05-24Verifying a rewritten page by element counts when the screenshot caught it mid-render
- 2026-05-22Removing a self-serve purchase path from the page while leaving its integration wired underneath
- 2026-05-21Diagnosing an OS privacy control that stopped a preview server reading its own directory
- 2026-05-09Resolving merge conflicts by file semantics: derived state overwritten, append-only history unioned
- 2026-05-07Replacing mutable counters with metrics derived live from systems of record
AI safety & governance · 31
- 2026-10-14Tests that keep a generated document unsigned by any person and free of its internal verdict word
- 2026-10-09Analytics that load only after consent, with decline clearing what was already stored
- 2026-10-08Replacing a generated stand-in for a real person with a cut-out from their own photograph
- 2026-10-06Marking a generated product image as a reconstruction before anyone sells from it
- 2026-10-04Declining a growth tactic that scanned others' sites, building lawful distribution instead
- 2026-10-03Running a newly adopted compliance linter against the product itself, then publishing the gap
- 2026-10-02Adopting an external compliance linter instead of building a jurisdiction-by-jurisdiction rules engine
- 2026-09-25Writing publication rules first, then bringing every public statement into line with them
- 2026-09-21A sealed public record of every time an agent was stopped, and by what
- 2026-09-21Letting an automated publisher name its own agents while client work stays unnamed
- 2026-09-20A narrow standing authorisation letting an agent correct facts about itself in a public thread
- 2026-09-19A guardrail refusing three drafts that asserted events absent from the case record
- 2026-09-16Correcting which EU AI Act obligations already apply before the claim reached public copy
- 2026-09-13Matching a published review method to the external-reviewer standard a frontier lab proposed
- 2026-09-12Running a public agentic-security checklist against our own fleet, after auditing each source licence
- 2026-09-09Publishing the measuring system as the first entry in its own public register
- 2026-09-08Revoking an agent's authority to send left already-queued messages armed
- 2026-09-03Grounding a compliance auditor in verbatim official text, then checking the law had not been amended
- 2026-08-31Blocking an automated post whose citation markers resolved to nothing
- 2026-08-26Bringing enforced gates in line with the oversight the product promises in public
- 2026-08-25Reading the AI Act article by article after the headline said the rules were delayed
- 2026-08-23Catching an example that cited a penalty tier the regulation does not contain
- 2026-08-21Writing privacy and terms pages only from the data flows the product actually has
- 2026-08-13Putting the agent disclosure above the hero, before any claim the site makes
- 2026-07-30Reframing an audit tool's claim from violation to deadline after checking when the obligations apply
- 2026-07-21Retracting agent-published findings until a person had verified them
- 2026-07-16Letting a blocked production deploy stay blocked until a person authorised it, then verifying the release
- 2026-07-15Auditing a product spec line by line against a standing safety policy and its data model
- 2026-07-12Accepting a guardrail on a privileged write and finishing the task with no credentials at all
- 2026-07-04Matching an organisation's map against public regulation with deterministic rules rather than a model call
- 2026-05-09Keeping merge to main outside the agent's reach so deploys carry a human signature
Data & security · 36
- 2026-10-15Reading multi gigabyte archives over the network in pieces, with no local copy
- 2026-10-15Diagnosing failed data sources by record census and by running each adapter alone
- 2026-10-13Reconstructing site visits from a geotagged media archive, triaged by hand before any record is created
- 2026-10-09Auditing candidate ground-truth datasets by licence, then mapping how blocked ones could be obtained
- 2026-10-07Stating the exact window a forest record covers, and treating a missing source as silence
- 2026-10-06Missing files read as loss turned out to be a curator's selection, so the restore was reversed
- 2026-10-06Completing a dataset migration with a declared, capped transfer verified by a range read
- 2026-10-05Auditing a public geodata service's licence before wiring it into a product
- 2026-10-04An explicit field allowlist on a publish path that had been leaking a private cost field
- 2026-10-02Cookieless analytics reporting every visit as a single page because no session identity persisted
- 2026-09-25Creating a webhook through browser automation without the secret ever entering the transcript
- 2026-09-21Counting machine readers of a public endpoint while excluding your own traffic, visibly
- 2026-09-20Auditing every version a repository's history can reach, not just the files at the tip
- 2026-09-18Resolving organisation names to domains under an acceptance rule, then a read-only census
- 2026-09-14Rotating a credential so its value never enters the agent's transcript
- 2026-09-06Entering a secret without it reaching shell history, the process table or a transcript
- 2026-09-03Tracing a local secrets file overwritten with a placeholder, then blocking the command that did it
- 2026-08-31A group report that stays closed below four participants, and capture limited to the score
- 2026-08-28Catching silent fallback and absence-as-fact in an open-data evidence pipeline
- 2026-08-27Deriving every collected signal backwards from the one screen that reads it
- 2026-08-27Verifying user-submitted URLs server-side, including the mapped forms that slip past a regex
- 2026-08-25Trusting the address validator over the assumption that a message had been delivered
- 2026-08-24A weekly threshold check was querying a key prefix nothing writes any more
- 2026-08-19Tracing a flooded booking list to overlapping availability rules rather than to the code
- 2026-08-16Keeping an assistant out of the credential path and shipping a connection check instead
- 2026-08-02Defaulting a missing collection to empty so a partial snapshot cannot leave stale state
- 2026-08-01Routing payment details through the private operator channel so they cannot reach public tables
- 2026-07-21Tracing repeated file corruption to a shell heredoc and moving binary writes off the shell
- 2026-07-14Writing follow-up notes into the record layer that owns them, after reading current state
- 2026-07-05Separating estimate templates from recorded actuals in a monthly operating ledger
- 2026-07-04Backing a sovereignty claim with customer-held model keys and one graph per organisation
- 2026-06-29Two independent state fields collapsed into one summary number, so the label overstated it
- 2026-06-23Merging a recurring spreadsheet drop into a pipeline without creating duplicates
- 2026-06-18Snapshotting a device-local operator store to private storage on every publish
- 2026-06-16Taking a deployment credential from the clipboard into config without printing it
- 2026-06-11Fusing seven keyless open data sources into one scored dossier for any coordinates on Earth