What shipped, and what the research found.

Signals leads say when they happened

Readings for agent clients that start the tool

Three releases in a day

• 1.12.0Live NVIDIA GPU readings and process memory
• 1.11.1Access-token files are written consistently
• 1.11.0A reader token for agents

Multithread v0.4.18

Multithread v0.4.17

Wordship has its own home, and deleting your account deletes your sign-in.

Deletion says exactly what it removes.

Muse agent adapter and native collaboration improvements

Search finds names and numbers, and says what it covers

semantic_search now ranks every record three ways at once, by the words in its text, by its title and by meaning, and fuses the three lists.

Windows updates wait for the old process to exit

Reversal training carries to familiar pairs only

Teaching the small model to handle reversed order on half of its familiar option pairs carried over to the other half in all four runs, and in none of the four to pairs it had not seen as options.

• Limit It shows a bounded training effect on one task and does not establish a general rule or an internal pointer. The miss on new pairs may reflect their difficulty.

Grouping a team's archived runs beats a randomly chosen member on average

Across 29 submission groups and 30 topics, averaging a team's rankings before adding one fixed reference run beats adding that run to a uniformly chosen single member, raising mean top-ten precision by 0.24 to 6.35 points under every completion of missing labels.

• Limit This covers the current archive files on 30 reused topics (78 of 81 differ from their listed checksums), is not a new method, and only 13 of 29 teams are positive on their own.

Acquisition stopped before evaluation

The first archived run's bytes did not match the publisher's listed checksum, so the frozen 81-submission comparison stopped before any evaluation.

• Limit This is a data identity failure. It gives no evidence for or against averaging submitted outputs.

An 81-run test frame and a counterexample on control design

Official metadata yields 29 eligible teams with 81 automatic submissions for a grouped test, and an exact synthetic case shows the earlier control design allows effects of either sign.

• Limit Metadata and synthetic work only: no new ranking was downloaded or evaluated.

Steadier turns, and your data in your hands.

Space Atlas, an on-demand disk navigator

Crash keeps conflicting reports out of stop evidence

Output format decides whether a learned choice survives reordering

Every version learned the task in its trained order. When the two options were listed in reverse, answers given in fixed positions stayed accurate, while answers written as option labels fell short of the accuracy bar, though they partly adapted.

• Limit It is a result about what the task demands, and it leaves open both the internal mechanism and whether reversal can be learned in general.

The small model could not pick the asked-for row at this budget

Asked to select one of two displayed probability rows, the unchanged small network missed the target in every run, reacting about equally to the row it was asked about and the one it was not.

• Limit The result holds for this model, task and training budget, and does not show the design is incapable of it.

Natural alignment loses to the mean of shuffled controls

The variants' natural fusion scores about 1.6 points of top-ten precision below the mean of 256 predeclared shuffled controls under every completion of missing labels, excluding the proposed advantage.

• Limit The controls are built from relevance labels, so their higher mean serves as a diagnostic and cannot be used as a retrieval policy.

The exact-pipeline transfer is unavailable

None of four later TREC-COVID rounds supplies all four original retrieval roles, so the transfer stopped before any rankings or judgments were acquired.

• Limit This is a feasibility result. It is neither a negative retrieval result nor proof that transfer is impossible.

Two related variants at a fixed weight: main comparison unresolved

With two documented variants in place of two repeated copies, the hybrid beats the neural-reranked source alone by 1.7 to 2 points of top-ten precision, but its comparison with the existing hybrid stays between minus 1 and plus 2 points.

• Limit These are development results on the same 30 reused topics. No untouched query was tested.

A relevant admission can still be an unhelpful exchange

On 30 TREC-COVID topics, swapping one minority-source result into a two-source hybrid's top ten cannot raise top-ten precision, because its seven relevant admissions displaced seven relevant results.

• Limit The planned later-label phase failed its preservation check and the follow-up was post hoc. The result covers one fixed policy on 30 reused topics and leaves open whether minority evidence is useful.

Equal weights keep results found only by the neural source out of the top ten

On SciFact, 21 known-relevant documents appear only in the neural source's list, yet on every query ten lexically supported documents outrank every result only the neural source retained, so none reaches the fused top ten.

• Limit This describes existing lists after the fact and tests neither reweighting nor reserved slots.

Your Run log can travel with you.

Deduce plays like the word game you know, and keeps your record.

Both games open with their own art, and Salvo is easier to follow on a phone.

32 releases in a day

• 1.9.29Signals handoffs carry each input's recorded scope
• 1.9.28Crashes recognizes stops recorded only by Windows
• 1.9.27Signals records which inputs each rule family inspected
• 1.9.26Saved captures can be inspected without a new reading
• 1.9.25On-demand change history before a stop
• 1.9.24Investigations keep their place between views
• 1.9.23Cited events open beside their Signals lead
• 1.9.22Signals leads name the event sample behind them
• 1.9.21Captures reuse the readings Signals already took
• 1.9.20Exact re-reads of one event by record ID
• 1.9.19Agents start from the reading the question needs
• 1.9.18Redacted answers name what they could not replace
• 1.9.17Readings carry the version that produced them
• 1.9.16The live header checks the connection without streaming records
• 1.9.15Connection health includes activity counters
• 1.9.14Reading durations include time spent waiting
• 1.9.13A busy launch slot reports busy, not a failed start
• 1.9.12Busy answers name their cause
• 1.9.11Capture finishing keeps other work responsive
• 1.9.10Saved-file work keeps readings responsive
• 1.9.9Relearning machine names no longer freezes other clients
• 1.9.8Handoffs report whether their prompt was included
• 1.9.7Saved-file failures say what went wrong
• 1.9.6One capture holds one version of the investigation
• 1.9.5Events reads forward from a moment
• 1.9.4Answers lead with the verdict and caveats
• 1.9.3Fault reports on both sides of a moment
• 1.9.2Kernel-WHEA reports on both sides of a moment
• 1.9.1The changes reading checks its window before claiming absence
• 1.9.0Stack answers return a compact index
• 1.8.0Leaner hardware-error timelines
• 1.7.0Precise timestamp boundaries

Daily Salvo and friend matches, open to everyone.

15 releases in a day

• 1.6.0Exact WHEA time windows for agents
• 1.5.0Hardware errors opens from a bounded preview
• 1.4.1Citable references for storm reports
• 1.4.0Storm readings count earlier-session reports
• 1.3.9Older storm windows on request
• 1.3.8Older WHEA report windows on request
• 1.3.7Bounded moment windows for events and faults
• 1.3.6Stack tells a new observation from a retry
• 1.3.5Log readings show how far back the log reaches
• 1.3.4Crash moments carry retention evidence
• 1.3.3Captures appear only when complete
• 1.3.2Kernel-WHEA reports listed under their timeline
• 1.3.1Exact WHEA record lookup
• 1.3.0Kernel-WHEA gets its own report timeline
• 1.2.0Hardware errors reads both WHEA channels

Play first. Sign in later, if you want to.

Four releases in a day

• 0.4.15More precise timed-out Claude peer receipts
• 0.4.14Clearer hook feedback when the worker fails
• 0.4.13Support for repositories before their first commit
• 0.4.12More accurate waiting feedback for Claude streaming calls

A new look: Chart Room, and clearer Salvo hits.

The moment, answered in one call and one click

The file looks after itself

Maps of the whole region, measured rather than drawn

Every covered county now has a map: 25 jurisdictions across Lake, Orange, Osceola, Polk and Seminole, projected from Census TIGER boundaries, with the worst deviation from the source at 0.62 rendered pixels.

Desktop, phone and agent access

trace_connection: routes between named things

A new MCP tool finds the routes between two entities, meetings or named patterns.

Five releases in a day

• 1.51.0The sketch keeps its key in place, and bend actions move to the bend's card.
• 1.50.0A clearer view of each bend and controls within reach.
• 1.49.0A bend strip within reach and edits that keep your place.
• 1.48.0A clearer 3D workspace and easier measurement inspection.
• 1.47.0Smoother entry, consistent inspection and original manual pages.

Private residents are no longer named

People who appear in the record only as public-comment speakers, commenters or neighbouring owners are now "a resident", and their home addresses "[address withheld]", across every reading and meeting record.

Six releases in a day

• 1.46.0Steadier entry, clearer diagrams and native frosted chrome.
• 1.45.0Clearer measurement choices, readable diagrams and calmer entry.
• 1.44.0Find your preferences in clearer Settings groups.
• 1.43.0See how your corner placement changes the measurement.
• 1.42.0Clearer measurement choices, steadier editing and visual help.
• 1.41.0Choose measurement points right in each bend card.

Nine releases in a day

• 1.40.0Measure a complete run around a rectangular column.
• 1.39.0A clearer run strip, with corrections that keep your place.
• 1.38.0A compact run strip keeps nearby bends within reach.
• 1.37.0Smoother length entry and steadier 3D controls.
• 1.36.0More reliable fraction entry and updates when you’re ready.
• 1.35.0Choose the direction your run starts.
• 1.34.0Groundwork for the iOS and Android test builds.
• 1.33.0A clearer run sketch and simpler direction choices.
• 1.32.0Plan directions and choose the fractions you work with.

Three releases in a day

• 0.4.11Offline version discovery and caller stop reasons
• 0.4.10Recovery advice that matches where progress stopped
• 0.4.9Attributed provider diagnostics

Nine releases in a day

• 1.31.0Standard keyboard focus for length entry.
• 1.30.0A calmer start to each length.
• 1.29.0Less clutter. More room to enter your run.
• 1.28.0More room for entering lengths.
• 1.27.0Fraction entry stays with the measurement.
• 1.26.0Fractions at your fingertips. Your bend stays in place.
• 1.25.0More room to read. Clearer controls.
• 1.24.0Choose a direction. See where the next leg goes.
• 1.23.0Consistent offset precision. Small offsets stay usable.

Six releases in a day

• 0.4.8Easier follow-up and discovery
• 0.4.7Waiting feedback for peer calls
• 0.4.6Peer answers survive malformed measurements
• 0.4.5Pending work stays discoverable in full briefs
• 0.4.4Clearer installation, readiness and follow-through
• 0.4.3Support reports from retained peer calls

24 releases in a day

• 1.22.0Room for every mark. Useful actions when a run is logged.
• 1.21.0A clearer review. A run you can share.
• 1.20.0Results when they are useful. A clearer Hide gesture.
• 1.19.0Steadier entry. Clearer diagrams.
• 1.18.0A smoother ruler. Clearer choices.
• 1.17.0Steady controls. Help at the right place.
• 1.16.0Less to scan. More within reach.
• 1.15.1Explore one detail at a time.
• 1.15.0See, understand and edit your run.
• 1.14.0Turn the tube in your hand.
• 1.13.1A steadier tube selector.
• 1.13.0See the tube you are working with.
• 1.12.1A lighter logo.
• 1.12.0The Bendr name, logo and look.
• 1.11.0Works with no signal.
• 1.10.0A frosted title bar.
• 1.9.0A dark mode.
• 1.8.0How fine the marks read, and two defaults.
• 1.7.1Undo and Redo look like Undo and Redo.
• 1.7.0A Settings button, and these notes.
• 1.6.0Undo survives the phone closing the page.
• 1.5.1Small things a review caught.
• 1.5.0Start new no longer loses the run.
• 1.4.0A run is opened, not just typed into.

Five releases in a day

• 0.4.2Clearer peer timeout observations
• 0.4.1Clearer setup and native launch reviews
• 0.4.0One product name from installation to collaboration
• 0.3.0Native collaboration in either direction
• 0.2.0Simpler setup and native collaboration

SciFact transfer: adding a source helps, simpler baselines lead

Adding SPLADE++ to four lexical lists raised mean binary nDCG@10 from 0.667 to 0.687 across 300 queries, but SPLADE++ alone (0.704) and a simple two-source hybrid (0.705) scored higher.

• Limit One collection with conventional binary scoring supports no new fusion method, generalization or significance claim.

Logged runs.

The trained models did not follow the local route weights

Rereading the saved predictions, the two models' choices sat further from each graph's actual route weights than a uniform guess would.

• Limit It is a look back at results already seen, with no new model runs, and identifies no internal mechanism.

The trained models barely respond when the goal changes

Across 512 paired prompts, swapping the requested goal shifted the models' route choices by about a hundred-thousandth of what an exact reference shifts, and their routing stayed slightly worse than uniform.

• Limit The tiny responses that remain do not show useful goal use, and cannot say whether they are noise or come from any internal mechanism.

A shortcut that could produce high success, and a test to expose it

Analysis showed that copying the requested ending and following the list of legal moves reaches 97.73 percent success while ignoring the graph's route weights, and a paired audit was designed around that shortcut.

• Limit Designed and not yet run. The shortcut is a worked control on reduced models, and it was not fitted to the trained ones.

Two small models reach the goal but do not route by it

Two small neural networks trained on the graphs reached the requested ending with 97.5 and 94.6 percent success, computed exactly over every graph and goal tested, yet their route choices were further from the ideal than unguided walking.

• Limit Reaching the goal often does not show the models learned goal-guided routes, and interpretation was deferred.

Preparing a transfer that can answer the question

The cycle fixed a new scientific-abstract collection, five systems and the metric bridge before inspecting any effectiveness result.

• Limit Preparation only: no retrieval effectiveness was measured.

Adding SPLADE++ lifts both query sets, with query reversals

Under a fixed top-ten diagnostic metric, adding SPLADE++ to four-source rank fusion raises the mean on both query sets under every allowed completion of missing judgments, though 9 queries favor the original fusion.

• Limit This is a finite result on heavily reused development queries. It shows no general superiority and gives no basis for routing queries.

No justified agent experiment in the existing material

An inspection of existing retrieval data and one real provenance question found that direct lookup and hash checking already answer it, so no new solver or multi-agent experiment was run.

• Limit This is a retrospective inspection by investigators who knew the project history. No preregistered experiment was run.

Correct answers and validity judgments under copied support

A single language-model verifier gave 4 of 4 correct answers and 16 of 16 correct validity judgments whether valid support was the minority or the majority, leaving no gap here for a multi-agent stage to repair.

• Limit Direct program solving could produce every correct output, so the result does not show the verifier relied on the evidence.

A balanced panel rules out the named answer shortcuts

Across six calls, a language model returned all 18 trace judgments correctly, a pattern no named position, constant or endpoint-only shortcut reproduces.

• Limit The panel reuses six traces from one arithmetic program, so it shows correct outputs on that program and does not measure general verification ability.

Correct certificate judgments with an unresolved shortcut

A language model correctly accepted one valid execution trace and rejected two invalid ones, before and after copies of a bad trace were added.

• Limit A simpler positional policy also scores perfectly here, so the result does not show step-by-step checking.

Designing the first trained-model test

A bounded test of whether a small neural network trained on these graphs learns to steer toward a requested goal was designed, and a separate check with no training showed its passes stay within an ordinary processor's limits.

• Limit Nothing was trained, so no model had learned anything yet, and all six predictions about internal mechanisms remained untested.

The option to change course later does not explain the split

Rechecking the saved values, all 216 pairs still called for different first moves when the final move was forced to be the same, so a later chance to revise does not explain the difference.

• Limit It holds for one problem family, horizon and reset price. Six single decisions changed when the next state could be seen, but that did not pass the controlled test.

When a memory summary is not enough over two episodes

Across all 27,225 reachable states of a small two-episode problem, 216 pairs share a compressed summary and the same best single move yet need opposite first moves once reinforcement carries into a second episode.

• Limit These histories were prescribed, so they say nothing about how often this happens.

Bounding what a compressed memory can know

An exact calculation of the range of goal chances consistent with a compressed memory settled the right action in 508 of 512 cases at one set of prices.

• Limit It uses one fixed price profile on one graph and supports no claim about average efficiency or the general value of extra information.

Recovery from a wrong ending when route memory is kept

On a small diamond-shaped graph, searches that first learned the wrong ending recovered in 231 of 256 runs when their route memory was kept, against 255 of 256 when the preference for that ending was erased.

• Limit It covers one small graph with fixed settings, and which intervention looks best depends on how recovery is measured.

Trying to reproduce a dropped value, without success

Sixty-four fresh language-model calls that varied the suspected causes produced no omissions, and all 256 retained values were correct.

• Limit Every task returned to its starting state, so copying the input would also score perfectly, and the earlier failure stays unexplained.

A model keeping a running record, with and without an archive

Over 48 language-model sessions every answer the record supported came back correct, but both conditions dropped the same known earlier value on identical input, and access to an archive repaired it at the next step.

• Limit It is one failure pattern on one pair of identical inputs, and gives no error rate.

A language model pulling records out of a history

Extracting records from 11 fixed histories, a language model reached the 55 percent ceiling that a two-record lookup allows under both prompts, so the added guidance showed no gain.

• Limit The task hit its ceiling, so it cannot show whether guidance helps, and it shows narrow model behavior, with no mechanism identified.

Pricing what a search should remember

An exact cost model comparing 20 fixed policies over 160 traces showed the cheapest one depends on the prices assumed: look information up when an archive is cheap, keep it when the archive is expensive.

• Limit The costs are declared assumptions, and nothing here tests a learned summary or a neural network.

How quickly a search recovers from memory built on a wrong answer

As reinforcement grew stronger, searches whose route memory started on the wrong answer recovered in 247, then 59, 3 and 0 of 256 runs. A sharp final answer appeared even without memory, so sharpness alone does not show memory at work.

• Limit It covers this exact model only, and did not test for a sudden tipping point or for trained networks.

Correcting the first graph results

Rereading the saved runs showed the combined system held on to a found route far more reliably than the comparison with a degraded guide (0.999 against 0.409) but found it more slowly, so the claim of faster discovery was withdrawn.

• Limit No new experiment was run, and those numbers came from an early controller that used the true goal to decide what to reinforce, changed before wrong goals were tested.

Loop-length panel built, then parked before any model calls

Eight matched programs passed their construction checks, but an answer imbalance meant an always-false guess could mimic a difficulty effect, so the solver batch was parked.

• Limit No model responses were collected in this cycle.

Complete collection, no observed error variation

Two prompt variants answered all eight small programs correctly, so the gate requiring observable errors and disagreement failed and the setup was parked.

• Limit This is a negative feasibility result for one small task and says nothing about solver accuracy in general.

Noisy provenance, incomplete empirical comparison

Three valid responses closely matched the noise-aware answers, but the provider refused the fourth call and the fixed workflow stopped with four inputs unsent.

• Limit With five of eight planned inputs lacking a valid response, the primary comparison is null.

A language model matches the exact Bayesian answers

With the collector fixed, a language model returned probabilities within 0.00000004 of the exact Bayesian answers on all 20 fixed requests, including the difference between copied and independent agreement.

• Limit This covers one small, fully specified synthetic model. It does not measure real-world calibration or general reasoning.

An instrument failure, preserved

The collector misread empty agent counters as activity and stopped the batch after its first response, leaving the planned test incomplete.

• Limit The stop came from a measurement bug in the collector, and no model reasoning failure was involved.

What evidence origin changes, and what it does not

In an exact synthetic model, knowing which reports share an origin improves probability estimates, while always protecting the dissenting minority makes them worse.

• Limit These are properties of a deliberately constructed model. No language-model gain was measured and no new aggregation method is claimed.

The observed association does not clear the next-step gate

Neither year's link between corroboration and relevance cleared the preset ten-point threshold across resampling, so the protocol stopped this branch and moved to a model with known truth.

• Limit Stopping was a resource decision and leaves open whether specialist sources are useful.

Better observations reveal a different ambiguity

SPLADE++ supplies many more judged examples than the lexical sources of top-ten documents no other source ranks highly, but for those documents, support from other lists cannot be separated from whether the old lexical candidates contained them at all.

• Limit This cycle counted candidates and judgments only; no ranking metric or fusion result was computed.

What counts as independent evidence?

After correcting the rank offset and tie handling, a tuned k still helps one six-ranker setup, but it lowers the original four-ranker setup and both transfers between the two query sets lose.

• Limit Results come from 43 and 54 reused queries, generated lexical rankers and partial relevance judgments.

Three releases in a day

• 1.2.0Read a leg to a wall.
• 1.1.1The insert circle, cleaned up.
• 1.1.0Insert a bend between two, and the version under the title.

The calculator as it first reached the phone.

Three releases in a day

• 2026.08.30.3Your first workout, guided
• 2026.08.30.2Lift Cards, even without a signal
• 2026.08.30.1Friends, with clear boundaries

Six more cities, SB 180 read again, and a faster site

Orlando, Kissimmee, Mount Dora, Eustis, Tavares and Montverde join the corpus: 83 harvested documents, six place dossiers, 25 cities in all.

The plates are already waiting

On a plate-loaded exercise, type a working weight and see a live plate summary before you leave the keypad.

Your drafts stay put

Set values now save as you type, whether or not the set is marked complete.

Three releases in a day

• 2026.07.31.3Finish on your time
• 2026.07.31.2Picture the work
• 2026.07.31.1Every workout tells its story

Rest that stays true

See a truthful end time and open detailed controls over your workout without stopping the countdown.

Set logging, dialed in

Weight, set type, muscle focus, and duration now live where you expect them, while the active row stays visible as you type.

Exercises, without guesswork

Browse and add exercises through one clear library, with muscle and training context before you choose.

Progress, while you train

Keep records and improvements visible in a live Progress Ledger, beside a calmer rest controller that belongs to the workout.

Four releases in a day

• 2026.07.25.5Start, in your style
• 2026.07.25.4Settings, one language
• 2026.07.25.3Display, composed your way
• 2026.07.25.2Lift Cards, lined up

Four releases in a day

• 2026.07.23.5Lift Cards, clear at every layer
• 2026.07.23.4Lift Cards, dialed in
• 2026.07.23.3Your Lift Card, your way
• 2026.07.23.2Custom exercises travel

Settings, made yours

A calmer Settings control center keeps your profile, training, privacy, support, and account controls clear without showing everything at once.

The League, up to date

Pull to refresh every League view, with standings that make the current week and a fresh start clear.

Your display, composed

Display now starts with one complete preview and reveals each layer of customization only when you want it.

Three releases in a day

• 2026.07.20.4Every update tells its story
• 2026.07.20.3Your training month, at a glance
• 2026.07.20.2Friends belong in the League

Rest cues, engineered

Rest timing now feels intentional: a rising three-count, a decisive finish, and cues composed around how you train.

First graph runs: steering toward a goal, and how it goes wrong

The first runs on one family of constructed graphs were summarized at the time as four predictions confirmed, one measuring instrument failed, and a mixed verdict on how a wrong goal signal leads the search astray.

• Limit That summary was later superseded: the faster-discovery prediction was unsupported, and the early controller had used the true goal when deciding what to reinforce.

The Water Line, and an agent surface rebuilt

The corridor's binding constraint moves from roads to water.

The atlas widens: Winter Park, Mascotte, communities

Ten corpus releases in two days.

The Mirror: The Observatory Observes Itself Being Observed

The agent surface now logs every tool call, accepts feedback from agents through a new MCP tool, and publishes what it reads at /agent-observatory.

The Attorney + Broker Projections: Lens Architecture Reconciled to Canon

The lens values agents use to request a stakeholder's view had drifted from the canonical roster, and this release restores it: synthesis plus seven stakeholder lenses, including broker and attorney.

Four releases in a day

• 0.13.0The Atlas Doubles: Constellation Expansion Across Three Counties
• 0.12.1The Visible Surface: Closing the Discoverability Loop for Browser-Class Agents
• 0.12.0The Mirror Tools: The Tool Surface Mirrors the Content Surface
• 0.11.0The Citable Contract: State, Definitions, and Claims Become Citation-Stable

Four releases in a day

• 0.10.0The Decoder Index: Every Named Entity Becomes Citable
• 0.9.0The Pattern Atlas: Named Patterns Become First-Class
• 0.8.0The Watch Resolves: Forecast Accuracy Surfaces
• 0.7.0The Watch: The Field's Forward Calendar

The Meeting Record Opens

The full standardized meeting record from January 2024 through May 2026 is now readable: 92 meetings of the Clermont, Leesburg, Minneola and Groveland planning and zoning boards, each at its own URL with its items, a public-comment summary, key signals and a source trail.

Five releases in a day

• 0.5.0Your Feed: Curation as Conversion Primitive
• 0.4.0Watchlists + Auth: Phase 1 Complete
• 0.3.0Semantic Search + MCP Agent Discovery
• 0.2.0The Field: Index Pages and Lens System
• 0.1.5The Corpus Opens: Depth + Design System

Four releases in a day

• Salvo, tuned from the first games between friends.
• 0.3.0Salvo rebuilt: hunt the fleet, then deduce each word.
• 0.2.0Streaks, stats and head-to-head records.
• 0.1.0Wordship opens with Daily Deduce, practice and friend matches.

Zoning Signal Observatory: First Publish

The Zoning Signal Observatory is live.