Repository navigation
Releases
Poisson Tasks 0.4.2
Fixes
-
In a narrow terminal nothing folds back to the left edge that could keep its indent, and nothing is split inside a value (#33). In both looks, on a terminal, a line breaks only at a space, keeps its indent when it wraps, and never splits a path, a URL, a command or a quoted value; a piece too long for any line is left whole for the terminal to fold. At 60 and 80 columns: the verdict line,
error:andnote:lines, the lines of a verb's questions (such as the note underinit's question on the model's name) and every line of the guided session keep their indent in the pretty look, where the terminal started them again at the left edge, and the session wraps its rows and its numbered steps in both looks, where it wrapped nothing; the reference solution's call and its reply, a JSON value on your own tools among them, break only between their members, where a quoted string was cut in two; what the model is shown, the tools and the wrong answers wrap at their indent; a command to copy, broken before its flags at 78 columns, is broken to fit a narrower terminal, a flag never parted from its value; and in the plain look the examples and exit codes of--helpwrap at their column. Piped output is unchanged, and--jsonis never wrapped. -
An unnamed set of your own tools is called by its file's name (#33). The settings line,
init's closing line and the guided session's summary saidyour tools (<the path in tasks.toml>), which a long path ran to two lines on each; they now name the file (your tools (petstore-openapi-3.0.json)), anddoctorgives the whole path once. A named set is called by its name, as before. -
A run stopped before its first model call no longer prints a raw
Event loop is closed(#52). When the start-up watchdog, a cap or Ctrl-C cancels an episode, the harness kills its processes and does not wait for them. If one of them still held its output pipes when the run ended, they were closed later, after the run's event loop had closed, and three raw lines (Exception ignored in …, a traceback,RuntimeError: Event loop is closed) reached stderr, how often depending on how busy the machine was. A run now waits for each process it started to be reported ended, and closes it, before its event loop closes. -
The versions Poisson Tasks is tested with are in the repository, and CI installs them (#40). Only
verifiersandtiktokenare pinned, so a fresh install takes the newest release of everything else the package allows, which can be newer than what a release was tested with.constraints.txtnow pins every package a checkout installs at the version the whole suite last passed on, and each release carries it at its tag:pip install poisson-tasks -c constraints.txtinstalls those versions and no others (docs/before-a-run.md). The package's own requirements are unchanged, floors and no ceilings, so installing into an environment you already have resolves as before, and every published Prove still installs beside it. CI's jobs without Prove install from the file; the jobs with Prove do not, so a new release of a dependency is met there first.
Docs
- The README leads with install, a first run and reading the card (#41). It is about 15 kB where it was about 19 kB. The reference material it held moved to the docs pages it links: each dial's levels and depth's full numbers to
docs/knobs-and-arms.md, and the cost and footprint list to a new page. It keeps the dials with their evidence, the published cards, the commands and what is not claimed, and links to poissonlabs.ai. - New page, for the site's sidebar:
docs/before-a-run.md, "Before a run: what it costs, and what it changes on your machine". Every item a first run writes outside your folder (the tokenizer file, uv for the harness, verifiers' cache and the temporary files), the limit on open files a run raises, the commands verifiers installs besidepoisson-tasks, and what goes over the network. - New page, for the site's sidebar:
docs/own-tools-reference.md, "Your own tools: how the responses are made, and the contract hash". The reference half ofdocs/own-tools.md, moved whole: the answers to calls, reads and writes, the errors in each format's idiom, the store, the reads on the bad-data dials, and how a tool file's contract hash is computed and what moves it. - The densest docs pages are in shorter sections (#51).
docs/knobs-and-arms.mdgains the fourteen dials at a glance, which recipes reach which dial, the turn budgets as a table, and checked blocks for the pin ledger on the block knob and for the two refusals ofbaddataandlagged.docs/own-tools.mdkeeps the guide (what you can bring, a first run, the refusals, the dials, the settings and the card) in subsections, lists and tables, with checked blocks for the[own]tableinitwrites, a flow task andverify's refusal, and links the reference page for the rest. No fact or figure was dropped.docs/calibration.mdis unchanged: it is rendered by the card code, and a shorter rendering waits for a release that changes that code. - The contract hash's example is corrected. Among what does not move the hash, the page read "
éagainsté", the same character twice; it now reads "\u00e9againsté", a JSON escape against the character it stands for.
Tests
- Existing tests were edited, each in its own commit: the test that holds the exact set of pages under
docs/gainsbefore-a-run.md, and thenown-tools-reference.md; the test that looks for the504and429statuses on the own-tools page looks for them on the reference page, where the errors table moved. - More existing tests were edited, each in its own commit, with no assertion changed: the two tests that compare the pretty look's first line and error lines on a terminal with the plain look's, and the two that read the guided session's
latest cardrow, read a line that wraps as the one line it is; and four tests of the model-B install start its driver from the package the test imported, not the checkout'ssrc/, so they pass where the package is installed from a wheel beside the checkout's tests, as they did not (#53). - New tests: the wrap rule, both looks and the guided session at 60 and 80 columns (
tests/test_fx_layout.py), a run stopped while its bootstrap's pipes are still open (tests/test_fx_run_reaps_children.py), and the tested versions and which CI jobs install them (tests/test_fx_constraints.py).
Poisson Tasks 0.4.1
Fixes
exportsays what theevaldisplay means (#30). Theevalcommand thatexportprints is verifiers', and so are the words it shows:[success ]beside an episode means it ended without an error, whatever it scored (the score is thereward=beside it), and the usage line counts each conversation once, so it shows less than the run spent, because each model call sends the conversation again and is billed for it.exportnow says so in a second note (and in--json, asnotes), anddocs/verifiers-export.mdsays where the real figure is. The display itself is verifiers' and is not changed.- Reading a card no longer builds it, a finished run leaves no lock file, and a run keeps the uv you have (#34). In the guided session, "Read the card", the word
cardand the run view read a built card by runningcard --force, which wrote its four files again. They now draw the run's owncard/card.mdas it stands, run no verb and write nothing; a card the run has gone on from (made from what ran, and the run finished since) is still built again.runremoves itsrun.lockwhen it ends, whether it finishes, fails or is stopped with Ctrl-C, and only the run that holds the lock removes it: a run that is still going keeps its lock, and a run that finds the folder held leaves the holder's file alone. verifiers' bootstrap ensures the newest uv before the first episode of every run, and with pip kept out of your user site (whichrundoes on purpose) that meant astral's installer was fetched on every run, with a uv already installed, and replaced it. A run now uses the firstuvon the harness'sPATH(~/.local/binfirst) when it takes--scriptinuv syncand inuv python find, the two commands the harness runs, which is uv 0.6.10 or later, and fetches the installer only without one. The plan'sharnessline,doctorand the README say so. - A run's card says what the run recorded (#35). The card of a run made with
runno longer says a bare not recorded for concurrency and budget: for the concurrency it says that the figure is in the run'srun.log, one line for each part of the run, and for the budget it says why there is none (a run has a spend or token cap and not a budget, and no file of the run folder records the cap). It no longer carries the paragraph on resumed runs when the run was never resumed, and the card of a run on your own tools says what an intent is where it uses the word.initsays that the estimate it shows before the cap, on your own tools, is the published card's usage on the built-in tasks scaled to the size of your tools, and so does theestimaterow of the planrunprints, where it said only that the figure was the card for the model. The published cards and their page are unchanged.
Poisson Tasks 0.4.0
Reads that say ok and are wrong
- Two new dials, each experimental until a model has been measured on it.
baddatais how little a bad read gives away: one read's first answer saysokand is wrong, and the same read made again is right from then on. At level 1 the answer is for something other than what was asked, and shows it; at 2 it is another value with nothing on it to say so; at 3 it is right but for one small thing, a stage one short or a number off by one (levels 0 to 3;baddata:b2).laggedis whether the first read after your write still shows the record as it was, so a model that believes the read over the write's own answer writes again (0 and 1;lagged:g1); it is offered on your own tools, since the built-in ledger has no read of a record and refuses a repeated stage call. Asked for in a comment on the launch. - What the model sees. On the built-in tasks a
baddatarequest names both records of the ledger and two reads, where other requests state the record and the goal: the first read's answer names the record and the second the stage to raise it through. On your own tools the request leaves one argument of the write unstated and says which read's answer holds it; alaggedrequest is the write, then a read to check that it took. Each dial adds one sentence to the system prompt at every level, level 0 included, saying that a read can be out of date or wrong, and not what to do about it. - What catches it. Nothing new: the end state. A model that trusts
baddataadvances the wrong record, stops a stage short, or sends the wrong value, and fails strict and leaky. A model that trustslaggedwrites twice, and fails strict; leaky passes it, as it does a duplicate onlanded, so alaggedtable is read on strict. - What a model that checks does. Reads again. Nothing in an episode charges for a read, and reading until two answers in a row agree is right at every level. Neither dial has a control arm: a model that trusts the first answer makes exactly level 0's calls, and the one extra read is made only by an episode that checks, so level 0 is the control.
- The card shows what the model did: acted on bad data and read again after a bad answer, the last two columns of every table on a card that holds either dial, and on no other card. Neither gates a card.
card --jsoncarries their counts and shares in each level that has them. - On your own tools, what a set needs.
baddataneeds a read whose answer can hold a value a write takes; level 1 also needs an answer that names what it was asked for, so the Petstore example reaches 0, 2 and 3.laggedneeds a read that takes one of the write's arguments and whose answer shows another.initanddoctorsay so on a set that does not have it. On these two dials only, a read's answer also says what it was asked, where the tool's output has a place for it; on every other dial reads answer as before. landedis narrowed. It needed a read that takes one of the write's arguments; that read's answer must now also be able to show one of them, since on some writes the read answered the same before the write and after it, and no model could have told whether the write landed. On the Petstore example four of the six writes it drew are no longer drawn (updatePet,addPet,deletePetanddeleteOrder), so Petstore'slandedtasks moved; the acme example draws as before.- Every task id moved, and no other task did. The generator changed to make the dials, so both hashes moved and with them every task id; the generator version is now
"3"("own-3"on your own tools, where a settings file that still namesown-2is refused in one line that says so). Every other task keeps its request, tools, goal, reference solution and answers under its new id. A run folder started with 0.3.x cannot be resumed with 0.4.0 (runsays so before anything is spent);card --comparebetween the two says the generators differ; a 0.3.x card still builds and checks. The published cards are unchanged and still not current. - Reaching them.
generate --armandexport --to verifierstake every new arm,lagged's on your own tools;[measure] preset = "custom"withknob = "baddata"runs and cards it on the built-in tasks, wherecontrols = trueis refused with the reason it has no control arm, and on your own toolslaggedtoo, at the levels the set reaches. No recipe changed; a later release measures both on current models. generatesays why it refuses an arm.generate --arm lagged:g1on the built-in tasks is refused in one line that names your own tools.generate --arm landed:l1has answered "an unexpected failure inside poisson-tasks" since 0.3.0, with its reason only in the log; it now prints that reason as its one line too.- Library names added:
arms.EXPERIMENTAL_KNOBSgains two, the record fieldsacted_on_bad_dataandread_again_after_bad(each left out of a written record where it does not apply),simulator.State.lagged,generator.BADDATA_NEEDS_A_CARRIER,generator.BADDATA_NEEDS_A_NAMED_KEYandgenerator.LAGGED_NEEDS_A_SHOWING_READ, and pin-ledger row N36, which holds the two reads of abaddatatask and what each answers.
Tests
- The two dials, what their reads answer, the narrowed
landedand the card's columns are held by tests written before the build (tests/test_bd_*.py): the arms and their refusals, the request, the answers and the reference solution at every card seed on both splits, the episodes that trust a bad answer and the ones that read again, the own-tools tasks on each example set, the figures, the experimental surfaces, and the docs. The generator's hash is recorded again at its new value, with a kept file of the two dials' digests (tests/test_bd_kept.py), and the own-tools digests are recorded again too, Petstore'slandedones among them. - Existing tests were edited, each in its own commit, where one named a count, a label or a list this release moves: twelve dials to fourteen, thirty ledger rows to thirty-one, the generator labels to
"3"and"own-3", four experimental dials to six, the knobs thatcustomanddoctorlist, the two-word labels ofdoctor's new rows, and two tests over every knob's request, which now treatbaddata's as they treatspread's.
Poisson Tasks 0.3.3
Fixes
- The Prove route no longer reads the text of
diff.scripts/release/prove_e2e.pyandexamples/prove_sweep/run.pylooked fortruncated=in whatpoisson-prove diffprints, and Prove's next release moves that line behind a flag. They now readtruncatedfromdiff --json, which Prove 0.1.0 to 0.1.3 all carry. The example also stops if the field is missing, where it used to print nothing. - The
proveextra has no ceiling.pip install "poisson-tasks[prove]"now installspoisson-prove0.1.0 or later, where it stopped before 0.2. Prove runs this route's check, from the latest release of Tasks, before it ships, on a wheel under a local version, and the ceiling would have refused that wheel at 0.2. The route passed on each of 0.1.0, 0.1.1, 0.1.2 and 0.1.3. No Prove 0.2 exists yet to measure.
Nothing about the tasks changed: every task id, both generator hashes and every published card are what 0.3.2 had.
Tests and tooling
- The release wheel gate's driver writes no bytecode. It runs with
python -B -I, because-IignoresPYTHONDONTWRITEBYTECODE, and the driver was leaving__pycache__folders in the release worktree. The wheel is unaffected. - An existing test was edited, in its own commit:
test_pi6_3_the_extra_stops_before_prove_0_2is nowtest_pi6_3_the_extra_has_no_ceiling. The test count does not change, so CONTRIBUTING's sentence on the suite stands.
Poisson Tasks 0.3.2
Fixes
- A mistyped arm id gets the nearest arm back (#31).
generate --armandverify --armused to say only what an arm id looks like, and listed nine of the knobs. They now saydid you mean depth:d3?when the text is near an arm (detph:d3,distractor:r8), and name every knob the built-in tasks take,timeoutsandratelimitsamong them; the sentence isarms.arm_from_id's own, so the library and both commands give the same one. On your own tools the refusal also says which of the plan's arms is nearest. - A depth the generator does not build is refused (#31).
depth:d9,depth:d33anddepth:d0are well-formed ids, and ended as an unexpected failure with exit 3. They are now a usage error, exit 2, in one line that says depth goes from 1 to 8 steps. --seednames a seed of the plan, and one outside it is refused (#31).generate --show --seed 42andverify --seed 42judged any integer without a word that 42 is not one of the plan's seeds. A seed the plan does not hold is now refused, exit 1 as--task 99is, in one line that names the plan's seeds (or how many it has, when more than ten) and says--task Ncounts them by place. The help text anddocs/commands.mdsay "one of the plan's seeds" where they said "a seed you chose".
Nothing about the tasks changed: every task id, both generator hashes and every published card are what 0.3.1 had.
Tests and tooling
- CI runs on a pinned runner image (#36). Every job of both workflows runs on
ubuntu-24.04, where it ran onubuntu-latest, which GitHub is moving to Ubuntu 26. A test holds it for every job of every workflow. - The stop-button tests pass however the suite is started (#37). The three tests that press the stop button failed when the test process was started with
SIGINTignored, as a background job started from a script is. They now setSIGINTto its default for the process they drive. - Three checks from 0.3.0 are held by tests (#38). A second write to the same tool with other arguments the schema accepts is not counted as a duplicate; after the target write, a read with no checkable argument (
list_channelswith no arguments on acme) answers as before the write; and each experimental dial's note ininitequals the reasonarms.EXPERIMENTAL_KNOBSgives for that dial, character for character, dial by dial. Each test was seen to pass on the code and to fail on the change it is written against. - The wording about the published cards is held by a test (#32).
initand the card already say a later release measures the published cards again, as the docs do (0.2.2 changed it from the next release); nothing held the words, and a test now does. - The release wheel gate waits for a prompt to settle before it types (#39). The first run of the
built-in-prettyscene in 0.3.0 lost itsyat the run's Start prompt, because the driver typed into the gap between the prompt being drawn and being read.scripts/release/drive.pynow waits until the prompt is the last line on the screen and nothing has been written for three seconds. It also answers a question once: the first run with only that wait stopped at the same prompt, because the setup loop, finding the session slow to show its next question, answered the last one twice, and the extra Enter reached the Start prompt as an empty answer. - Existing tests were edited, each in its own commit: the two tests that gave
--seeda seed of their own outside the plan now give one the plan holds; the three stop-button tests setSIGINTfor the process they drive; and the test that reads CONTRIBUTING's count sentence accepts 0.3.2's.
Poisson Tasks 0.3.1
Fixes
- The Prove route works with Prove as it is published. Prove went public as the distribution
poisson-prove, whose package ispoisson_prove. Every release up to 0.3.0 imported it under its earlier name, so a Tasks installed beside the published Prove could not use the route.poisson_tasks.adapternow takesEpisodefrompoisson_prove, and importing it without Prove raises aModuleNotFoundErrorthat names the way to install it. Nothing else changes for anyone who does not use the route:import poisson_tasksstill never reaches the adapter. - A
proveextra.pip install "poisson-tasks[prove]"installspoisson-provefrom PyPI, from 0.1.0 up to, and not including, 0.2. The floor was measured, not guessed: the route passed on each of 0.1.0, 0.1.1 and 0.1.2, and the yanked 0.0.1, which holds the name and no code, fails. The ceiling is the next minor release, because Prove is 0.x and its command line can move in one. A plain install does not bring Prove. - The example and its page follow Prove's command line.
examples/prove_sweep/run.py(it wasexamples/proofload_sweep/) reads the run id frompoisson-prove sweep --json, because the text form names a run by its name and no longer prints an id line, and its diff section prints the line that says no episode was cut short by the horizon. The page isdocs/prove.md, and sayspoisson-provewhere it said the earlier name. - The licence is an SPDX expression. The package page showed the whole MIT text in its License field; it now shows
MIT, and the wheel and the sdist carryLICENSEandNOTICEas licence files. TheNOTICEentry for Prove names the published distribution and the licence read from its wheel. - The output blocks in
docs/own-tools.mdanddocs/python.mdshow 0.3.0's generator. They still showed generator version 1, an older task id and adoctorlist of seven dials, so the docs check, run from the wheel, failed on them; 0.3.0's pages shipped with those blocks.
Nothing about the tasks changed: every task id, both generator hashes and every published card are what 0.3.0 had.
Tests
- The route is run end to end from two published packages.
scripts/release/prove_e2e.pyinstalls this tree's wheel with itsproveextra into a fresh environment, so thatpoisson-provecomes from PyPI, then runs the example and the realpoisson-provecommand on the built-in tasks, reads the episodes back, and checks that Tasks still imports, and the adapter is refused with the extra's name, once Prove is uninstalled. Every release runs it, the fast lane of the pull request check runs it, and the test suite runs it at the newest Prove the extra allows and at the lowest. - CI installs Prove from PyPI. The jobs with Prove no longer check out a private repository with a token; they install the
proveextra, and now run on pull requests as well as on main. - Existing tests were edited, each in its own commit, because Prove's command line moved or its name changed: the scripted path's helper that read the run id from a line
sweepno longer prints; the refused-target test's exit code (1, not 2) and its count of run folders (the store root now holds a.poissonmarker folder); the helper that runs the command (python -m poisson_prove); the test of the one import the adapter may make from Prove; the two tests that importEpisodein their bodies; the docs test that names the Prove page; the test that reads 0.3.0's changelog section, which is no longer the first; and the examples test that names the example folder.
Poisson Tasks 0.3.0
Faults, part two: timeouts, rate limits, and a write that landed
- Three new dials, each experimental until a model has been measured on it.
timeoutsis how many calls on the way time out, with nothing landing, and must be made again (levels 0, 1, 2 and 4;timeouts:o2).ratelimitsis how many answerrate_limitedwith a retry-after of one second (0, 1, 2 and 4;ratelimits:r2).landedis whether the write goes through but answers a timeout, so a blind retry is a duplicate (0 and 1;landed:l1); it is offered on your own tools, whose store accepts a second write as yours would. On the built-in ledger a repeated stage call is refused and changes nothing, so a duplicate there has nothing to catch, andlandedis not offered on the built-in tasks. Asked for in replies to the launch. - What the model sees. A timeout says only that whether the call went through is not known; a rate limit says to retry after one second. Tasks has no clock, so the retry-after is a hint the model reads and nothing waits. Each dial adds one sentence to the system prompt at every level, as faults does. On your own tools the answers come in your file's idiom: a 504 or 429 for an OpenAPI document, an error object with
retry_afterfor an OpenAI tools file, anisErrorresult for MCP. - The card shows what the model did after the fault: gave up after a fault (now for each kind), retried after a landed timeout, and on your own tools, duplicate writes. None of them gates a card. A card's tables also appear for every dial it holds a level of: before, a card of the faults, history or shape dial printed no level table.
- A write on your own tools now answers with what it was sent: every argument under its name, into the matching field of the tool's output schema or beside
okandidwithout one. Before, a write answered with a made-up record, so a correct call could read as a wrong one in the replies view. Reads answer as before until a write lands: after it, a read that takes one of the write's arguments shows the written values, so an agent can check whether a timed-out write went through before making it again.landedneeds such a read, andinitanddoctorsay so on a set that has none. - Every task id moved, and no built-in task did. The generator changed to make the three dials, so its hash moved and with it every task id, on the built-in tasks and on your own tools; the generator version is now
"2"("own-2"on your own tools). Every built-in task keeps its prompt, tools, goal and reference solution under its new id, as in 0.1.0. On your own tools a task keeps its request, tools and calls, and its reference answers now carry the echo. A run folder started with 0.2.x cannot be resumed with 0.3.0 (runsays so before anything is spent);card --comparebetween the two says the generators differ; a 0.2.x card still builds and checks. The published cards are unchanged and still not current. - "Experimental" is said with its reason wherever a dial is named: in the card's "what it does not claim", in
doctor's rows on your own tools, ininit's choices, in the export's arm help, and asexperimentalincard --jsonanddoctor --json. The decoy-tools dial keeps its reason (it did not make tasks harder for either calibrated model); the three new dials say no model has been measured on them yet. - Reaching them.
generate --armandexport --to verifierstake every new arm;[measure] preset = "custom"withknob = "timeouts"or"ratelimits"runs and cards them on the built-in tasks, with the control pairtimeouts:o0+len(o4)undercontrols = true, and on your own toolslandedtoo. No recipe changed; a later release measures the three on current models. - Library names added:
arms.EXPERIMENTAL_KNOBS,surface.fault_kind,simulator.State.writes,calibration.experimental_knobs,generator.LANDED_NEEDS_A_CHECK_READ, and the record fieldsretried_after_landedandduplicate_write, each left out of a written record where it does not apply.
Tests
- The three dials, the echo, the reflection and the card are held by tests written before the build (
tests/test_f2_*.py): the arms and their refusals, the lines and the plan at every card seed on both splits, the counting episodes and the live replay, the own-tools answers in each format's idiom, the figures, the experimental surfaces, and the docs. The generator's hash is recorded again at its new value, with a kept file of the timeouts and ratelimits digests at the card seeds, and the own-tools digests are recorded again too.
Poisson Tasks 0.2.2
Fixes
initrefuses a key that would go over plain http to another machine, in one line. With a key variable named and a--base-urlthat is plainhttpand not loopback,initnow exits 1 with anerror:line and anext:line that is the command as typed plus--allow-plain-http-key, writes nothing and opens no connection. Before, it stopped with exit 3, as an unexpected failure, whenever it checked the key: with the key set, and with the key unset, which it checks even under--no-check-key. The guided setup now says the refusal at its key question and refuses after the last one.- A card no longer promises that the next release measures the published cards again. The line on a card that cannot be read beside the published cards, because they were measured on an earlier build of the generator, and
init's note on the published card now say a later release measures them again. Neither 0.2.0 nor 0.2.1 did.
Tests
init's plain-http refusal is held: under--yeswith and without--no-check-key, with the key set, and after the guided setup. Each case checks the exit code, theerror:andnext:lines, that nothing is written and that no connection is opened, beside loopback andhttpsendpoints that write the file as before.
Poisson Tasks 0.2.1
Fixes
card --checkandcard --publishedsay a card is current from what its run recorded.currentnow agrees with the card's own stale line: it reads the generator the run recorded when the run recorded one, on the built-in tasks and on your own tools. Before, it read the generator hash in the card's block against the built-in generator, so a card built by a newer install than its run could saycurrentwhile its stale line said it was stale, and every card on your own tools said not current. The published cards record no run and read as before.- The coloured look keeps a
made withrow inside the terminal. A line of a row made of items joined by·now leaves room for the·that ends it. Before, a line could run two columns past the terminal, which then broke it inside an item.
Tests
- What 0.2.0 built and no test held is now held: a resume on your own tools refused when the run was started with another own-tools generator or on another tool contract; a run started before 0.2.0 and resumed writing one
resumedline, its card saying it was started before 0.2.0 and resumed with this version;runrefusing such a resume before the plan and its cost question;card --checkreadingcurrentfrom the run's record; themade withrow breaking only between its items; and thebeside the published cardsrow leaving the label column as wide as the other rows need, in both looks.
Poisson Tasks 0.2.0
Every card says what produced it
- A card names what made its numbers, so a change between two cards can be traced to a changed tool contract or a changed generator rather than to the model. Asked for on the launch post.
- The run records it before its first request, in a new file in the run folder,
produced_by.jsonl: the package and verifiers versions, the generator's hash (and on your own tools the own-tools generator's hash), the hash of your tool file's contract, the plan's hash and its seeds. A resume adds a line.plan.jsonis unchanged, so every run folder is found and resumed as before. - The tool file's contract hash is the SHA-256 of the file as canonical JSON: keys sorted, no spaces, UTF-8. Whitespace, key order, and JSON against YAML for the same document do not change it; any change to a tool, a description or a value does. It is computed from the file you gave, with no domain prefix, so you can recompute it yourself;
docs/own-tools.mdshows how. - The card shows it: a
made withrow in the result, aWhat produced itsection incard.mdwith the full hashes, andproduced_byincard --jsonand in the card'sartifact.json.card --checkrebuilds a card that carries it byte for byte. card --comparetakes another run's folder. When the two runs measured the same tasks the same way it sets the cards side by side incard/compare.md, with one note when only the package or verifiers version differs. When a generator or the tool contract changed it says which, with both hashes, and sets no figure beside another, since the tasks differ: the line says the shift between them is not the model's alone.--compare publishedis unchanged.- A resume is refused when the run was started with another generator or tool contract than this install's, in one line, before anything is spent: its tasks are not the ones this install makes.
export --to verifierswritesproduced_by.jsonbeside its configs; the configs are unchanged.- Runs made before 0.2.0 recorded none of this, and their cards say
not recorded; they are read, carded and resumed as before. The two published cards are unchanged, byte for byte, and say the same. - No task moved. The generator is unchanged, so every seed gives the task it gave in 0.1.0, with the same id.
- A card written by 0.2.0 is checked with 0.2.0 or later;
card --checkin 0.1.0 does not know its new field.