ThesisAnalysisOS

Evidence

Every claim, read back

This is docs/CLAIMS.md, with one addition. When the page was built, each row was checked against the code: a tested row names code and a test that exist, and a measured row quotes a number its recorded run contains and that run still stands. 42 of 42 hold. Had one not, the build would have stopped.

Check it yourself in the repository with python3 evals/claims.py.

What a record stood on

ClaimCodeCheckStanding
A recorded result is answered holds, fails, stale or unknown, with a reason; unknown is never turned into another answertaos_core/basis.pytests/test_standing.pyholdstested
A passed check holds for the bytes it tested and stops holding when one changes, however recent it isgates.standingtest_a_pass_does_not_cover_a_file_edited_afterwardsholdstested
The answer names the file that changedbasis.comparetest_closed_work_stands_until_a_file_changes_and_the_reason_names_the_fileholdstested
No route moves a task to review or done unless every check holds, or a waiver with a reason is recorded and shownTaskStore.transitiontest_done_cannot_be_reached_by_a_route_that_skips_the_check, test_a_waived_closure_is_unverified_until_the_check_runsholdstested
A result is shared by every task whose files are the same bytes, and is executed oncegates._key, .taos/gate_results.jsontests/test_standing.py::TestSharedResultholdstested
A change to a file no domain claims is unknown; a check naming an undeclared domain is unknownbasis.compare, gates.standingtests/test_standing.py::TestDomainsholdstested
A declared domain holds what its patterns match on disk, so an input git ignores is part of what a check stood onbasis.on_disk, basis.snapshottest_an_input_git_ignores_is_still_part_of_what_a_check_stood_onholdstested
A check marked isolated can read only the files its result is filed under, and what it writes is discardedgates._isolatedtest_a_check_cannot_pass_on_a_file_it_did_not_declare, test_an_isolated_check_sees_only_what_it_declared_and_leaves_nothing_behindholdstested
Only a person accepts a result, and the acceptance is of one versionTaskStore.accepttest_acceptance_is_of_one_version_and_does_not_carry_overholdstested
A work product is current while its declared inputs are the same bytes; with none declared it is unknownadd_evidence, standing._artifactstest_a_work_product_is_current_while_its_declared_inputs_are_the_same_bytesholdstested
A new or lapsed instruction makes work closed under the old ones stalerules.binding, rules.changes_sincetest_a_new_instruction_makes_the_closure_stale_and_is_quotedholdstested

Instructions, lessons and the other records

ClaimCodeCheckStanding
The instruction that governed a scope on a past day is still answerable after it was replacedrules.governingtest_a_later_replacement_does_not_erase_the_earlier_answerholdstested
A replacement governs from its own date, inside its own scope, until it lapsesrules.displaced_bythe four other tests in TestInstructionsKeepTheirTimeholdstested
What governs a task and what was learned reach the agent at start, in the original wordslifecycle.start, capsule_texttests/test_standing.py::TestDeliveryholdstested
A reminder shown is an exposure; only a judged outcome is a repeatlessons.at, lessons.outcome, lessons.repeatstest_a_reminder_shown_for_correct_work_is_not_a_repeated_mistakeholdstested
A setup is credited only when it was observed to runcontrol.rebuildtest_an_unobserved_profile_earns_no_creditholdstested
A lapsed lease is overClaimStore.heartbeat, acquiretests/test_counterexamples.py::TestALeaseEndsholdstested
A refused start takes nothinglifecycle.starttest_a_refused_start_takes_nothingholdstested
A label is read strictly and a missing one is counted apartreplay.labeltest_the_text_false_is_not_read_as_true, test_an_unlabelled_case_is_not_counted_as_harmlessholdstested
A decision-bank run keeps the answers it scored and their hash, so the score can be recomputedbank.scorethe bank test in tests/test_science.pyholdstested
A change is not confirmed by a number with no source, and a prediction is written oncechanges.score, anchors.predicttest_a_change_is_not_confirmed_by_a_number_from_nowhere, test_a_prediction_cannot_be_rewritten_once_madeholdstested

The repository's own numbers

ClaimCodeCheckStanding
A number quoted here is stale once the code it was measured on changes, and the answer names the fileevals/measured.pytest_a_recorded_number_goes_stale_when_the_code_it_measured_changesholdstested
An isolated check sees a fixed environment, not the caller's_fixed_environmenttest_an_isolated_check_does_not_see_the_environmentholdstested
Each isolated result records what was closed, found by trying; where the network and the rest of the workspace can be closed they are, and a strict check does not run where they cannotclosestest_the_record_says_what_was_closed_and_it_washoldstested
The setup answers can say which files each check reads_config_fromtest_setup_can_say_what_a_check_readsholdstested
No row of this file names code or a test that is not there, or a number its recorded run does not containevals/claims.pytest_no_claim_names_something_that_is_not_there_or_misquotes_a_recorded_runholdstested
The demo runs from a clean copy and prints each idea with real outputtaos_core/demo.pytest_it_runs_from_a_clean_copy_and_shows_each_idea_with_real_outputholdstested

Measurements

ClaimNumberReproduceStanding
Earlier releases had the counterexample cases: every one at the first, five at the second22 of 22 fail at 071a348; 5 of 22 fail at 89f36ce; 22 of 22 pass herepython3 evals/at_revision.pyholdsmeasured
The tests notice when a protection is removed. The first run, with 16 faults, caught 15; the missing test was added20 of 20 seeded faults caughtpython3 evals/mutations.pyholdsmeasured
Reuse by basis relied on no stale result and executed what an oracle would60 executions, 0 wrong, in 360 results relied onpython3 evals/reuse_frontier.pyholdsmeasured, one workload
Age limits trade reruns against stale reuse180, 90, 48 executions with 30, 83, 161 wrongsameholdsmeasured, one workload
In the open, under-declared inputs are reused wrongly150 wrong in 360 when every check hid one foldersameholdsmeasured, one workload
Isolated, a check that reads what it did not declare never passes, and correct declarations cost no extra executions. Predicted before the code was written0 of 366 runs passed when every check hid one folder and ran isolated; 60 executions, 0 wrong with everything declaredsameholdsmeasured, one workload
Asking costs time that grows with the tree15.4, 40.6, 305.4 ms for 48, 480, 4,800 filessameholdsmeasured, one laptop
Isolating a check costs milliseconds. Measured in pairs that differ only in the isolation, order alternating14.0 ms added, the median of 40 paired runs (95% interval 13.6 to 14.7 ms)sameholdsmeasured, one laptop
The suite passes on Python 3.9 and 3.12, Linux and macOS151 tests./taos selftest, and CIholdstested

Not claimed

  • That research done with this system comes out more correct. Nothing here measures research outcomes.
  • That the cost laws of physical transport networks apply to agent work. docs/METHOD.md says what is shared with that work (a two-term ledger and the question it asks) and what is not (the exponents).
  • That the controller in policies/CONTROL.md allocates effort optimally. It is a bounded one-step rule, and it credits only what it observed.
  • That an open check's declaration of what it reads is verified. It is trusted, unless the check is marked isolated.
  • That an isolated check is a correct check. Isolation makes its result depend only on the files it is filed under and a fixed environment. The network and the rest of the workspace are closed on macOS only, and a file outside the workspace can be read by its absolute path anywhere.
  • That anything outside the workspace is seen: the network, the clock, the environment, another machine.

The reuse experiment in full

60 edits, 6 checks over 12 folders, seed 7. After each edit every check is relied on, 360 results in all. A wrong reuse is a result relied on after a file it reads changed. Python 3.12.10 on macOS-27.2.1-arm64-arm-64bit.

Policy, as recordedExecutions Wrong reuses
rerun everything3600
rerun when older than 2 consultations18030
rerun when older than 4 consultations9083
rerun when older than 8 consultations48161
TAOS, 1 domain(s) declared3600
TAOS, 2 domain(s) declared1800
TAOS, 3 domain(s) declared1200
TAOS, 4 domain(s) declared1200
TAOS, 6 domain(s) declared600
TAOS, 12 domain(s) declared600
oracle (knows what every check reads)600
TAOS, three tasks asking600
TAOS, each check hides one folder it reads29150
TAOS, checks isolated, declared correctly600
TAOS, checks isolated, each hides one folder3660

Predictions, written before the runs

  • held P1. with what each check reads declared correctly, no stale result is ever reused, at any granularity
  • held P2. the number of executions equals the count derived from the edit trace exactly: nothing extra, nothing missed
  • held P3. three tasks asking the same questions execute no more checks than one task does
  • held P4. finer declarations never cost more executions, and stop helping once they match what the checks read
  • held P5. a check that reads a folder it did not declare does reuse a stale result: the declaration is the limit
  • held P6. no age window reaches both the oracle's execution count and zero stale reuse
  • held P7. isolated, a check that reads a folder it did not declare never passes, and correct declarations cost nothing extra

The record itself · The script that wrote it