Compose vision with gates, and cover what a workcell needs to keep running - #5
Open
di-omics wants to merge 2 commits into
Open
Compose vision with gates, and cover what a workcell needs to keep running#5di-omics wants to merge 2 commits into
di-omics wants to merge 2 commits into
Conversation
…nning Four layers for the part of an autonomous lab that starts once more than one instrument is involved: whether anything would catch a failure, whether the instruments are still entitled to be trusted, whether an expert has taught the operation, and whether a control loop can close. coverage is the composition and the reason the other three are worth having. vision knows what a camera can and can never see, qc knows which gates exist and whether they fire, and recovery knows which failures destroy material. Nothing joined them, so nothing could answer the question a lab actually asks: for every failure that destroys material, is there anything at all that would catch it? On the reference protocol the answer is 14 failure modes, none covered, 9 of them destructive. The useful part is that they do not split into one problem. Six are visible and their checks are declared and undeployable -- a CV project with a known payoff, and bead_pellet_aspirated is among them. One is invisible but already sits behind a declared gate that cannot fire because the reader is broken, so no camera is involved and a repair closes it. Seven are invisible with no assay reading the quantity they change, and no capability upgrade moves them at any price. Confusing those three is how a CV budget gets spent against a problem no model solves. sota_lift therefore returns both lists, because a report showing only what a capability lifts is a purchase justification. Two things refused as coverage: a vision check that exists but whose requirements are unmet, and a gate that exists but cannot be evaluated. Counting either reports a lab as covered while material is quietly destroyed. durability tracks entitlement rather than predicting failure. It computes whether a planned campaign crosses a service or calibration boundary mid-run, because a plate that started before an expiry and finished after it has an ambiguous provenance and nothing downstream repairs that. No MTBF and no remaining-life estimate: that needs population reliability data this package does not have, and the invented figure becomes a specification. teaching models the transfer. A demonstration is data, not authority -- a scientist doing something twice is two observations with a spread, and the spread is the information. An envelope refuses a tolerance from one demonstration and a machine with one good run reports as indistinguishable from unmeasured. The operations with no envelope are the real backlog of an automation programme and demonstration_queue ranks them the way unlocks() ranks decoding work. feedback asks whether a loop closes, with latency as the load-bearing check: a sensor downstream of its own actuator is a post-mortem wearing the costume of a control loop. No gain, no settling time -- that needs a plant model and measured response data. What it does report is how many plates are already committed between sensor and actuator. Each module was built, adversarially audited on three lenses, and fixed. Includes a fix for attainment() naming the untaught operation in its own refusal, which is the common case and was reading "no expert has demonstrated 'unknown'". 156 new tests, 284 total.
Three audit lenses ran against each of the four modules -- one hunting defaults that read as a PASS when they are really a SKIP, one checking that no fact is recomputed where a sibling module already owns it, and one actively trying to extract a flattering answer. Between them they found the same class of defect this package exists to refuse, committed by the modules written to refuse it. The pattern repeats across all four. A gate reading a measurement nobody declared counted as coverage. A demand was met by a measurement set that says nothing. A duplicate gate row silenced an unsatisfiable one because the lookup returned the first match, which also made the report depend on row order. A workcell declaring no instrument at all reported as a clean lab. A NaN accrual was read as a measurement. A negative accrual minted budget that was never earned. A campaign landing exactly on a service boundary was reported as not crossing it. A coverage report holding nothing destructive reported as complete. Two are worth naming separately because they are cross-module rather than local. A recovery report built against a different gate report was accepted and silently mixed two workcells' worth of evidence; it is now refused. And a gate path this module had already refused was being republished as a residual detection path, so a failure this layer called uncovered appeared one line later as watched by the thing that could not watch it. Every fix carries a regression test that fails before it and passes after. No existing test was weakened to make anything pass. 63 new tests, 347 total. ruff clean.
di-omics
force-pushed
the
feat/coverage-durability-teaching-feedback
branch
from
July 29, 2026 17:39
d531723 to
3330cda
Compare
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Four layers for the part of an autonomous lab that begins once more than one instrument is involved.
coverage.py-- the composition, and the reason for the restvisionknows what a camera can and can never see.qcknows which gates exist and whether they fire.recoveryknows which failures destroy material. Nothing joined them, so nothing could answer the question a lab actually asks: for every failure that destroys material, is there anything at all that would catch it?On the reference protocol: 14 failure modes, 0 covered, 9 destructive. They do not split into one problem:
bead_pellet_aspiratedis here -- the failurerecoveryalready calls the worst silent oneenzyme_activity_lostsits behind an unsatisfiable gate because the reader is broken. Repair the instrumentmandatory_gates(). A better model buys nothing. They need an assay the protocol does not havesota_lift()returns both lists, because a report showing only what a capability lifts is a purchase justification.The other three
durabilitytracks entitlement, not prediction: whether a planned campaign crosses a service or calibration boundary mid-run. No MTBF, because that needs population data nobody has.teachingholds that a demonstration is data, not authority. A scientist doing something twice is two observations with a spread. An envelope refuses a tolerance from n=1, and a machine with one good run reports as indistinguishable from unmeasured.feedbackrefuses any loop whose sensor sits downstream of its actuator: a measurement taken after the material is consumed is a post-mortem wearing the costume of a control loop.The audit earned its keep
Three adversarial lenses found the class of defect these modules exist to refuse, committed by them: a gate reading an undeclared measurement counting as coverage; a duplicate gate row silencing an unsatisfiable one because the lookup returned the first match; a workcell declaring no instruments reporting as clean; a NaN accrual read as a measurement; a campaign landing exactly on a boundary reported as not crossing it. All closed with regression tests.
156 new tests. Recreated after the PR history was reset; the branch is unchanged.