Hand off the detector-design thread: the surrogate IS the design - #523
Merged
Conversation
A blue-sky session proposed a coordinated-event detector that needs no labels, had it stopped by its own eleven-role review, and spent the rest of the day stocking the literature that would have prevented the error. Nothing is half-done -- #519, #520 and #521 all landed. What is open is a decision. The handoff distils what survives into the next step, rather than restating the run record. Three things it carries that the record does not: THE SCREEN, not a choice. Tony: "why pick one replacement? we have compute to test all of them." Right, and this project already has the pattern -- the tube 2x2 was built to be measured rather than argued about. Eight surrogate candidates in a grid, three tiers cheapest first, with the KILLED uniform dither kept in as the known-bad positive control: a screen that fails to flag it is broken. The decisive tier needs no model at all, because the leak that killed the proposal was found by counting. WHY THE COST IS NOT COMPUTE. Pattern jitter is a dynamic program, joint-ISI dithering needs a per-cell 2-D histogram, operational-time dithering needs a rate estimate that is its own parameter choice, and each needs its own support-violation tests because a subtly wrong surrogate produces a leak that looks like a finding. Generation cost also sits in the training inner loop -- fresh negatives every epoch -- so it may bind at training time while being free in the screen. THE TRAPS, most of them this session's own. Three fabricated PMCIDs written immediately after the review that exists to catch fabricated citations. Seven claims about bugarach contradicted by bugarach -- and on the nulls the glossary was right and the reading was wrong, which is what INDEX.md is for. A document written after a murderboard has not been murderboarded. Corrections deliberately NOT made, each left for a change with its own review: the Gregers Hansen canonical surname across three docs; whether detector_history.md section 4 is right to head its 1980 row "origin", which nobody can answer until someone reads the 1973 paper; and the export contract's silence on field steps. Root HANDOFF.md gets an ADDITIVE pointer block only. The loop/MAHICE thread and its #466 are untouched and explicitly not superseded -- two live threads now, and the note says to delete only the one that lands. That third option is the one whose absence cost four days in August. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
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
2 participants
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.
A blue-sky session proposed a coordinated-event detector that needs no labels, had it stopped by its own eleven-role review, and spent the rest of the day stocking the literature that would have prevented the error.
Nothing is half-done — #519, #520 and #521 all landed. What is open is a decision, not a task.
What the handoff carries that the run record doesn't
The screen, not a choice. Tony: "why pick one replacement? we have compute to test all of them." Right — and this project already has the pattern; the tube 2×2 was built to be measured rather than argued about. Eight surrogate candidates as a grid, three tiers cheapest first, with the killed uniform dither kept in as the known-bad positive control: a screen that fails to flag it is broken. The decisive tier needs no model at all, because the leak that killed the proposal was found by counting.
Why the cost is not compute. Pattern jitter is a dynamic program; joint-ISI dithering needs a per-cell 2-D histogram; operational-time dithering needs a rate estimate that is its own parameter choice. Each needs its own support-violation tests, because a subtly wrong surrogate produces a leak that looks like a finding. And generation cost sits in the training inner loop — fresh negatives every epoch — so it may bind at training time while being free in the screen.
The traps, most of them this session's own. Three fabricated PMCIDs written immediately after the review that exists to catch fabricated citations. Seven claims about bugarach contradicted by bugarach — and on the nulls, the glossary was right and the reading was wrong, which is what
INDEX.mdis for.Corrections deliberately not made
Each left for a change with its own review: the Gregers Hansen canonical surname across three docs; whether
detector_history.md§4 is right to head its 1980 row origin (unanswerable until someone reads the 1973 paper); and the export contract's silence on field steps.Root HANDOFF.md
Additive pointer block only. The loop/MAHICE thread and its #466 are untouched and explicitly not superseded — two live threads now, with a note to delete only the one that lands. That third option is the one whose absence cost four days in August.
🤖 Generated with Claude Code