Two handover items existed only in a machine-local board — and one of them was wrong

docs/detector_history.md

@@ -3,6 +3,76 @@
 *Written 2026-08-22 from two interface2 reports, the bugarach tree, and the
 darkroom literature shelf.*
 
+> **Revised 2026-08-30: two of the three "his" detectors were built with AI
+> sessions, and the independence claim is withdrawn — replaced by a timeline that
+> does not depend on anyone's memory.** This corrects the block immediately below,
+> which is five days older and says the wrong thing in one place and an
+> unestablishable thing in another.
+>
+> **What is wrong.** Tony, 2026-08-30: *"we built the first version of loco and
+> coact detect in matlab together."* Together means with AI sessions in this
+> estate. So *"`rate_detect`, `coact_detect`, `loco_detect` **are his**"* overstates
+> for at least two of the three. interface2's own record agrees, and the dates are
+> checkable:
+>
+> | | first committed | AI-assisted commit |
+> |---|---|---|
+> | SpikyDetect 1–3 (synchrony) | 2026-04-23 → 04-28 | **no** |
+> | `rate_detect` (`SpikyDetect4` → `RateDetect`) | 2026-05-13 | **yes** |
+> | `CoactDetect` (renamed from *"local coinc"*) | 2026-07-14 | **yes** |
+> | `detect_loco` | 2026-07-14 | **yes** |
+>
+> The first Claude-assisted commit anywhere in interface2 is **2026-05-06**. The
+> synchrony lineage predates it; **everything rate- and coordination-shaped comes
+> after.** ⚠ Commit dates are a **lower bound** on when work happened, and Tony
+> notes his repo hygiene was poor for years, so uncommitted earlier work cannot be
+> ruled out. What the dates do establish is which commits carry a co-author trailer,
+> and for `rate`, `coact` and `loco` they all do.
+>
+> **What is unestablishable.** The block below rests §4's *"convergence, not a
+> derivation"* on *"I was totally unaware when I designed them."* For a human that
+> is real evidence — a person's knowledge is roughly what they have read. **It is
+> not available as evidence when a model is a co-builder**, because a model has no
+> introspective access to whether its output derives from CFAR-shaped training data,
+> and neither collaborator can inspect that from the outside. Tony, 2026-08-30, on
+> the lit search that found CFAR: *"i think it surpised both of us."* Genuine, and
+> consistent with **both** independent convergence and unrecognised reconstruction.
+> **Surprise cannot separate them.** So the claim is withdrawn as unsupportable, not
+> as false — nothing here says it is wrong, only that nothing available can show it
+> is right.
+>
+> **What replaces it, and it is stronger.** A timeline, in git, in two repositories,
+> depending on nobody's recollection:
+>
+> > The detectors were built **2026-04 to 2026-07**. The CFAR literature was found
+> > on **2026-08-22** — `ff1db4d`, *"Three of the six detectors judge a moment using
+> > a window that contains it"* — and the resemblance was recognised then, five weeks
+> > to four months after the fact, in a search run for another purpose. `rate`,
+> > `coact` and `loco` were built in AI-assisted sessions.
+>
+> That **rules out deliberate derivation**, which is the only charge worth
+> pre-empting. It does not assert independence from the literature, and should not:
+> with a model in the loop that is not a claim this project can support, and §7
+> already concedes four literatures were never searched.
+>
+> **This costs nothing that was being defended.** Tony closed priority on 2026-08-24,
+> in the block further down — *"most researchers would be kind of thrilled with the
+> link … it's a tool and it's useful"* — and the ruling there was *acknowledge the
+> origins and take the engineering*. The independence claim was never load-bearing;
+> it is simply the one sentence here that would not survive a reader who asked how it
+> could be known. **§5's list is untouched, and it is what this document is for.**
+>
+> Also unchanged: **port fidelity to 1e-9**, which is about code and not about ideas;
+> the SCE root (Cossart 2003) and the Kreuz position, both settled in the blocks
+> below; and `locust` as the one true port.
+>
+> Consequence outside this file: the README and `CITATION.cff` are wrong in the
+> *other* direction — they say "six detector ports" when one is a port — so the repo
+> now has an error each way. Filed as
+> [the repo underclaims its own authorship](todo/2026-08-30-the-repo-underclaims-its-own-authorship.md),
+> **which must be re-read against this revision before anyone acts on it**: it was
+> written before this correction and its title overstates.
+
 > **Revised 2026-08-29: the author said where each one came from, and interface2
 > is frozen on the topic.** Every lineage row in this document was assembled from
 > interface2's reports and this tree. Tony supplied the missing half — who wrote

docs/todo/2026-08-30-the-rail-still-costs-97px.md

@@ -0,0 +1,46 @@
+---
+status: open
+filed: 2026-08-30
+---
+
+# The rail's last ~97px cannot be trimmed — it has to be argued
+
+> **Not murderboarded** — a work item. Filed by a session that did **not** do the
+> measurement, from `bugarach-63`'s handover at its session end on 2026-08-30.
+
+## What already landed
+
+**PR #411 (`0188362`)** took the viewer's canvas top from **514px → 351px** at
+1440×900: the header went, its privacy promise moved into the nav, `#demoNote`,
+`#meta`, `#bar` and `#wins` merged into one row behind a four-state provenance chip,
+the rail was **tightened, not redrawn**, and the diagnostic page now leads with its
+figure with the legend in a closed `<details>`.
+
+## What is left, and why it stopped
+
+About **97px** of rail remain, and the session that measured it stopped rather than
+take them, because **the remaining trim is not a trim.** Getting them requires
+collapsing two grid rows, and the rows are not decoration — **collapsing them changes
+what the diagram asserts.**
+
+That makes it a claim about the pipeline, not a layout tweak, and it is the reason
+this is a todo and not a follow-up commit. A session optimising for pixels would take
+them without noticing the assertion moved.
+
+## What is actually undecided
+
+- **What do the two rows currently assert**, and is that assertion load-bearing for a
+  reader who has not seen the pipeline before?
+- Is 97px worth a weaker diagram? The screenshot that started this had the raster at
+  **43% of its own page**; after #411 it is not, so the pressure that justified
+  aggressive trimming has largely been spent.
+
+⚠ **Do not treat this as a continuation of #411.** #411 was cosmetic by construction
+and said so; this one is not, and inheriting its "cosmetic" framing is how the
+assertion would get changed silently.
+
+## Not authorized
+
+Handed over, not started. Tony pulled that session up for scope on 2026-08-30 —
+*"address this critique does not mean run wild"* — so the existence of this file is a
+record, not a go-ahead.

docs/todo/2026-08-30-who-wrote-the-detectors-the-repo-is-wrong-both-ways.md

@@ -0,0 +1,101 @@
+---
+status: open
+filed: 2026-08-30
+---
+
+# Who wrote the detectors — the repo is wrong in both directions, and the fix is smaller than it looks
+
+> **Not murderboarded** — a work item. Verified against `main` at `0188362` and
+> against interface2 at `c711e737` on 2026-08-30; file:line given for each claim.
+> **If any of this reaches an outside reader as prose, murderboard that artifact.**
+
+> ⚠ **REWRITTEN 2026-08-30, hours after filing.** This file was first called
+> *"the repo underclaims its own authorship"* and that title was itself an
+> over-claim. Tony then said *"we built the first version of loco and coact detect
+> in matlab together"*, which makes the repo wrong in the **other** direction too.
+> **Read [`docs/detector_history.md`](../detector_history.md)'s 2026-08-30 revision
+> first** — it is the authority and it moved. The original framing survives only for
+> the README/`CITATION.cff` half.
+
+## The two errors, in opposite directions
+
+**1. The front door under-claims.** It says the repo is six ports of other people's
+MATLAB. It is one.
+
+**2. The lineage doc over-claimed.** It said three detectors were Tony's alone. Two
+of the three were built with AI sessions, and its independence-from-CFAR claim rested
+on a mental state that is not available when a model is a co-builder. **Corrected
+2026-08-30** — see that file's own revision block; this item does not restate it.
+
+**The repair is not "claim more."** It is to say accurately what was written where,
+and by whom, in both directions.
+
+## Error 1 — what the front door says
+
+| where | text | the problem |
+|---|---|---|
+| `README.md:7` | *"bugarach **lifts six** coordinated-event detectors out of MATLAB"* | **accurate about the mechanism** — six Python implementations were made from six MATLAB originals. It omits **whose** MATLAB, and that omission is the whole defect |
+| `README.md:79` | *"**Six detector ports**"* — first row of **What is built** | reads as six ports of *other people's* methods. Its own body then links to `detector_history.md` *"for which is which"* — so the summary is wrong and defers the correction to a file nobody opens. **The summary is the part a reader keeps** |
+| `README.md:73–74` | interface2 *"is named here as the source of a port and never linked"* | true, and **says nothing about who wrote the originals** — the gap a stranger fills the ordinary way |
+| `CITATION.cff:15` | *"Browser-based viewer and **Python ports of six** neural-coordination detectors (… locust …)"* | **the machine-readable one.** GitHub renders it in *Cite this repository*; citation managers quote it verbatim, where no next paragraph sits beside it. Also names `locust`, which the public build withholds |
+
+⚠ **The word "port" is doing two jobs and the README conflates them.** (a) MATLAB →
+Python, a code path, true of nearly all of them. (b) Someone else's method → this lab,
+an intellectual path, true of **one**. A reader sees (b) and the repo means (a).
+
+**The first version of this file got that wrong**, saying *"five of six were not
+lifted."* They were lifted — mostly out of Tony's own MATLAB. Omission, not error.
+
+## What the record actually supports, with its caveats attached
+
+From [`detector_history.md`](../detector_history.md), **including** its 2026-08-30
+revision. These caveats are load-bearing and belong wherever the correction is
+written, not in a footnote:
+
+| detector | origin | caveat that must travel with it |
+|---|---|---|
+| `rate+context` | designed here; commit body states it in Tony's terms — *"excess = primary rate (1s) − context rate (60s)"* | **AI-assisted commit** (2026-05-13). 08-24 also calls it cell-averaging CFAR |
+| `CoactDetect` | built here | **built with AI sessions** — Tony, 2026-08-30 |
+| `LoCo` | built here | **built with AI sessions.** `maxlt` is GO-CFAR (Hansen 1973); percentile-of-pool is kin to OS-CFAR |
+| `SPIKE-synch` | Tony's detector on Kreuz's measure | **"Not novel either"** — Kreuz's own lab published the same two-knob detector (Kreuz et al. 2022) |
+| `binned SCE` | **not a port** | published root is **Cossart 2003, from Yuste's lab**, predating CICADA |
+| `locust` | **the** port, modified at port | the one case where "port" means (b) |
+
+**So the honest correction is narrow: say who wrote the MATLAB originals.** It is
+**not** "these are novel", and for two of the six the record says the opposite.
+
+## Two traps for whoever writes it
+
+1. **README and the site describe different detector sets, and both are right.** The
+   README documents the repo — six. The public site **withholds `locust`** — five. The
+   attribution paragraph is **written twice**, not copy-pasted.
+2. **It meets the count problem in the `<h1>`.** `tools/build_site.py:171` reads
+   *"Six coordinated-event detectors ported from MATLAB"* — **wrong about the count and
+   wrong about "ported", in one sentence.** That is
+   [the site types what a token could substitute](2026-08-30-the-site-types-what-a-token-could-substitute.md)
+   item 3 meeting this item at the same line. Both sessions that looked reached that
+   independently; doing them together beats sequencing.
+
+## Scope
+
+**A wording fix backed by an existing document**, now that the document is correct.
+It does not re-derive who wrote what, does not weaken the **port-fidelity to 1e-9**
+claim (about code, not ideas — `README.md:79`'s body is careful and should survive
+intact), and claims priority over nobody. Tony closed priority on 2026-08-24:
+*"most researchers would be kind of thrilled with the link … it's a tool and it's
+useful."*
+
+## Provenance
+
+Handed over at `bugarach-63`'s session end on 2026-08-30, which found it and reserved
+the branch name `who-wrote-these` **without filing anything** — it existed only in
+cross-session messages and a name that was never cut. That session did not start the
+work; Tony pulled it up for scope: *"address this critique does not mean run wild."*
+
+**The existence of this file is not authorization.** The paths are free; the go-ahead
+is Tony's.
+
+The outside critique that started it: a stranger read the four public repos
+adversarially for a short-course page, concluded bugarach was *"a Python re-typing of
+somebody else's MATLAB algorithms"*, and **published that** before being corrected.
+That critique was right that the front door misleads. It did not know about error 2.