The state of the work

Chapter 6 promised future work. This is that work, being built.

If you have not read Start from zero yet, read that first — it explains the problem this is all about. This page is the status of the software: what exists, what it fixed, and what comes next.

DocumentFinal · submitted

Phases done8 of 9

Tests433 passing

API keys used3 · phase 4

Where to go

Start from zero

start

The whole problem explained assuming no background at all: what the benchmark is, why a correct fix can score zero, and what the experiment showed. Twelve minutes, no code.

What changed since you submitted

viva

Four findings that post-date the document — including one correction to how the central mechanism is described, and a check confirming your classification was right all along.

The result, as a reminder

Explained in full on the entry page; repeated here as a picture, because every number below refers back to it.

The fourth format-class task, astropy__astropy-14182, was excluded: the model refused to generate for it at all, so there was never a patch to evaluate. That leaves three evaluable, and all three recovered.

So why build anything?

Because the original study had three problems that no amount of writing could fix.

Problem	What the build does
A failed attempt left no record at all — it printed “Not applied” and wrote nothing, so 8 of 24 cells simply vanished	Every cell is enumerated before anything runs, and must end in a recorded outcome — one of eight, never silence
Conditions lived in 30 separate scripts with no entry point, so nobody could state what had been run	One config file, one command, a checksum over the config that changes if any setting does
The headline numbers were typed by hand into tables and figures, so nothing checked them	They regenerate from the raw logs, and the regenerated table is compared byte-for-byte with yours

The whole thing runs with no API key, no network and no Docker.

That was a deliberate constraint, not a limitation we worked around. It means anyone can verify the results on any laptop — which is the difference between a claim and a result.

The nine phases

Each phase depends on the one before it, so the order is real rather than administrative. A page is written when a phase finishes, so nothing here describes work that does not exist yet.

Phase	What it built	State
0 · Provenance	Rescue the original data, freeze the 300-task benchmark	done
1 · Foundations	The shared vocabulary, the config schema, the pre-flight check	done
2 · Edit layer	The part your finding actually lives in — 138 tests	done
3 · Replay	Chapter 5 regenerated from raw logs, offline	done
4 · Providers	The first live model call, with a token budget that refuses to overspend	done
5 · Strategies	Your three recovery conditions, prompts byte-verified against the originals	done
6 · Sweep	1,308 cells, cost-checked before spending, with no missing cells	done
7 · Dashboard	Six views over the results; states the finding in a sentence derived from the data	done
7 · Dashboard	The visual layer	next
8 · Extensions	The L3 and L4 rungs your section 8 proposes	planned

Nothing here changes your document

Self.docx is final and is not edited again. Where the build found something the document states differently, it is recorded as a numbered decision in the repository — never applied silently to the text. There are sixteen of those decisions so far; the four that matter to you are on the findings page.

Built on top of your own logs and predictions, preserved at Lagadnakul/Dissertation@c8872d0. Every figure on these pages is computed from that data, not copied from the document.