Repository navigation
docs(interview): data inventory and bilingual resume material - #17
Merged
Merged
Conversation
…ges have a before and after A read-only pass over the brief's §4 ledger and the git history, answering two questions: which quotable figures have not reached the resume material, and which changes can be told as "I made a change, here is the outcome". Ledger usage, 77 rows: 46 already used, 23 unused but safe, 8 unused and unsafe. Worth knowing that rows 2 and 3 -- the edge counts, the figure an outside reader would most readily take for a headline -- appear in the resume material only inside the forbidden-phrasings table, as an example of what not to write. That is deliberate. The unused figures are re-tiered by whether they can carry a bullet on their own (A), need a qualifying sentence (B, interview ammunition only), or are scale figures (C). Tier C is completed from the whole repository rather than only from the ledger, since scale is what a non-technical reader reads first and the resume material used it least. Improvement chains are only counted when all four elements are present: a sourced before-state, a change with a file:line or commit, a sourced after-state, and causality recorded in code or in a commit rather than inferred afterwards. Each is classified, and the classification is the point: - Engineering, 4. Unifying the symbol-resolution rule (verified citations 48-54 -> 62-67, redaction markers 17-21 -> 9-13, symbol-form citations surviving for the first time). Hydrating graph evidence (L3 call-graph contribution -0.0067 -> +0.0226). Filtering test code out of retrieval (47% of retrieved candidates -> 0%). And call resolution, where the commit measured the downstream effect itself and recorded it as zero -- that one cannot be told as an outcome. - Measurement correction, 5. Symmetric scoring, the GRAPH_TOOLS set, the B3.5 arm, dropping groundedness from the composite, and scoring an uncited answer zero. - Parameter change, 4. One of them is unusable: the comment behind the rerank passage cap quotes latency figures that were observed once and never benchmarked, and this project quotes no latency anywhere. Engineering chains are few and that is the project's shape rather than a gap in the search. The dense iteration is almost entirely about how things are measured. Measurement corrections are not moved into the engineering column to improve the ratio -- calling one an optimisation invites "how much did it improve", and the honest answer is that the metric did not move, the earlier measurement was wrong. Seven chains are listed separately as "suspected, evidence incomplete" so they stay out of the resume material.
…e guide The same facts as the technical brief, reorganised for a mentor who will edit a resume and who reads once. Chinese and English are complete independently rather than one being a translation of the other. Figures come only from the brief's §4 ledger and the data inventory. Where the ledger attaches a qualifier, the text either carries it or uses a phrasing that does not need one. Lead order is deliberate: citation verification, then how the agent is kept in check, then the evaluation methodology. The call graph does not open, because my own control arm measured it down and opening with it means opening on the weakest link. The non-ML boundary is stated outright -- no training, no loss, no RL, every model an off-the-shelf call, and the static call-graph resolution labelled as program analysis. Two Mermaid diagrams, drawn only from components that exist. `chunks.tsv` and `index_runs` are dashed and marked NOT WIRED so nobody reads them as part of the data flow. The section that makes this usable is "what each number proves", in three groups. Group one is safe on a resume, with a column saying what each figure means to a reader rather than restating the figure. Group two lists the qualifier each number needs and the specific follow-up question that lands if it is missing. Group three stays verbal, with the reason it stays off the resume: needs full context, is a diagnostic excluded from the verdict, is configuration a reader cannot judge, or is not an outcome. Bullets carry only ledger-backed figures or mechanistic results. There are no throughput, latency, token-cost or accuracy-lift numbers, and the notes to my mentor explain per category why adding any would mean inventing them. Six phrasings that collapse under the first follow-up are listed as must-not-reintroduce, since an editing pass would otherwise restore them. Two disclosures are written where they cannot be missed. Filtering test code also mechanically moves the metric toward the ground truth, and the commit that made the change says so; I have to say it first rather than have it read out of the commit. And the citation counts come from single-baseline runs carrying no verdict, whose own report says the counts are the evidence and the means are weak, so the means are never quoted. Seven bullets is too many for one resume, so the file ends with a by-channel selection instead of a deletion: four for technical applications, three for HR screens, with a reason for each.
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
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.
Two interview-preparation documents built on top of the brief that #16 landed.
Documentation only — nothing outside
docs/interview/changes, and no code,evaluation artifact or run was touched.
2e1c41aDCODE_DATA_INVENTORY.md— which ledger figures are unused, and which changes have a real before and afterce272feDCODE_RESUME_ASSETS_CN.md/_EN.md— bilingual resume material with a per-figure usage guideThe inventory
A read-only pass over the brief's §4 ledger and the git history.
Ledger usage, 77 rows: 46 used, 23 unused but safe, 8 unused and unsafe.
Rows 2 and 3 are the edge counts — the figure an outside reader would most
readily take for a headline — and they appear in the resume material only
inside the forbidden-phrasings table, as an example of what not to write.
Unused figures are re-tiered by whether they can carry a bullet alone (A),
need a qualifying sentence (B, interview only), or are scale figures (C).
Tier C is completed from the whole repository rather than only from the
ledger, since scale is what a non-technical reader takes in first and the
resume material used it least.
Improvement chains are only counted when four elements are all present:
a sourced before-state, a change with a
file:lineor commit, a sourcedafter-state, and causality recorded in code or in a commit rather than
inferred afterwards.
citations 48–54 → 62–67, redaction markers 17–21 → 9–13, symbol-form
citations surviving for the first time). Hydrating graph evidence (L3
call-graph contribution −0.0067 → +0.0226). Filtering test code out of
retrieval (47% of retrieved candidates → 0%). And call resolution, where
the commit measured the downstream effect itself and recorded it as
zero — that one is explicitly not tellable as an outcome.
GRAPH_TOOLSset,the B3.5 arm, dropping groundedness from the composite, and scoring an
uncited answer zero.
passage cap quotes latency figures observed once and never benchmarked,
and this project quotes no latency anywhere.
Engineering chains are few, and that is the project's shape rather than a
gap in the search. Measurement corrections are deliberately not moved into
the engineering column to improve the ratio — calling one an optimisation
invites "how much did it improve", and the honest answer is that the metric
did not move, the earlier measurement was wrong.
Seven further chains are listed as suspected, evidence incomplete, so
they stay out of the resume material.
The resume material
Same facts as the brief, reorganised for a mentor who reads once. Chinese
and English are each complete rather than one being a translation.
Figures come only from the §4 ledger and the inventory. Where the ledger
attaches a qualifier, the text carries it or uses a phrasing that does not
need one.
Lead order is deliberate: citation verification, then how the agent is kept
in check, then the evaluation methodology. The call graph does not open,
because the project's own control arm measured it down. The non-ML boundary
is stated outright — no training, no loss, no RL, every model an
off-the-shelf call, and the static call-graph resolution labelled as
program analysis.
Two Mermaid diagrams drawn only from components that exist.
chunks.tsvand
index_runsare dashed and markedNOT WIRED.The section that makes the file usable is "what each number proves", in
three groups: safe on a resume, needs a qualifier (with the specific
follow-up question that lands if it is missing), and verbal only (with the
reason it stays off — needs full context, is a diagnostic excluded from the
verdict, is configuration a reader cannot judge, or is not an outcome).
No throughput, latency, token-cost or accuracy-lift numbers, and the
notes to a mentor explain per category why adding any would mean inventing
them. Six phrasings that collapse under the first follow-up are listed as
must-not-reintroduce, since an editing pass would otherwise restore them.
Two disclosures sit where they cannot be missed. Filtering test code also
mechanically moves the metric toward the ground truth, and the commit that
made the change says so. And the citation counts come from single-baseline
runs carrying no verdict, whose own report states that the counts are the
evidence and the means are weak, so the means are never quoted.
Seven bullets is too many for one resume, so the file ends with a
by-channel selection rather than a deletion: four for technical
applications, three for HR screens, with a reason for each.
Checks
Style and fact discipline were verified mechanically after each pass:
banned phrasings at zero, every figure traced to the ledger or the
inventory, the per-figure "what it proves" column within its length cap,
and the diff confined to the intended sections.