How Descant elevated the model’s work

Which corrections are removed

The original 53.8/100 calculation removed independent-review corrections only. It remains available unchanged. The headline comparison also includes the plan and implementation corrections made before independent review.

Plan findings have no structured file location. Their category assignments are disclosed analyst judgments; build and independent-review records retain the original file-path rule. The middle row shows the result with those plan judgments excluded.

The scoring rule caps each category’s deduction at two anchor points. Once a category reaches that cap, further findings do not lower its score. This limits the numerical effect of adding the author-side evidence; the cap and weights were not changed.

What is being scored

Weighted rubric points normalized to 100. These are not percentages of tasks passed, benchmark scores, probabilities or percentiles. Each category has an anchor from 0 to 5. Engineering has weight 51; process has weight 45. The rated total is 96.

Cost and delivery efficiency, with weight 4, is excluded: Completion cost is unknown; the projected assessment leaves this domain unrated.

Group score = 100 × sum of (category weight × anchor) ÷ (5 × included weight). The table shows both weighted sums so the normalization can be checked.

The deduction rule and all 15 categories

Per retained finding, the declared deductions are 0.5 anchor points for blocker, 0.25 for high, 0.1 for P2 and 0.05 for P3. Sum the deductions in a category, cap them at 2, then subtract from the projected anchor with a floor of 1. These are chosen scoring assumptions, not measured effects.

Five process categories are set to zero by assumption: independent review quality, author response and convergence, follow-up quality, issue grooming, learnings and their grooming. This does not establish that another model cannot perform those activities. Excluding these categories removes them from both sides of the overall calculation above.

The rule applies to the six engineering categories and requirements traceability. Decomposition, scope and communication stay at their projected anchors. A zero finding count means no retained thread was assigned to that category, not that it has no defects.

blocker and high correspond to the source reviewer severity labels. P1 as a rubric category means traceability and must not be confused with a reviewer severity label. Display one decimal, round half up; retain exact fractions in the calculation output. Displayed endpoints can therefore differ by 0.1 from the displayed gap.

Review records and source scope

This assessment uses 233 finding threads with retained corrections across 70 merged PRs. The broader source corpus has 114 PRs and 103 merged PRs: 99 implementation, 2 design and 2 human preparation or scope PRs.

The recorded-results section uses 196 independent-review fixes on merged implementation PRs, excluding Minor findings. 196 + 24 + 12 + 1 = 233. The broader modeled-scenario input includes Minor and design findings and one retained earlier correction beyond the implementation-only non-Minor case-study measure. Those are overlapping views of the independent-review records, not additional findings.

The expanded calculation adds 585 author-side records: 560 from implementation PRs and 25 from the two design PRs. That is 411 plan corrections plus 174 build corrections, all recorded before the first independent review. The combined assessment uses 818 records. Records can overlap in the underlying problem they describe.

Author-side findings are published as sanitized excerpts from retained private records, with their recorded corrections. Linked PRs and commits identify the work, not public copies of every internal review. A recorded fixed answer is not an independent audit of every patch; a corrected plan is not necessarily a defect that reached code.

Retain the 232 roots whose final substantive author disposition is fixed, plus one earlier explicit correction that source inspection confirms remained at merge and the evaluated revision even though the final arbitration response declined further change as already correct. Every included root has a fix reply naming a commit. This establishes the recorded correction claim, not an individual patch audit of all 233 threads. The one retained earlier correction was inspected in its fix commits, merge and evaluated revision. 6 retained correction threads on unmerged PRs are excluded.

Original ordered file-path rules for independent and build review, plus disclosed per-record analyst judgment for 411 plan findings without structured file locations. Top-level inline review finding thread, not necessarily a unique defect or a unique corrective commit.

• File location is a proxy for rubric relevance, not a semantic classification of the defect or its effect.
• For example, a boundary defect found in a test file is assigned to tests rather than architecture.
• One thread receives one category even when its consequences span categories; repeated or overlapping findings need not represent independent defects.
• The exact rule order matters. A prose-only reconstruction assigned some engineering findings differently. The supplied final row assignments and ordered regex table, rather than the abbreviated prose, define this calculation.
• The public aggregate inputs reproduce score arithmetic; the separate public census and final row assignments support reproduction of root counts, dispositions and category totals. Reproduction is not empirical validation of the scoring choices.
• Plan assignments reflect the substantive engineering consequence described in the finding and correction; no source path is invented.
• The 411 plan assignments are judgments by one analyst. The plan-excluded sensitivity removes them; classification ambiguity is not a statistical confidence interval.

Sensitivity to the scoring rule

Every row keeps the same projected base: 83.9/100 for engineering and 79.3/100 overall. Only the declared penalties and cap change. These scenarios are not confidence intervals or performance bounds.

Assumptions and interpretation

• Use the same projected completion anchors for every reported sensitivity; do not substitute the observed snapshot or the later process-only reassessment.
• Treat the retained fixed findings as absent fixes and translate their reviewer severities to anchor deductions. This is an analyst-selected scoring rule, not a measured causal effect or a calibrated prediction.
• Set independent review, author response, follow-ups, issue grooming and learnings to zero in the overall scenario by definition. Model sessions can perform such activities; their absence is not established by another run.
• Hold decomposition, scope and communication at the original projected anchors. The author records include decomposition and scope concerns, but this scenario preserves the pre-existing held-category rule.
• Keep traceability under the finding-penalty rule. Both learning findings are retained in the source count; their category is set to zero and their individual penalties are not separately applied.
• Keep deferred and declined findings outside the additional penalty. They are already part of the assessed history; the model does not newly subtract them.
• Do not infer that removing fixes creates an executable or feasible completed system; no fixes were actually reverted and no resulting system was evaluated.
• Do not estimate hypothetical self-correction, initial investigation and planning benefits, interactions among fixes or the effect of unavailable records. Those unmeasured effects do not establish a net direction of bias.
• This is one projected assessment and a family of constructed scoring scenarios, not a matched experiment or a measured head-to-head comparison.
• The severity coefficients, cap, floor, original anchors, path mapping and plan-category judgments are assessment choices. Arithmetic correctness does not validate them as estimates of another system.
• Testing selected sensitivities provides no claim about all defensible rules, statistical uncertainty or performance bounds.
• A capped category can remain unchanged after some parameter changes; that is a property of the chosen formula, not proof that missing evidence or alternate categorization cannot matter.
• A fair outcome comparison would require actual comparable runs with common requirements, information, resources and evaluation criteria.

Download the method and evidence

The data includes category weights, anchors, finding counts, scoring rules and calculated results. The source records and recomputation script make the arithmetic inspectable.

• Complete evidence and reproduction bundle
• Assessment data and calculation results
• Full methodology
• Recompute the all-review comparison
• Author-side finding records
• Author-side category assignments
• Retained correction records
• Source pull requests and review comments
• Category table
• Sensitivity table
• Source and reproduction guide
• Original independent-review-only assessment
• Original independent-review-only methodology