corteX Lens iconcorteX Lens
  • Home
Sign inSign up
corteX Lens iconcorteX Lens
corteX Lens iconcorteX Lens

Your workspace is waiting on a wider screen

Lens lays out a lot of signal at once, so it is built for the desktop. Open it on a laptop or a larger display and everything will be right where you left it.

corteX Lens iconcorteX Lens
  • Home
Sign inSign up
corteX Lens iconcorteX Lens
corteX Lens iconcorteX Lens

Your workspace is waiting on a wider screen

Lens lays out a lot of signal at once, so it is built for the desktop. Open it on a laptop or a larger display and everything will be right where you left it.

Reproducible savings methodology

Self-graded savings, with the math.

Lens computes a headline savings number for every customer. That number is the same in every report your team reads. This page publishes the math chain, the paired-trial protocol, the parity gate, and the weekly numbers corteX Lens runs on itself. The notebook reproduces the result bit-for-bit on your own audit-log export.

The math chain (Section 1) and the parity gate (Section 3) carry the trust contract. The headline below is a paired-trial median with a 95 percent CI, not a marketing roll-up.

Combined headline (weekly)

43.14%

Combined input tokens removed, token-weighted across 1 pinned benchmark, run 2026-W25.

Per-session median savings 25.26%. 95 percent CI: 0.00% to 0.00% (stratified bootstrap, dataset id is the stratum). Wilcoxon p = 0.21140 (n = 27).

real-developer-transcripts

rev claude- (n=11/42)

25.26%

Run cost: $0.00 of $50.00 budget. Generated 2026-06-18T00:00:00Z.

Run history

Every measured run, in order.

The headline above is the current run. This timeline shows the full measured progression so you can see how savings and parity moved across model and rule changes. Each row links to the document that carries its math. Forecast-only runs are excluded; this is measurements only.

  1. v1

    Baseline (pre-patch)

    Narrow gate
    Median savings
    23.07%
    Parity (judged)
    17.84%
    Rows attempted
    200
    Wilcoxon p
    < 0.0001

    Highest savings of any run, but parity was the weakest at 17.84 percent. The gate stayed narrow because parity was the open worry.

    Worker gemini-2.5-flash, judge gemini-2.5-flash, 2026-05-31. source

  2. v2

    Patches A-E (model upgrade)

    Partial improvement

Per-level summary

Measured outcomes per rule profile

Sealed runs only, grouped by the Removal Catalog lever their source documents state. A run appears in a group only when its committed document names the rule configuration; nothing here is inferred.

parity-first rules2 sealed runs

  • R5-off: 8.92 percent median savings, 62.87 percent parity, worker gemini-2.5-flash, 2026-06-01.
  • gate: 8.85 percent median savings, 49.50 percent parity, worker gemini-3.5-flash, 2026-06-08.

savings-max rules1 sealed run

  • v2: 10.91 percent median savings, 47.79 percent parity, worker gemini-3-flash-preview, 2026-06-01.

Lever not stated: v1, v3, realdata-cc. Their source documents do not name the rule configuration, so they are not grouped; absent means the document does not say, never a guess.

Same-stack caveat from the R5 ablation row, verbatim: Run on the v1 model stack, so savings is not directly comparable to v2 and v3. No two sealed runs on different levers share a worker model, so per-level savings are not directly comparable across levers.

The current published run measures Lens on real Claude Code developer transcripts, with the R7 transcript-tool-result-window rule as the savings lever. The headline is 43.14 percent of input tokens removed (token-weighted aggregate; 43.17 percent on large-context sessions), with a 25.26 percent per-session median. Quality shows no measurable change (composite quality delta 0.000), and the parity sample is indistinguishable from the baseline noise floor at this size (Wilcoxon p=0.21), so the verdict is NARROW, not GO. Raw transcript content is private and withheld. Earlier Agent-FLAN runs remain in the timeline as measured history but are no longer the published number.

Open methodology

The math chain in plain language.

Six chained steps from raw tokens to the dollar number on your dashboard. Each step has a tested implementation in TypeScript and a byte-stable Python reference. The methodology is signed per commit and the Removal Catalog version. Roughly a three-minute read.

Read the methodology

Verify on your data

Run the same math against your own audit log. The notebook prints a report that matches the dashboard headline for the same window, and the full walkthrough lives on the verify guide. Verify it yourself.

What is saving the tokens

What is saving the tokens

The proven lever, and where the saving shows up.

This run has one measured savings lever, not a stack of estimates. The bars below group the same paired trials to show where across the run the saving lands. Every figure is read from the published run cl-realdata-cc-2026-06-18, and a cohort below the privacy floor is withheld rather than shown.

Proven savings lever

R7

Transcript tool-result windowThe one change that produced the saving on this run. There is a single measured lever here, named and traceable, not a roll-up of many small estimates.

Windows oversized tool-result bodies in Claude Code transcripts so the largest-context sessions carry most of the input-token saving.

Token-weighted savingsThe share of total input tokens removed across the whole run, weighted by size. This is the aggregate cut a budget owner feels on the bill.

43.14%

Per-session median savingsThe middle session's saving. Half of sessions saved more and half saved less, so one very large session cannot inflate the figure.

25.26%

Where the saving shows up

By input complexity

Each bar is the measured per-cohort median input-token saving for this grouping. A cohort is a different slice of the same paired trials, so the bars show where the lever pays off, not an additive share that sums to the headline.

Loading the savings breakdown chart

Trust check

Verdict, quality, and parity floor

Gate verdictThe gate's grade for this run. NARROW means the saving is real and proven on this dataset, but the claim is scoped to the workload it was measured on rather than promised everywhere.

NARROW

Quality readWhether the saving cost answer quality. The transformed arm is compared against the baseline replayed against itself. Quality reads held only from a published composite quality delta at or above parity; a parity figure that merely fails to separate from the noise floor at this sample size is inconclusive, not proof that quality held.

Inconclusive

Inconclusive at this sample size: the transform-versus-floor Wilcoxon p is 0.2114 against the 0.05 bar, so quality is neither proven to hold nor shown to move.

Parity vs noise floorThe rate at which both arms produced an equivalent answer, read against the baseline-versus-baseline noise floor rather than against 100 percent. The Wilcoxon p is the real read: above the 0.05 bar the difference from the floor is unresolved at this sample size, which is inconclusive, not proof that quality held.

26.20%

vs 36.10% baseline noise floor. 42 paired rows, Wilcoxon p 0.2114

Parity is a two-run differential: the baseline arm is replayed against itself to set a quality noise floor, then the lens transform arm is tested against that floor with a paired Wilcoxon signed-rank test. The paired Wilcoxon p of 0.2114 does not reach the 0.05 bar, so the transform arm is not distinguishable from the baseline floor at this sample size.

Every figure on this panel is read from the published run cl-realdata-cc-2026-06-18. The savings and parity numbers come from /proof/latest-numbers.json, and the gate verdict traces to the sealed gate document. The math chain is on the methodology page. No synthetic numbers, and cohorts below the k-anonymity floor are withheld.

Rendered 2026-08-15T19:50:30.337Z. Source: docs/proof/ in the QuestoM/cortex-lens repo (methodology.md, reproduce.ipynb, weekly-numbers/). The weekly files are written by the cost-efficiency-ide benchmark harness on its Monday cron. No synthetic metrics. If a source is unreachable the panel reports it.

Median savings
10.91%
Parity (judged)
47.79%
Rows attempted
200
Wilcoxon p
< 0.0001

Parity lifted to 47.79 percent on the judge-completed denominator (65 of 136 paired). The conservative read across all 200 attempted rows is 32.5 percent. Savings fell from v1 mostly because of the model swap, not the patches.

Worker gemini-3-flash-preview, judge gemini-3.1-pro-preview, 2026-06-01. source

savings-max rules

  • v3

    Block C-1 (all fixes applied)

    Narrow gate
    Median savings
    10.65%
    Parity (judged)
    47.06%
    Rows attempted
    200
    Wilcoxon p
    < 0.0001

    Prior published headline, superseded by the CL-PG-3 sealed gate. Parity 47.06 percent on the judge-completed denominator (64 of 136 paired); 32.0 percent across all 200 attempted. Wilcoxon p below 0.0001. Gate stays narrow because savings is under the tens-of-percent bar.

    Worker gemini-3-flash-preview, judge gemini-3.1-pro-preview, 2026-06-01. source

  • R5-off

    R5 ablation (R5 disabled)

    Narrow gate
    Median savings
    8.92%
    Parity (judged)
    62.87%
    Rows attempted
    200
    Wilcoxon p
    < 0.0001

    Highest parity of any run at 62.87 percent (167 paired) by turning off the tool-result-trim rule. The trade was about 2 percentage points of savings for about 15 points of parity. Run on the v1 model stack, so savings is not directly comparable to v2 and v3.

    Worker gemini-2.5-flash, judge gemini-2.5-flash, 2026-06-01. source

    parity-first rules

  • gate

    CL-PG-3 sealed gate (retry + judge-deadlock fixes)

    Narrow gate
    Median savings
    8.85%
    Parity (judged)
    49.50%
    Rows attempted
    200
    Wilcoxon p
    < 0.0001

    Superseded headline. This Agent-FLAN sealed gate (replay_failed 0 of 200, parity 49.50 percent, savings 8.85 percent, Wilcoxon p below 0.0001) was the published number through mid June. It is no longer the published headline: the current run is the real Claude Code developer-transcript run below. Kept here as honest measured history, not shown as current anywhere.

    Worker gemini-3.5-flash, judge gemini-3.1-pro-preview, 2026-06-08. source

    parity-first rules

  • realdata-cc

    Real developer transcripts (Claude Code)

    Published now
    Narrow gate
    Median savings
    25.26%
    Parity (judged)
    26.20%
    Rows attempted
    42
    Wilcoxon p
    0.21140

    Current published run, measured on real Claude Code developer transcripts from the owner's machine via the lens-verify kit, with the R7 transcript-tool-result-window rule as the savings lever. Headline 43.14 percent of input tokens removed (token-weighted aggregate; 43.17 percent on large-context sessions over 8K input tokens, n=31); 25.26 percent per-session median; tiny prompts under 2K tokens save about 0 percent. Quality is unchanged: composite quality delta 0.000 across 41 paired rows, and the rigorous transform-vs-floor parity Wilcoxon is p=0.2114 (n=27), not statistically significant, so quality is indistinguishable from the baseline noise floor at this sample size. Funnel: 40 rows attempted, 35 replayed cleanly, 5 replay_failed; savings and parity are over the 42 valid replayed pairs (parity 11 of 42). An offline cross-check over all 5000 real Claude Code rows reads 43.95 percent token-mass reduction by the same rule. Raw transcript content is private and withheld from the public artifacts. Verdict NARROW, never GO.

    Worker gemini-3.1-pro-preview, judge gemini-3.1-pro-preview, 2026-06-18. source