Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
65 commits
Select commit Hold shift + click to select a range
65215cb
uorc: the performance vocabulary (step measure, codec point, corpus κ)
claude Nov 11, 2026
9db6cf9
uorc: an x86-filtered frame costs the same steps as its profile
claude Nov 11, 2026
0562c78
Prism model: performance as theorems (Uorc.Performance, UC-PERF-02..06)
claude Nov 11, 2026
b5cd9a6
Merge main (#132 Pages page model) into the performance theorems branch
claude Nov 11, 2026
544e85a
tests: one native test per performance row (CM-02 names UC-PERF-02..06)
claude Nov 11, 2026
7f6c72a
Uorc.Performance: bounded search, undominated choice and the potentia…
claude Nov 11, 2026
ca3eff8
Restore the generated client crates cargo fmt touched
claude Nov 11, 2026
f1eed54
Merge remote-tracking branch 'origin/main' into claude/project-thread…
claude Nov 11, 2026
9439a8a
Add kappa-named model states to the codec
claude Nov 11, 2026
69a0135
Merge remote-tracking branch 'origin/claude/project-thread-5x220g' in…
claude Nov 11, 2026
e8c6146
Merge remote-tracking branch 'origin/main' into claude/project-thread…
claude Nov 11, 2026
023b75c
Choose by bytes then steps; exclude x86 trials with zero hits
claude Nov 11, 2026
3ad7412
Merge lane 1 theorems; export with tie rule and x86 exclusion
claude Nov 11, 2026
05d1240
Merge main f94cce6
claude Nov 11, 2026
8273457
Keep the learned trial under a sequential bound; report warm-up progress
claude Nov 11, 2026
13f75ae
Write state-free frames over the v0.4 basis; add state 2; test over s…
claude Nov 11, 2026
cb5fafb
Merge main 2a66cdb; code each family's v0.4 pair whatever the weights
claude Nov 11, 2026
2ebae4b
ncm streams probe: streams per frame over an order-free merged state …
claude Nov 11, 2026
200d6e6
Table bits as a recorded facet: step the context table on the chosen …
claude Nov 11, 2026
b31e23b
ncm streams: shipped-form coder (decoder, per-stream lazy tables, par…
claude Nov 11, 2026
c8c90ea
Test: recorded table fields are the tables the frame codes with
claude Nov 11, 2026
880bc77
ncm streams: per-thread working tables, sorted deltas, k-way merge, i…
claude Nov 11, 2026
88480ff
Codec point names the state's kappa; tests on the exported model with…
claude Nov 11, 2026
99d54f2
Export the model with each family's v0.4 pair (export 11)
claude Nov 11, 2026
ffd377e
Decode host as the page holds it: 512 MiB per worker, decoder plus sp…
claude Nov 11, 2026
cc4440a
Keep the v0.4 profiled blocks: freeze the v0.4 coder, append the curr…
claude Nov 11, 2026
9dcecfb
Merge #143 (export 11) into the table lane
claude Nov 11, 2026
3c997b4
Merge main 7ac966d
claude Nov 11, 2026
48909e0
Window depth and match table bits as stepped facets: generic window e…
claude Nov 11, 2026
d7c2b6e
ncm streams: frame word (streams at bits 37-52, epoch at 53-57), nati…
claude Nov 11, 2026
168e4ed
One page host statement: Uorc.Page::pageWorker sets PAGE_WORKER, Clie…
claude Nov 11, 2026
c13dca0
Merge the table lane (cc4440a, #143 export 11); route recorded facets…
claude Nov 11, 2026
51859b5
Land the table step after #143: a word that records table bits is wri…
claude Nov 11, 2026
4208cdb
Merge #143 (main 7ac966d) into the table lane
claude Nov 11, 2026
a8ad207
Format the page host check
claude Nov 11, 2026
af8d266
Merge the table lane (gen_ncm fix, profiled era routing, page host)
claude Nov 11, 2026
33499c9
ncm streams: derive streams per frame from the declared decode host (…
claude Nov 11, 2026
012dbfa
Page verify reports its decode passes on the bar
claude Nov 12, 2026
cd2836a
Rely on the portfolio's window for references; deeper frames are the …
claude Nov 12, 2026
432affb
Registry projections for UC-PERF-16 and UC-PERF-17
claude Nov 12, 2026
93a112d
Portfolio test: the unappliable entry records a table field below TAB…
claude Nov 12, 2026
4d46d35
Merge main 86a6fd3; search conformance on a frame that continues no s…
claude Nov 12, 2026
4976103
Merge the table lane (93a112d test fix, page verify progress)
claude Nov 12, 2026
e0ef1af
One shared pass for twin strategies that both learn
claude Nov 12, 2026
4036fdf
ncm streams in the IR model (ncm_streams.py): encoder and decoder byt…
claude Nov 12, 2026
3bf6a30
Merge #143 head (e0ef1af, main 86a6fd3) into the window lane for the …
claude Nov 12, 2026
ae15939
Merge the streams lane into the joint export branch
claude Nov 12, 2026
961dfb3
Plan every strategy the frame can code, by weight: no budget
claude Nov 11, 2026
2edf2e1
Export Client.search as its own root; native compress with steps runs…
claude Nov 12, 2026
01660db
Streams in the codec library: dense regions, merge steps, archive dis…
claude Nov 12, 2026
2501a64
Streams in the model: host threads, derived streams facet, epoch and …
claude Nov 12, 2026
0264f59
Merge main b29ae74 (#143, squash of ea75e13)
claude Nov 12, 2026
3c6c70c
Tests: drop frame codings that repeat another test's check
claude Nov 12, 2026
ef9de7c
Export fixes; UC-PERF-08 states the shipped rule; full-plan time as u…
claude Nov 12, 2026
225b744
Keep every warmed model state the declared host has room for
claude Nov 12, 2026
4f33e8f
Regenerate the page projection
claude Nov 12, 2026
7f8f14a
warm-keep on the joint host: frames decoded at once are workers
claude Nov 12, 2026
90ebdc0
Observed axioms of the sizing helpers
claude Nov 12, 2026
88e42f6
Merge main cd43ca8 (#149)
claude Nov 12, 2026
55b6e3d
Uorc.Performance: the plan's facet and the two unmet time distances
claude Nov 12, 2026
ac243b9
Drop the duplicate full-plan facet (planCompressTime, stateReplayTime…
claude Nov 12, 2026
fe6c08e
Twins share one pass everywhere: state twins replay the state once
claude Nov 12, 2026
f9f0f96
Joint export: streams unshipped in v0.5, exported client rebuilt
claude Nov 12, 2026
8311a74
Merge remote-tracking branch 'origin/main' into joint-wip
claude Nov 12, 2026
b1dc33f
Keep the page worker's memory within its share while encoding
claude Nov 12, 2026
File filter

Filter by extension

Filter by extension


Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
8 changes: 6 additions & 2 deletions CONFORMANCE.md
Original file line number Diff line number Diff line change
Expand Up @@ -94,10 +94,12 @@ The three honesty levels (R2):
| `UC-PERF-05` | `build` | A trial carries its defect, the count of bytes at which its decode differs from the input. Admission fails every trial with a nonzero defect, keeping never raises a defect above its operands', and the trial kept from any admitted list has defect 0 (bestOfAdmittedIsExact): whatever codec point or named model state produced a trial, it can change the archive's size and never its meaning. |
| `UC-PERF-06` | `build` | A candidate dominates an incumbent when it is no larger in bytes and no larger in steps, and its gain is the drop in bytes plus steps. The gain never exceeds the drop in cost (dominanceLowersCost), no trial gains over itself, gains compose (a positive gain of a over b and of b over c gives at least their sum of a over c), and a positive gain implies no growth in bytes or in steps: dominance is a strict partial order, so consolidating winners cannot cycle. |
| `UC-PERF-07` | `build` | A trial abandoned at a bound is modelled as a trial of at most the bound's bytes. For every list of trials and every raw length, the portfolio of trials abandoned at the raw length settles to the same size as the portfolio of complete trials (boundedPortfolioSettlesTheSame), and a trial no smaller than one already kept changes neither the kept bytes nor the kept steps (skippedTrial theorems): a search that stops a candidate once it cannot win, or skips one known to be no smaller, writes an archive of the same size for fewer compress steps. |
| `UC-PERF-08` | `build` | Choosing by fewer bytes and then fewer steps, from the raw frame and any list of trials, keeps a trial that is no later in that order than every trial in the list and than the raw frame (chosenIsLeast, chosenIsNoWorseThanRaw), has exactly the size the bytes-only rule settles to (chosenKeepsTheSize), and is dominated by no trial it is no later than (leastIsUndominated); a trial no earlier than the kept one changes nothing (excludedTrial theorems). The rule lowers decode steps on byte ties at no cost in size. The released client breaks byte ties by schedule order until its model export adopts this rule. |
| `UC-PERF-08` | `build` | Choosing by fewer bytes and then fewer steps, from the raw frame and any list of trials, keeps a trial that is no later in that order than every trial in the list and than the raw frame (chosenIsLeast, chosenIsNoWorseThanRaw), has exactly the size the bytes-only rule settles to (chosenKeepsTheSize), and is dominated by no trial it is no later than (leastIsUndominated); a trial no earlier than the kept one changes nothing (excludedTrial theorems). The rule lowers decode steps on byte ties at no cost in size. The client's choice is this rule, in the model (Client.pick) and natively (compress_is_the_chosen_trial_uc_perf_08). |
| `UC-PERF-09` | `build` | The potential of a codec point is its bytes plus its steps. Consolidating a candidate into an incumbent lowers the potential by exactly the candidate's gain (consolidationSpendsItsGain), which is positive exactly when the candidate dominates, and over any sequence of candidates the potential never rises (potentialNeverRises). The potential is a natural number, so an engine that consolidates only dominating candidates never regresses and consolidates at most as many times as its first potential. |
| `UC-PERF-10` | `build` | The encoder's stop is a ledger over its coder events: its count of emitted bytes saturates at the cap (ledgerSaturatesAtTheCap), so a payload below the cap is exactly the unbounded payload, four header bytes, the emitted bytes and one final byte (stopKeepsEveryPayloadBelowTheCap), and a candidate is abandoned exactly when its emitted bytes reach the cap and at no other point (stopAbandonsExactlyAtTheCap). A coded bit at a 12-bit probability narrows the coder's width by at least one and by at least one part in 4096 (aOneNarrowsTheWidth, aOneSpendsAFraction, aZeroNarrowsTheWidth, aZeroSpendsAFraction). No exact rule decides a raw frame earlier than its ledger does (exactRawDecision): until a candidate's ledger reaches the cap, a continuation the model codes in few bytes can still make the coded frame the smaller one. Measured: random bytes code to 0.28 to 0.30 percent over raw under every shipped profile, so on random input the earliest exact stop comes at about 99.7 percent of the frame. |
| `UC-PERF-11` | `build` | Input the model cannot shrink costs one coder pass per scheduled candidate, so that cost falls only through passes that write the same bytes in less time. Strategies whose profiles differ only in mixer selection read the same contexts, slots, match and repeat state and mixers 2 and 3, so when no core is spare they are coded in one shared pass that writes each strategy's own payload (ncm::encode_pair_below). It changes no byte: the shared pass equals the two separate passes in both directions of the pair and when either side is abandoned, and the client's archives do not depend on the thread count. |
| `UC-PERF-16` | `build` | A frame's window facets step on its chosen archive (Client.stepWindow): its coding is recoded over one more earlier frame of its block at a time, then with one more match table bit at a time, and each step is kept only while the archive gets strictly shorter and the decoder fits the declared decode host; each facet stops at the first step that does not pay or does not fit, so no depth or table size is fixed. A frame is therefore the portfolio's archive byte for byte or a strictly shorter one, and it decodes to the frame. |
| `UC-PERF-17` | `build` | The profiled window entry over k earlier frames is one form for every k of at least 1 (Archive.profilePrefix, Archive.profileTail): three graph nodes per frame concatenate basis roots 1 to k in order, and the call to the coder reaches its payload and profile word three nodes further back per frame. For one and two frames it is the fixed entry byte for byte, and the decoder reads k from the entry, so window depth is a facet of the archive rather than a constant of the format. |

## product-interfaces

Expand Down Expand Up @@ -187,8 +189,10 @@ The three honesty levels (R2):
| `UC-PERF-05` | `build` | `kernel_theorem` | `not_measured` | A trial carries its defect, the count of bytes at which its decode differs from the input. Admission fails every trial with a nonzero defect, keeping never raises a defect above its operands', and the trial kept from any admitted list has defect 0 (bestOfAdmittedIsExact): whatever codec point or named model state produced a trial, it can change the archive's size and never its meaning. |
| `UC-PERF-06` | `build` | `kernel_theorem` | `not_measured` | A candidate dominates an incumbent when it is no larger in bytes and no larger in steps, and its gain is the drop in bytes plus steps. The gain never exceeds the drop in cost (dominanceLowersCost), no trial gains over itself, gains compose (a positive gain of a over b and of b over c gives at least their sum of a over c), and a positive gain implies no growth in bytes or in steps: dominance is a strict partial order, so consolidating winners cannot cycle. |
| `UC-PERF-07` | `build` | `kernel_theorem` | `not_measured` | A trial abandoned at a bound is modelled as a trial of at most the bound's bytes. For every list of trials and every raw length, the portfolio of trials abandoned at the raw length settles to the same size as the portfolio of complete trials (boundedPortfolioSettlesTheSame), and a trial no smaller than one already kept changes neither the kept bytes nor the kept steps (skippedTrial theorems): a search that stops a candidate once it cannot win, or skips one known to be no smaller, writes an archive of the same size for fewer compress steps. |
| `UC-PERF-08` | `build` | `kernel_theorem` | `not_measured` | Choosing by fewer bytes and then fewer steps, from the raw frame and any list of trials, keeps a trial that is no later in that order than every trial in the list and than the raw frame (chosenIsLeast, chosenIsNoWorseThanRaw), has exactly the size the bytes-only rule settles to (chosenKeepsTheSize), and is dominated by no trial it is no later than (leastIsUndominated); a trial no earlier than the kept one changes nothing (excludedTrial theorems). The rule lowers decode steps on byte ties at no cost in size. The released client breaks byte ties by schedule order until its model export adopts this rule. |
| `UC-PERF-08` | `build` | `kernel_theorem` | `not_measured` | Choosing by fewer bytes and then fewer steps, from the raw frame and any list of trials, keeps a trial that is no later in that order than every trial in the list and than the raw frame (chosenIsLeast, chosenIsNoWorseThanRaw), has exactly the size the bytes-only rule settles to (chosenKeepsTheSize), and is dominated by no trial it is no later than (leastIsUndominated); a trial no earlier than the kept one changes nothing (excludedTrial theorems). The rule lowers decode steps on byte ties at no cost in size. The client's choice is this rule, in the model (Client.pick) and natively (compress_is_the_chosen_trial_uc_perf_08). |
| `UC-PERF-09` | `build` | `kernel_theorem` | `not_measured` | The potential of a codec point is its bytes plus its steps. Consolidating a candidate into an incumbent lowers the potential by exactly the candidate's gain (consolidationSpendsItsGain), which is positive exactly when the candidate dominates, and over any sequence of candidates the potential never rises (potentialNeverRises). The potential is a natural number, so an engine that consolidates only dominating candidates never regresses and consolidates at most as many times as its first potential. |
| `UC-PERF-10` | `build` | `kernel_theorem` | `not_measured` | The encoder's stop is a ledger over its coder events: its count of emitted bytes saturates at the cap (ledgerSaturatesAtTheCap), so a payload below the cap is exactly the unbounded payload, four header bytes, the emitted bytes and one final byte (stopKeepsEveryPayloadBelowTheCap), and a candidate is abandoned exactly when its emitted bytes reach the cap and at no other point (stopAbandonsExactlyAtTheCap). A coded bit at a 12-bit probability narrows the coder's width by at least one and by at least one part in 4096 (aOneNarrowsTheWidth, aOneSpendsAFraction, aZeroNarrowsTheWidth, aZeroSpendsAFraction). No exact rule decides a raw frame earlier than its ledger does (exactRawDecision): until a candidate's ledger reaches the cap, a continuation the model codes in few bytes can still make the coded frame the smaller one. Measured: random bytes code to 0.28 to 0.30 percent over raw under every shipped profile, so on random input the earliest exact stop comes at about 99.7 percent of the frame. |
| `UC-PERF-11` | `build` | `build_evidence` | `not_measured` | Input the model cannot shrink costs one coder pass per scheduled candidate, so that cost falls only through passes that write the same bytes in less time. Strategies whose profiles differ only in mixer selection read the same contexts, slots, match and repeat state and mixers 2 and 3, so when no core is spare they are coded in one shared pass that writes each strategy's own payload (ncm::encode_pair_below). It changes no byte: the shared pass equals the two separate passes in both directions of the pair and when either side is abandoned, and the client's archives do not depend on the thread count. |
| `UC-PERF-16` | `build` | `build_evidence` | `not_measured` | A frame's window facets step on its chosen archive (Client.stepWindow): its coding is recoded over one more earlier frame of its block at a time, then with one more match table bit at a time, and each step is kept only while the archive gets strictly shorter and the decoder fits the declared decode host; each facet stops at the first step that does not pay or does not fit, so no depth or table size is fixed. A frame is therefore the portfolio's archive byte for byte or a strictly shorter one, and it decodes to the frame. |
| `UC-PERF-17` | `build` | `build_evidence` | `not_measured` | The profiled window entry over k earlier frames is one form for every k of at least 1 (Archive.profilePrefix, Archive.profileTail): three graph nodes per frame concatenate basis roots 1 to k in order, and the call to the coder reaches its payload and profile word three nodes further back per frame. For one and two frames it is the fixed entry byte for byte, and the decoder reads k from the entry, so window depth is a facet of the archive rather than a constant of the format. |
| `UC-WEB-02` | `build` | `build_evidence` | `not_measured` | The product shows the codec as a point of its modelled space: uorc inspect, the library's Info JSON and the Pages client name each frame's codec point (basis, ncm profile word, consolidated model state, lanes and its κ) and the exact steps its decode takes, as Uorc.Performance defines them, from the header alone; the page's model view shows the release's performance theorems, its closed claims with codec point, corpus κ, bytes, steps and the factory's decision, and the engine's tick, rule30 register and weights, exactly as the release publishes them in model.json. |
5 changes: 5 additions & 0 deletions Cargo.lock

Some generated files are not rendered by default. Learn more about how customized files appear on GitHub.

3 changes: 3 additions & 0 deletions Cargo.toml
Original file line number Diff line number Diff line change
Expand Up @@ -57,3 +57,6 @@ opt-level = 1

[profile.dev.package.uorc_decompress]
opt-level = 1

[profile.dev.package.uorc_search]
opt-level = 1
6 changes: 5 additions & 1 deletion client/gen/axioms.json
Original file line number Diff line number Diff line change
Expand Up @@ -131,5 +131,9 @@
"Archive.v04Profile1XTail": [],
"Archive.v04Profile2Tail": [],
"Archive.v04Profile2XTail": [],
"Archive.libraryPrefix": []
"Archive.libraryPrefix": [],
"NcmDecode.tbOf": [],
"NcmDecode.mbOf": [],
"NcmEncode.tbOf": [],
"NcmEncode.mbOf": []
}
Loading
Loading