pi-canon Part 2 · Pricing Recall

The Question

One end-to-end number, and the refusal it earned

Part 1 ended on 19 of 20 traps avoided against 8 of 20 for a floor with no memory, and it was careful to say what that number was not: package-level by construction, unable to attribute the result to the journal, the spine, or surfacing separately. The last section of that page named an ablation as the experiment the claim needed. This page is that experiment, run until the parts had prices.

The frame that survived contact with the data is not components but channels. Stored knowledge is inert until something carries it back into a session, and this design has exactly three carriers. Touch is unsolicited: a tool call touches a governed asset and the governing article's one-line capsule arrives, unasked. Recommendation is unsolicited too, and reaches further: a retriever ranks the articles that touch cannot reach, the off-spine and declared-rule residue, against the session's active context, and a suggestion rides only when its score stands out. Search is the agent's own act, and it is the only channel that reaches the journal. Two channels push, one pulls, and they are separated here because they fail separately. The studies priced touch and recommendation directly; search has no study of its own, and what this page can honestly say about it comes from the arm that carried it.

The arms that recur through the studies are worth naming once, because every chart below leans on them. Bare is no memory at all. Flat is the same distilled knowledge handed to the session as one plain document. Journal is the project's full history behind search. Pull is articles behind addresses the session must ask for. Push is pull plus surfacing on touch. Flat, pull and push carry the same distilled content compiled three ways; bare and journal are the floors on either side of it, nothing and everything.

Everything on this page comes from the 0.2 paper and its deposit: six preregistered studies of 75, 165, 105, 124, 120 and 90 graded cells, a 128-session surface experiment, and a 52-turn replay on a real store. A verifier shipped with the deposit recomputes 135 headline numbers from the raw artifacts. The worker throughout is one model at one setting, and section by section this page says what that buys and what it does not.

Addressing

Where reading the whole thing stops working

The cheapest objection to this design is that none of it is necessary: distill the knowledge, hand the session the file, done. Study 1 could not reject that objection, and the honest reading is that at 692 bytes it never could have: the flat comparator sat at ceiling on the four soundly graded decisions, leaving addressing no visible benefit to pay for. What study 1 did establish is the price of the opposite habit, treating history as memory. The arm that answered from the full journal burned a median of 340,119 tokens per session against the flat document's 21,309, for no more correctness. History is for appending, not for reading.

Study 2 made size the factor: the same distilled content, compiled at 0.7 KB, 52 KB and 259 KB, behind flat, pull and push access. The curves cross between 52 KB and 259 KB. At small, flat ties the addressed arms on correctness and wins on cost, 23,553 fewer tokens than pull. At large, the flat document scores 1.47 of 5 while the same content behind addresses scores 4.93, at a third fewer tokens than flat spends failing.

Why flat fails at scale took three corrections to state honestly, and the mechanism is not the one usually reached for. The answer strings still arrive: count them and every flat-large session got all five. What almost never arrives is the answer scoped to the thing it governs. This corpus asks the same question many times for different subsystems, so at 259 KB the document holds 201 occurrences of the answer pattern and only 69 of them govern the asset under test. The session's read window covered a fraction of the file with the governing sections past the cut, and the rival occurrences it did read arrive stripped of their scope. Strict scoped delivery came to 0.33 of 5; fourteen of fifteen cells got none of the five governed facts and still scored 1.21 on average, which is what guessing from plausible rivals looks like. An address is the unit that survives this: fetched by name, arriving with its scope attached.

The same study put a number on discovery, and it is the cleanest number in the programme. Pull requires the session to ask; 15 of 15 cells that asked for the decisive address scored, 0 of 30 that did not ask scored. Asking was perfectly predictive, and the gap between push and pull grew with store size, +0.07 to +0.87. Surfacing on touch is not a refinement of addressing. It is the thing that makes addressing work when the store is too big to already know.

Touch

Same store, same answers, three ways to reach them

Study 3 is the floor the touch channel stands on, and it is built from stores that all contain the answer. Every decision graded in it is answerable from every store in the study; what varies is whether anything makes the answer findable. No address and no surfacing lands at 0.67 of 5. One flat document lands at 2.53. Addressed and surfaced on touch lands at 4.87 against a bare floor of 0.00, with fewer median tokens than the flat document spends, 62,750 against 81,480, and an eighth of what the journal arm reads.

The comparison is a triangulation across arms built for other contrasts, matched on content rather than on bytes, and it is labeled post hoc exactly as the paper labels it. What it is for is the regime question: the difference between 0.67 and 4.87 is not storage, since the knowledge is present in both stores. It is representation, addressing and delivery, moving together. The addressed arms are told at session start that a store exists and where it lives; a flat file's presence tells a session nothing about what is inside.

Touch also had to survive subtraction. When the 2x2 further down found that presenting the memory costs first-pass correctness, the response was to delete surface pieces, and the deletions were priced here: study 3 detected no loss from removing the orientation line, and the whole result held with the reduced surface. What touch buys survived; what the wrapping cost did not.

Recommendation

The channel that pays in exactly one place

Recommendation is the programme's longest arc, and it opens with the channel losing. In study 3 the ranked residue had nothing to add: the arm carrying it ran beside the touch channel, which was already delivering the decisive articles, and the residue itself was filler drawn from the same value pool as the graded decisions. On decisions it was a null against push, and on the end-to-end check, an endpoint added post hoc, exploratory and uncorrected, it was the only memory arm to lose ground, at 27% more tokens. The channel shipped off, with that reason on the record.

Study 4 gave the residue something worth finding, and the same mechanism roughly quadrupled what the session walked away with. The store held declared rules governing no asset, the class of knowledge touch structurally cannot reach. Every recommendation arm beat push with search, +0.87 to +1.31 rule facts of 2. The preregistered claim was that the best of six arms beats push, and p=0.0002 is that selected best-arm contrast, unadjusted; all six beat push, the weakest at p=0.0015, and a Bonferroni correction across the six leaves the best contrast significant. Push carries the search action, so this is a beat over a search-carrying arm; five of its fifteen sessions did search and recovered the origin rule, so the contrast says recommendation increased rule-fact recall beyond that arm, not that search was unused or failed.

The same study killed the channel's original gate. The cutoff was an absolute score threshold, and an absolute score cannot ship: the retriever normalizes against a query-dependent ceiling, so the same quality of match scored 0.25 in study 3, 0.03 in study 4, and 0.30 in a synthetic probe. The shipped value would have been wrong for the next corpus and could silently deliver nothing. Its replacement is a ratio within the query, how far the top article stands out above the rest, and study 5 priced it on the capsule arms: at the shipped 1.4, the measured difference from the uncut channel was minus 0.07 rule facts at p=1.0, an n=15 null rather than an equivalence claim, while suggestions fell from 26.4 per session to 3.0 and the open rate rose from 0.17 to 0.82. Set the cutoff past the decisive reach, at 2.0, and the capsule arm goes silent, 0 rankings passed of 141, and hands back everything it had won, minus 0.73 at p=0.0015; the name arm at the same level passed one ranking of 132.

The arm that closes the argument is the decoy: the channel pointed at a store that genuinely cannot answer. It passed 0 of 139 rankings, suggested nothing, leaked nothing. A ratio cutoff is a relevance test, and a store with nothing relevant never stands out. That silence, more than any delivery number, is the case that the cutoff belongs on whenever the channel is.

The shipped default is 1.4, selected by three single-cell smoke tests at 1.0, 1.4 and 2.0 that were spent and reported before the protocol froze; the 120 cells then priced that frozen operating point rather than choosing it. The figure below is the argument in its rawest form: every ranking study 5's free-running arms performed, drawn at its measured standout ratio, with the cutoff in your hands. The decisive moments sit apart from the crowd; everything else sits at or below 1.33, with a median of 1.08. A cutoff below the gap passes noise. A cutoff above it silences the channel. The band between is wide, which is why one frozen constant worked at all.

The Surface

Presenting the memory costs more than anyone designed for

The most uncomfortable result in the programme has nothing to do with recall. A development observation suggested that merely presenting the memory system, before it stores or retrieves anything, was costing first-pass correctness: sessions with an inert memory surface shipped worse first attempts at the same task. The 2x2 below was frozen to localize it. Two factors, each inert: the tool schema, a description the model rereads every turn, and the orientation line, one steered sentence telling it to read before working. Four arms, 32 sessions each, one plant task, first-pass correctness graded deterministically.

With neither piece, 25 of 32 sessions shipped a correct first pass. With both, 8 of 32. Each factor's main effect is negative and significant, the tool schema at −1.05 log odds, the orientation line at −1.32, and the interaction straddles zero, so additive and worse-than-additive cannot be told apart at this size. One thing the design cannot say, and the paper's external review made it say so out loud: which piece costs more. That comparison reduces to 19 of 32 against 21 of 32, and it resolves nothing.

What followed was subtraction, executed by the numbers. The orientation line was removed, priced by study 3 detecting no loss without it. The check-back, a session-end nudge to revisit the store, was removed when study 6 measured it as a null at 0.96 power. The capsule on surfaced lines stayed, and the paper is explicit that this one is a judgment call rather than a verdict: one exploratory result favors capsules, a bare address spends a read round trip to teach what the capsule teaches on the line, and the data left it open. The caveat that cannot be subtracted: studies 4 through 6 ran with the old surface. Their internal contrasts are unaffected, and their absolute levels carry it.

The End-to-End Check

The rerun that had to disown its own recall numbers

Study 6 reran Part 1's chain shape under a frozen protocol, 30 repetitions, three arms, and its recall story begins with a contamination. The earlier harness left the prior sessions' transcripts on disk at recall time, and the bare arm, with an empty store, read at ceiling: 14 to 15 of 15. That number was not memory. It was the agent finding the transcripts. Study 6 removed them and forked recall from a hash-verified pre-probe snapshot, and the floor told the truth: bare fell to 2, 19, 0 and 1 of 30 across the four facts while the memory arms held 29 to 30 of 30. The one fact bare still gets, scheduler ids, is the one the checked-out code answers by itself.

The trap replicated at double the prior repetitions: a session two days later attributes a planted sweep to its actual job in 0 of 30 bare cells against 19 to 21 of 30 with memory, and every arm's own current task sat at 29 to 30 of 30 with every pristine check green. Memory arms do the work and carry the knowledge; bare does the work and structurally cannot.

The check-back died here, and the manner of its death is the method in miniature. The preregistered endpoint came back a risk difference of exactly 0.000 with an interval of −0.223 to +0.223, at 0.96 power for the effect it was designed to find. The post hoc trace made the null legible: in the nine canon plants that shipped broken, the nudge was delivered in eight, and the number of sessions that then touched the flagged file was zero. Delivered, and not acted on. The feature was removed the same day, and the removal is the study's contribution: a null at power, converted into one less thing every session pays for.

The Replay

A store it was never tuned on

Every constant above was priced on one corpus family, so before freezing, the shipped retriever was replayed over a real working store: 33 articles of live project memory, 52 real session turns, trace-level capture. The first thing the replay found was a failure. A store that small and saturated lets ranking find a regime the priced corpus never enters: once the still-eligible residue is a tail of near-zero scores, the standout ratio is a ratio over junk, and it explodes. A 4.79 standout on a 0.101-score match, the ratio anti-correlating with relevance, the whole 33-article store riding over the session down to scores of 0.002.

The fix is a consumed floor: while the cutoff is active, the crowd floors at the strongest already-delivered article the same query raises, a bar that moves with the query and the store rather than sitting at any fixed score. Re-run on the same trace, the rides drop from 33 to 15, the minimum observed ride score rises from 0.002 to 0.075, and the strongest ride improves, 0.184 to 0.264, because an article not spent on a junk query is still available for the query that deserved it. The bound on all of this is stated in the section title the paper gives it: none of these 52 turns needed history, so the replay measures the false-positive side only. It shows the drained-store failure going away on the same trace; it cannot show that surfacing helps when history is needed. Whether value ports is the next study, not this one.

The Verdicts

What shipped, and why, by the numbers that decided it

The 0.2.3 configuration is the studies' verdicts, executed. Four kinds of verdict, and every entry below names its evidence.

The Journey

Each study broke something in the one before it

Read in order, the eight runs are not a suite; they are a sequence, and the sequence is the finding. Study 1's null said the comparator was at ceiling, so study 2 made size the factor and found the crossover, and its review corrected the delivery instrument twice on the way. Study 3 shipped the recommendation channel off on a post hoc cost signal, mislabeled its own decoy, and saturated four cost measures, which taught the ceiling lesson the 2x2 then exploited to price the surface itself. Study 4 gave the channel something worth finding and paid for it: the run withdrew its own naming comparison through a re-offer defect, exposed why absolute cutoffs cannot ship, and lost a cell to a directory collision its correction later traced. Study 5 replaced the threshold with the standout ratio, sized its own internal null contrast, and rebuilt the decoy into a real control. The replay then found the regime the priced corpus never enters, and the drained-store guard closed it with one mechanism verified on the same trace. The three motivation runs before study 6 produced a chain-validity fix, the contamination discovery, and two observational hints; study 6 then quantified the contamination, replicated the trap at doubled repetitions, and returned the check-back null, and the day it reported, the check-back came out of the package. Every one of those breaks is in the record with a date on it.

The Fine Print

What none of this establishes, in the same breath

One worker model, one setting, throughout: every number above is a property of that pairing, and eligibility-style results move when the worker does. One corpus family built to have scoped rivals, and one task shape, sessions making graded decisions against planted knowledge; the crossover point and the standout gap are properties of that corpus, and the paper claims the shapes, not the constants. The 2x2 grades one plant task 128 times, which powers it without generalizing it. Studies 4 through 6 ran with the surface the 2x2 later priced. Study 4's attrition is accounted rather than summarized: 124 preserved of 135 launched, the loss differential over six arms, one grade dropped with its observation withdrawn. And the replay bounds the false-positive side only.

The review behind the numbers is the same discipline Part 1 used, grown teeth: five rounds of model review over the evidence base, two over the conversion, an external reviewer model over the built PDF, and verification rounds over every revision, each round certified by the next rather than by itself, every finding and disposition recorded in the claims map beside what replaced it. All of it is fact-checking by language models against the artifacts, a way of catching our own errors before they publish. None of it is peer review, and it substitutes for none. Nothing here is replicated elsewhere.

The deposit is the arbiter. Beside the paper: the per-cell artifact trail for all 679 study cells and the 128 surface sessions, worker transcripts and work trees and surfacing traces included, the frozen protocols with their amendment histories, the sealed corpora, both replay traces, and a claims map recording every review round. A verifier recomputes 135 headline numbers from those files under an identity header binding the run to the release. If a check fails, the paper and the deposit disagree, and the deposit is right.

Citations

The paper, the artifacts, and the work this stands on

The paper. Shane Conner, "pi-canon: Pricing Recall in Long-Term Memory for AI Agents", Zenodo, August 2026. doi.org/10.5281/zenodo.21960351. The deposit beside the PDF carries the artifact trail for every run above, the figure sources, the claims-to-evidence map, and the verifier that recomputes 135 headline claims from the files.

Part 1. The design, and the four-arm evaluation that measured it whole: shaneconner.com/projects/pi-canon. Its paper: doi.org/10.5281/zenodo.21890648.

pi-canon. The package: github.com/shaneconner/pi-canon, on npm as pi-canon, MIT licensed. The release this page describes is 0.2.3.

pi-fold. The companion package on the working tier, separate and optional in both directions: shaneconner.com/projects/pi-fold.

BM25. Stephen Robertson and Hugo Zaragoza, "The Probabilistic Relevance Framework: BM25 and Beyond", Foundations and Trends in Information Retrieval 3(4), 2009. The reference retriever behind the recommendation studies, and the normalization behavior that killed the absolute threshold.

The AGENTS.md convention. agents.md. The one-file shape the flat arms compile to.

Lost in the Middle. Nelson Liu and colleagues, 2023. arXiv:2307.03172. Positional degradation in long contexts; study 2's flat-large failure is a harder one, sections never entering context at all.